# Unbundling Work ## Comprehensive Research Plan for AI Workforce™, Labor as Code™, Native Alpha™, and Zero Meaningless Work Version: 5.0 Date: September 12, 2026 Purpose: Research context for GitHub specifications, AI engineering harnesses, product experiments, operating models, incentive and institutional analysis, transition and capability formation, risk and liability research, systemic security and resilience, causal evaluation, rights and knowledge ownership, and commercialization work. Trademark usage: Labor as Code™, AI Workforce™, Operational Truth™, Outcomes Acceptance Testing™, and Native Alpha™ are trademarks of Intellectual Frontiers LLC and are marked accordingly throughout this research plan. ## 1. Research area ### Working title Unbundling Work ### Core thesis The wrong question is: > How many people can AI eliminate? The better question is: > Which parts of work actually require human capability, and which parts have merely accumulated around humans because historically there was no practical alternative? The objective is not "Zero People." The objective is to minimize the unnecessary consumption of human capability. A useful shorthand for that objective is: > Zero Meaningless Work "Meaningless" does not mean boring, repetitive, or low status. Some repetitive work is essential control work. Some administrative work creates evidence, safety, compliance, continuity, or accountability. The research must therefore avoid labeling work as meaningless by intuition. For this research program, work is a candidate for elimination, automation, delegation, or redesign when human involvement does not materially improve one or more of the following: - judgment under uncertainty - authority or legal accountability - trust - empathy or relationship quality - creativity or problem framing - negotiation - physical presence or dexterity - exception handling - ethical responsibility - interpretation of ambiguous context - acceptance of consequential risk - generation of new knowledge - learning from unusual cases The research question is not whether a job can be automated. A job is usually too large and too historically contingent to be a useful unit of analysis. The unit of analysis should instead be the work. That means tasks, decisions, handoffs, authorities, evidence requirements, inputs, outputs, risks, outcomes, and exceptions. A further hypothesis of this research is that liability-bearing credentials are unusually important targets for unbundling. Where society requires malpractice coverage, professional liability insurance, errors-and-omissions coverage, bonding, surety, indemnification, regulatory capital, or similar instruments, it is signaling that the work carries consequential risk. That does not mean the credentialed person should be replaced. It means the work should be decomposed carefully enough to determine exactly which activities require the professional capability, authority, accountability, and liability envelope. The design objective should therefore include: > Use the minimum necessary credential, authority, and liability exposure required to safely produce an accepted outcome. This is not merely a labor-efficiency claim. It is a risk-engineering hypothesis. If work can be decomposed into narrower units with explicit evidence, authority, controls, acceptance tests, and accountability, then risk may become more observable, attributable, containable, and eventually more precisely financeable or insurable. Whether that actually reduces liability cost must be tested rather than assumed. ## 2. Why this research matters Most AI workforce discussions start with the wrong abstraction. They start with roles: - AI nurse - AI doctor - AI care coordinator - AI salesperson - AI project manager - AI analyst - AI compliance officer A role title hides more than it reveals. A role is a bundle of responsibilities assembled over time for reasons that may include regulation, organizational history, software limitations, information scarcity, coordination costs, labor availability, reimbursement rules, and habit. The unbundling argument makes this point directly: the role title is not the work. A "care coordinator," for example, may actually track patient status across systems, identify care gaps, reach out to patients, escalate high-risk cases, document in the correct system, and maintain continuity across encounters. Once these activities are made explicit, each can be designed, measured, delegated, assisted, or automated independently. This leads to a broader research claim: > Jobs are bundled abstractions. AI changes the economics of unbundling and recomposing them. Healthcare makes this especially visible because credentials, licenses, supervision rules, reimbursement structures, patient safety, professional liability, and clinical accountability force organizations to think more carefully about who may do what. Professional liability is not merely a compliance detail. It is an economic signal. If a profession is important enough that failures are routinely transferred, pooled, insured, bonded, indemnified, or otherwise financed, the surrounding work deserves close examination. A physician's malpractice coverage, an attorney's professional liability coverage, an engineer's errors-and-omissions exposure, or a director's D&O coverage each signal that consequential authority and risk have been bundled around a role. The research should ask whether the bundled role is actually the right unit at which that risk should be understood. The same logic should generalize to other regulated and consequential industries. High compensation may signal scarce capability. Meaningful liability instruments may signal consequential risk. The most attractive unbundling opportunities may occur where both are present and where the underlying work is observable and decomposable. ## 3. Research doctrine The research area should maintain several distinctions. ### 3.1 People are not the problem Human labor is expensive partly because human capability is scarce. The existence of a salary or labor cost does not mean the rational objective is to remove the person. In many settings the problem is the opposite: there are too few people with the judgment, trust, training, experience, authority, or physical presence required to do the important work. The research should therefore examine whether AI increases the effective supply of scarce human capability by removing work that does not need to consume it. ### 3.2 Headcount reduction can be an outcome, but should not be the governing objective A redesigned process may require fewer people. That is legitimate. But the research should distinguish among at least five outcomes: 1. Fewer people are required for the same output. 2. The same people produce substantially more output. 3. The same people deliver higher quality or safer outcomes. 4. Existing staff can perform work that previously required new hiring or outsourcing. 5. Entirely new work becomes economically feasible because AI reduces the cost of preparation, coordination, documentation, analysis, or execution. The first outcome should not be assumed to be the most valuable. ### 3.3 "Top of license" is a useful healthcare shorthand Healthcare already has language for this problem. A physician should spend scarce physician time on work that genuinely requires physician judgment, authority, accountability, and trust. An RN should not consume nursing capacity on work that could safely be completed by another worker, human or digital, when the nurse clinical capability is needed elsewhere. A pharmacist should not spend professional judgment time assembling information that can be reliably prepared for review. This research should use "top of license" where appropriate in healthcare, but generalize the idea outside healthcare as: > Top of capability The objective is to use people where their distinct human capability materially changes the outcome. ### 3.4 Unbundling comes before automation The research should resist the common sequence: Role -> AI agent -> deployment Instead use: Observed work -> decomposition -> evidence -> authority -> risk -> worker contract -> allocation -> acceptance test -> recomposition -> continuous verification This is one of the central propositions to test. ### 3.5 Recomposition matters as much as decomposition Unbundling is not the destination. After work is decomposed, it must be recomposed into a functioning operating system in which humans, AI workers, conventional software, physical systems, external services, and management controls work together. The unbundling-credentials argument frames this as recomposition rather than replacement. Human professionals retain judgment, ethical responsibility, final authority, and accountability while AI workers take on narrower capabilities such as recall, drafting, structuring, summarization, classification, and routing. The research should test how far this model can extend, where it fails, and where humans should retain direct execution rather than merely supervisory responsibility. ### 3.6 Credential exposure should be treated as a scarce resource A credentialed professional contributes more than minutes of labor. The organization is consuming a bundle that may include: - scarce professional time - demonstrated knowledge and skill - legal scope of practice - institutional privilege - professional authority - reputation and trust - ethical duties - supervisory responsibility - legal accountability - malpractice or professional liability exposure - organizational indemnification or insurance capacity The research should therefore measure more than labor minutes. It should investigate credential exposure per accepted outcome. Candidate question: > What is the minimum necessary credential exposure required to safely produce this outcome? This should not be interpreted as a mandate to remove credentialed professionals from work. In some workflows more direct professional involvement may be safer, more trusted, or economically superior. The purpose is to make the exposure explicit rather than accidental. A second principle follows: > Liability is evidence that work needs to be decomposed carefully, not evidence that it cannot be decomposed. A highly insured profession may be a poor automation target if its core work is inseparable, tacit, physical, or highly contextual. The actual target is liability-decomposable work: consequential work in which preparation, evidence gathering, execution, approval, communication, supervision, and accountability can be separated without losing safety or legitimacy. ### 3.7 Unbundling should concentrate Native Alpha™ Unbundling Work should not stop after separating human work from machine work. It should ask what unusual capability remains after commodity, transferable, routine, and automatable work has been stripped away. Native Alpha™ is the differentiated capability that a particular person, team, organization, workflow position, data asset, relationship network, intellectual asset, or operating system is unusually positioned to know, see, decide, or do. A credential is not proof of Native Alpha™. A job title is not proof of Native Alpha™. Seniority is not proof of Native Alpha™. High compensation is not proof of Native Alpha™. The unbundling process should test whether supposed differentiation survives decomposition. The research should therefore pair two design questions for every consequential work unit: > What is the minimum necessary credential, authority, and liability exposure required for a safe accepted outcome? And: > What differentiated human or organizational capability would materially improve this outcome? This creates a paired optimization objective: > Credential Exposure ↓ / Native Alpha™ Density ↑ The goal is not simply to remove professional minutes. It is to remove commodity work from scarce people while preserving, concentrating, and where possible scaling the particular human capabilities that change outcomes. This also means some technically automatable work should remain human. If a human interaction is where trust compounds, unusual judgment is exercised, proprietary learning occurs, referrals are generated, relationships deepen, or new knowledge is created, removing the person may destroy Native Alpha™ even if the task can be automated. The research should distinguish four dispositions for differentiated capability: 1. Preserve it because direct human participation creates value. 2. Amplify it by removing surrounding commodity work. 3. Encode portions of it where doing so increases organizational capability without destroying the source of differentiation. 4. Reject it as false differentiation when decomposition shows that the capability is commodity, reproducible, or irrelevant to outcomes. The deeper hypothesis is that unbundling destroys false differentiation and exposes real differentiation. ### 3.8 Work exists because of systems of incentives, not only because tasks need doing A work unit can look wasteful and still be rational behavior inside the current institution. Healthcare work may exist because of: - reimbursement rules - payer utilization controls - documentation requirements - defensive medicine - legal discovery expectations - credentialing rules - organizational budget structures - departmental boundaries - vendor limitations - contractual obligations - information asymmetry - fragmented ownership - patient-safety controls - internal performance measures - market power or cost shifting between parties The research must therefore ask not only what a worker does, but why the work exists and who benefits from its continued existence. A central principle is: > Before automating a burden, identify the mechanism that creates the burden. Some work should be automated. Some should be eliminated by removing the upstream cause. Some requires contractual, regulatory, reimbursement, or organizational redesign. Automating a bad institutional bargain can make the bargain cheaper without making the system better. ### 3.9 Capability formation is part of the work system Routine work can be meaningless repetition, but it can also be how professionals learn. Residents learn by seeing ordinary cases before unusual cases. Nurses develop pattern recognition through repeated encounters. Engineers learn failure modes by operating systems. Lawyers learn judgment through research and drafting. Coding, compliance, regulatory, and operational professionals develop tacit knowledge through repeated exposure. The research must therefore distinguish: - work whose repetition creates little additional capability - work whose repetition is deliberate practice - work that builds pattern recognition - work that teaches exceptions and failure modes - work required for credential maintenance or progression - work that creates institutional memory A useful principle is: > Do not eliminate a work unit until you understand whether performing it creates future capability. Zero Meaningless Work should not become Zero Apprenticeship. ### 3.10 Scaled AI creates correlated failure risk Human workers often make heterogeneous mistakes. Machine workers can make highly correlated mistakes because thousands of work units may depend on the same model, prompt, policy, data source, tool, integration, or vendor. The research must therefore treat common-mode failure, dependency concentration, and adversarial manipulation as first-class properties of the AI Workforce™. A worker that is 99.9 percent reliable at one case may still be unsafe if the same failure can affect 100,000 cases before detection. ### 3.11 Recomposition creates rights and ownership questions When unusual human judgment, operating know-how, workflows, evaluation cases, policies, evidence structures, or decision patterns are encoded into Labor as Code™, the organization may be creating intellectual assets. The research must ask: - who owned the underlying know-how before encoding - who owns the resulting worker contract, prompt, schema, evaluation set, workflow, data product, and operating procedure - what is background intellectual property versus newly created foreground intellectual property - what rights model providers, vendors, employees, contractors, affiliates, and customers receive - whether the resulting capability can be moved to another vendor or model - whether encoding Native Alpha™ strengthens control of the advantage or accidentally transfers it away Native Alpha™ can only compound commercially when the organization has clean, usable rights to the capability it is trying to scale. ## 4. Intellectual architecture The research program should distinguish between the destination, the method, the implementation discipline, and the operating model. ### Destination: Zero Meaningless Work Minimize work that consumes scarce human capability without materially benefiting from that capability. ### Discovery method: Unbundling Work Decompose roles, workflows, credentials, organizational structures, and software-mediated processes into the actual units of work. ### Implementation discipline: Labor as Code™ Represent work explicitly enough that it can be inspected, versioned, assigned, tested, governed, automated, and improved. Labor as Code™ is the method for moving from roles to tasks, tasks to workflows, and workflows to outcomes. ### Operating model: AI Workforce™ Recompose work across: - humans - AI workers - deterministic automation - conventional applications - databases and source systems - physical machines - outside service providers - governance and control mechanisms The AI Workforce™ is not a catalog of bots. It is the complete human-machine operating system that produces an outcome. ### Evidence layer: Operational Truth™ Continuously establish what is actually happening in the operation through queryable, inspectable, source-linked evidence. ### Trust layer: Credentials, authority, risk, liability, and accountability Determine what each worker is qualified to do, allowed to do, accountable for, prohibited from doing, and what legal or financial exposure attaches to the work. The trust layer should make explicit which person, organization, vendor, insurer, risk pool, or other party bears responsibility when a work unit fails. ### Institutional layer: incentives and work creation Explain why each major work unit exists, which party imposes or benefits from it, and whether the burden should be automated, eliminated, renegotiated, or redesigned upstream. ### Transition and capability layer: migration, learning, and workforce development Represent how the current operation moves safely toward the recomposed operation while preserving the learning pathways required to create future expert capability. ### Resilience layer: security, dependency, and graceful degradation Make explicit the dependencies, adversarial threats, common-mode failures, fallback modes, recovery procedures, and concentration risks that appear when work becomes machine-mediated. ### Rights layer: ownership and knowledge capture Track who owns the inputs, workflows, encoded know-how, evidence, outputs, evaluation assets, and foreground intellectual property created by the recomposed system. ### Recomposition principle: Native Alpha™ Use the decomposition to identify which human and organizational capabilities are genuinely differentiated, then design the recomposed system to concentrate those capabilities where they materially change the outcome. Native Alpha™ is not another worker type. It is the strategic allocation principle that tells the AI Workforce™ what should be preserved, amplified, or allowed to compound rather than treated as commodity labor. A useful research progression is: Observed work -> Unbundle -> Understand -> Identify Native Alpha™ -> Encode -> Allocate authority and risk -> Recompose -> Test -> Verify -> Compound ### Verification layer: Outcomes Acceptance Testing™ A worker or workflow is not successful because it executed without an error. It is successful when the intended real-world outcome occurred within defined quality, safety, timing, cost, and evidence requirements. This is especially important in healthcare. "The message was sent" is not the same as "the correct patient received the correct follow-up within the required time and the loop was closed." ## 5. Primary research questions The research should answer the following questions systematically. ### 5.1 What is the right unit of work? Investigate whether useful units are best represented as: - tasks - decisions - jobs to be done - state transitions - cases - work packets - outcomes - responsibilities - handoffs - events - obligations - exceptions Questions: - When is a task too large to automate safely? - When is a task too small to be operationally useful? - Does the right unit differ by role or industry? - Should work units be defined by action, outcome, risk boundary, authority boundary, or evidence boundary? - Can the same ontology represent both digital and physical work? - What work unit remains stable even as models and software change? ### 5.2 What makes human involvement valuable? Develop an explicit Human Capability Value model. Candidate dimensions: - uncertainty - novelty - ambiguity - interpersonal trust - emotional stakes - legal authority - clinical authority - ethical responsibility - accountability - physical presence - physical skill - tacit knowledge - cross-domain synthesis - creativity - negotiation - persuasion - exception handling - consequence severity - social legitimacy Research questions: - Which dimensions correlate with successful human retention? - Which dimensions can AI augment without replacing ownership? - Which dimensions can AI eventually perform but still require humans because society, law, insurance, or customers demand human accountability? - How much of "human required" is technical versus institutional? ### 5.3 Which work is actually meaningless? Do not define meaningless work by whether workers dislike it. Develop a classification such as: 1. Essential human work. 2. Essential control work that can be automated. 3. Essential preparatory work that can be delegated to AI. 4. Essential execution work that can become deterministic. 5. Duplicative work caused by system fragmentation. 6. Compensating work created because another process is unreliable. 7. Documentation work required only because systems cannot observe the event directly. 8. Work created by obsolete regulation or policy. 9. Work created by vendor or software limitations. 10. Work with no defensible current purpose. This classification should allow the research to distinguish: - eliminate - automate - AI-assist - delegate - standardize - consolidate - retain - redesign upstream cause ### 5.4 What does it mean to unbundle a credential? Credentials should be treated as trust and risk bundles rather than merely educational labels. A healthcare credential can encode: - body of knowledge - demonstrated competency - scope of practice - independent authority - supervision requirements - ethical duties - accountability - institutional privileges - licensure jurisdiction - continuing education - malpractice or professional liability coverage - claims history or underwriting implications - organizational indemnification - ability to supervise or delegate - ability to accept particular categories of risk A credential therefore represents more than permission to perform a task. It can represent a package of capability, legal authority, trust, accountability, and financial exposure. Research questions: - Which tasks associated with a credential legally require the credential? - Which require the credential only by institutional policy? - Which require review by a credentialed professional but not direct execution? - Which require professional accountability even if another worker performs the preparation or execution? - Which do not require credentialed involvement at all? - What liability attaches to preparation, recommendation, authorization, execution, communication, supervision, and failure to act? - How much credential exposure is consumed by each work unit? - Can preparation be separated from decision, and execution from approval, without obscuring accountability? - Can a task-level AI worker have an equivalent "credential" that specifies tested capability, data boundaries, escalation behavior, authority limits, auditability, and version? - How should human credentialing and AI worker credentialing interact? - Can organizations build machine-readable scope-of-practice and liability policies? - Which aspects of professional risk are legally indivisible even when the work is operationally decomposable? ### 5.5 What does it mean to unbundle liability? The traditional professional model often packages many different activities into one broad risk envelope: `professional + credential + scope + activities + decisions + supervision + errors + liability = one risk package` Unbundling Work creates a different research possibility: `professional -> work units -> authority requirements -> failure modes -> controls -> evidence -> accountable party -> liability exposure` The objective is not to fragment responsibility until nobody is accountable. It is the opposite: make responsibility more explicit and attach it to the smallest useful unit of work. Key research propositions: 1. Liability-bearing credentials are a discovery signal for consequential work worth examining. 2. Liability-decomposable work is the real opportunity, not high liability by itself. 3. Professional risk should be mapped separately across preparation, recommendation, authorization, execution, communication, supervision, and outcome ownership. 4. Operational Truth™, explicit worker contracts, evidence provenance, progressive authority, and Outcomes Acceptance Testing™ may make risk more observable and attributable. 5. Better decomposition may eventually support more precise underwriting, indemnification, risk sharing, or insurance products, but lower premiums should never be assumed without evidence. Research questions: - Which professional liability exposures arise from the core professional decision versus the surrounding workflow? - Can the same outcome be achieved with less total credential exposure without increasing harm? - Can task-level evidence improve attribution when something goes wrong? - Does explicit decomposition reduce ambiguous handoffs and failure-to-follow-up claims? - When does unbundling create dangerous gaps between parties instead of clearer responsibility? - How should organizational liability interact with individual professional liability? - What liability should remain with a healthcare organization when an AI worker performs preparation or execution? - What responsibility belongs to an AI vendor, model provider, system integrator, employer, supervising professional, or insurer? - Can underwriting eventually use work-unit performance, OAT results, control maturity, and incident history instead of relying primarily on broad role or organization categories? - Can risk be priced at the workflow, work-unit, worker-contract, or outcome level? - What legal, regulatory, actuarial, and contractual barriers make risk decomposition impractical? The research should explicitly study both risk transfer and risk reduction. Moving liability from an MD to a hospital, vendor, or insurer is not the same as reducing the underlying risk. ### 5.6 How should work be recomposed? Once the old role is unbundled, determine how work should be allocated. Possible worker types: - Human owner - Human executor - Human reviewer - Human supervisor - AI packet worker - AI recommendation worker - AI execution worker - Deterministic automation - Rule engine - Conventional software workflow - External service - Physical device or robot Research questions: - What allocation minimizes total cost while satisfying safety and outcome constraints? - What allocation maximizes human capability utilization? - When does supervision itself become meaningless work? - When can "review" be sampled rather than universal? - How should confidence thresholds change authority? - How should escalation work when evidence conflicts? - When does a packet worker earn execution authority? - Can authority expand automatically based on demonstrated performance, or must expansion always be explicitly approved? ### 5.7 Where is the Native Alpha™? After a role or workflow is decomposed, identify what remains that is genuinely differentiated. Questions: - Which work materially benefits from this particular person's unusual judgment, knowledge, relationships, credibility, context, or physical skill? - Which work benefits from a particular organization's data, workflow position, distribution, reputation, intellectual property, operating know-how, or trusted relationships? - Which activities appear differentiated only because they were historically bundled with a credential or job title? - Which technically automatable activities should remain human because they create trust, learning, relationship value, or strategic insight? - Can the differentiated capability be amplified by packet workers, evidence preparation, automation, delegation, or better decision surfaces? - Does the recomposed workflow increase Native Alpha™ Density, meaning a greater proportion of scarce human or organizational capacity is applied where unusual capability materially changes outcomes? - Does use of the differentiated capability create compounding advantage through data, learning, workflow position, relationships, distribution, reputation, intellectual property, or future product capability? - If the supposed advantage disappears after decomposition, was it ever Native Alpha™? Candidate principle: > Do not automate what merely looks expensive. Unbundle the work, identify what is commodity, identify what is genuinely differentiated, automate or delegate the commodity work, and concentrate Native Alpha™ where it matters. ### 5.8 Why does this work exist? Every meaningful work unit should have a work-origin hypothesis. Candidate origins: - intrinsic to producing the outcome - safety control - legal or regulatory obligation - reimbursement or payer requirement - contractual requirement - defensive documentation - fragmented systems - information asymmetry - coordination limitation - organizational habit - management reporting - vendor limitation - cost shifting - market power - obsolete policy - historical artifact Research questions: - Who requires the work? - Who benefits from the work? - Who pays for the work? - Who bears the delay or burden? - What behavior is the work trying to prevent or encourage? - Would the work disappear if the incentive, contract, reimbursement rule, system boundary, or information problem changed? - Does automation remove the cause or merely make the burden cheaper to absorb? - Could cheaper execution perversely cause more of the low-value work to be demanded? ### 5.9 What capability does performing the work create? Research whether a work unit contributes to: - skill acquisition - pattern recognition - tacit knowledge - credential progression - supervised practice - exception recognition - judgment calibration - institutional memory - professional identity - future Native Alpha™ Key question: > If AI takes this work away today, how will tomorrow's expert acquire the capability that used to emerge from doing it? The answer may be simulation, deliberate practice, sampled execution, rotation, supervised exceptions, synthetic cases, or intentionally retained routine work. ### 5.10 How can the recomposed system fail systemically or adversarially? Research: - common model dependency - common prompt or policy dependency - shared bad source data - synchronized automation errors - vendor or cloud outage - tool compromise - prompt injection - poisoned evidence - spoofed identities or events - malicious insiders - unauthorized tool use - credential theft - policy tampering - model drift - supply-chain compromise Key questions: - What is the maximum correlated blast radius? - Which dependencies can fail together? - Which failures remain invisible until many outcomes are affected? - What must continue to operate when AI is unavailable? - What evidence proves that a worker resisted adversarial input rather than merely performed well on normal cases? ### 5.11 How will the research know what caused an observed improvement? Outcomes Acceptance Testing™ establishes whether an outcome met its acceptance criteria. It does not by itself establish causality. Research designs should therefore include, where practical: - pre/post baselines - matched cohorts - phased or stepped-wedge deployment - randomized assignment where ethically and operationally appropriate - interrupted time series - difference-in-differences - case-mix adjustment - site adjustment - sensitivity analysis - negative controls Key question: > Did the recomposition cause the improvement, or did the outcome change because something else changed at the same time? ### 5.12 Who owns the recomposed capability? For every material worker or workflow, research: - background intellectual property - foreground intellectual property - employee and contractor invention obligations - affiliate rights - customer-contributed know-how - patient or customer data rights - vendor terms - model-provider rights - derived-data rights - prompt and worker-contract ownership - evaluation-set ownership - workflow and policy ownership - portability - license scope - field-of-use restrictions - confidentiality - trade-secret treatment - rights upon termination Key question: > Does encoding the capability increase the organization's control of its Native Alpha™, or make the capability easier for another party to capture? ## 6. Core concepts that should be retained and generalized ### 6.1 The Title Trap Do not design "AI versions" of professions. An "AI nurse," "AI care coordinator," or "AI pharmacist" is too vague because the title obscures scope, authority, evidence, failure behavior, and expected outcomes. General research rule: > Never begin an AI Workforce™ design with a job title. Begin with observable work. ### 6.2 Worker Contracts The worker contract should become a central research artifact. The worker-contract model identifies several important fields: - scope - inputs - ranked sources of truth - output shape - allowed actions - review gate - memory rules - failure behavior The research should expand this into a canonical Worker Contract Schema. Candidate schema: ```yaml worker_contract: id: name: version: purpose: owning_outcome: work_unit_type: work_origin: work_origin_owner: incentive_or_obligation: trigger: preconditions: inputs: evidence_requirements: source_priority: transformations: permitted_reasoning: prohibited_reasoning: output_schema: allowed_actions: prohibited_actions: authority_level: human_owner: review_gate: escalation_rules: confidence_policy: risk_class: credential_requirements: minimum_credential_exposure: liability_owner: indemnification_requirements: insurance_or_risk_instrument: jurisdiction: privacy_class: security_controls: adversarial_threats: dependency_map: common_mode_failure_modes: maximum_correlated_blast_radius: fallback_mode: graceful_degradation_policy: recovery_requirements: memory_policy: retention_policy: failure_behavior: audit_log_requirements: capability_formation_value: learning_exposure_requirement: rights_owner: background_ip_dependencies: foreground_ip_policy: input_rights: output_rights: portability_requirements: acceptance_tests: monitoring_metrics: rollback_behavior: expiry_or_recertification: ``` This is a research hypothesis, not a final technical standard. The research should determine which fields are universally necessary and which should be domain extensions. ### 6.3 Packet Workers before execution workers Healthcare organizations should generally begin with packet workers that produce structured output for review rather than workers that change live systems or contact patients directly. Research hypothesis: > The safest path to useful automation is often progressive authority, not immediate autonomy. Potential progression: Level 0: Observe only Level 1: Summarize Level 2: Prepare structured packet Level 3: Recommend action Level 4: Draft action for approval Level 5: Execute after explicit approval Level 6: Execute within narrow policy Level 7: Execute independently with retrospective audit Research questions: - Does this progression reduce risk and increase trust? - Which work should never progress beyond a certain level? - What evidence should be required before increasing authority? - How should model changes trigger recertification or reduced authority? - How should sampling and audit rates change as reliability improves? ### 6.4 AI credentials AI credentials should be treated as task-specific trust contracts rather than titles. The research should develop an AI Worker Credential artifact containing at least: - verified capabilities - benchmark or OAT results - approved data sources - data trust rules - safety and escalation behaviors - authority boundaries - auditability requirements - model and tool versions - jurisdictional constraints - recertification date - incident history - responsible human or organizational owner Key research question: > Can AI worker trust be represented as evidence about narrow capabilities rather than reputation around a broad vendor, model, or product? ### 6.5 Risk decomposition Unbundling reveals where risk actually resides. This should be retained, but the research should avoid assuming that broad "low, medium, high" labels are enough. Develop a multidimensional risk model. Candidate dimensions: - patient or customer harm severity - probability of error - reversibility - detectability - time to detection - time sensitivity - financial impact - privacy impact - legal consequence - regulatory consequence - reputational consequence - dependency on uncertain evidence - authority exercised - external communication - physical action - ability to obtain human review - blast radius - correlated failure potential - shared dependency concentration - adversarial susceptibility - recoverability - graceful degradation capability - dependency on one model, vendor, cloud, or data source Research question: > Is "risk" best treated as a scalar, a vector, or a policy rule over multiple dimensions? The research should add a liability decomposition view alongside operational risk. For each work unit, identify: - who can cause harm - who can prevent harm - who authorizes the action - who executes the action - who communicates the action - who supervises the action - who is legally accountable - who is contractually accountable - who indemnifies whom - what insurance or risk-financing instrument responds - what evidence would be available after a failure This should allow comparison between operational risk and financed risk. The party who performs the work, the party who owns the outcome, and the party whose insurance ultimately pays may be different. ### 6.6 Data readiness AI performance is downstream of the data and source systems that feed it. The research should distinguish: - data availability - data completeness - data correctness - data timeliness - provenance - source authority - conflict resolution - workflow capture quality - semantic consistency - accessibility - permissions - historical depth - event fidelity A data set is not "AI ready" in the abstract. It is ready for a specific worker contract and outcome. ### 6.7 Operational Truth™ Operational Truth™ is continuous, trustworthy, verifiable visibility into what is actually happening rather than what people, reports, or disconnected systems claim is happening. This should become foundational to the broader research program. Five concepts should be explored: 1. Queryable evidence. 2. Continuous verification. 3. Observable state. 4. Source priority and conflict handling. 5. Evidence supply chain from source event to operational claim or AI output. 6. Evidence authenticity: whether an apparently valid source event may have been spoofed, manipulated, poisoned, or produced by a compromised actor or system. The research should explicitly compare: - Reported state - Recorded state - Inferred state - Observed state - Verified state This may become one of the most important distinctions in Labor as Code™. ### 6.8 Outcomes Acceptance Testing™ Every unbundled unit of work should have a test tied to real-world outcome, not merely software execution. Candidate OAT schema: ```yaml outcomes_acceptance_test: work_unit: intended_outcome: population_or_scope: time_window: success_evidence: failure_evidence: safety_guardrails: quality_threshold: latency_threshold: escalation_threshold: false_positive_tolerance: false_negative_tolerance: human_review_requirement: source_of_truth: audit_sample: ``` Research question: > Can OAT become the equivalent of software tests for operational work? ## 7. Healthcare as the primary proving ground Healthcare should be the main research domain because it combines nearly every difficult element of human-machine work design: - credentialed professions - state licensure - institutional privileging - supervision rules - patient safety - professional liability - privacy - reimbursement - documentation requirements - fragmented source systems - physical work - emotional work - uncertain evidence - irreversible consequences - time-sensitive decisions - high coordination burden The goal should not be to produce a generic catalog of "AI use cases." The research should document what each role actually does, why the work exists, what capability or authority it consumes, what evidence it depends on, what risk it creates, what liability envelope surrounds it, and how it might be recomposed. Healthcare should also be used to test a specific discovery heuristic: credentials and roles associated with meaningful professional liability may be especially valuable places to look for poorly bundled work. The hypothesis is not that insured professionals are replaceable. It is that the existence of malpractice coverage, professional liability coverage, institutional indemnification, self-insurance, risk pools, and other instruments signals that scarce capability and consequential risk have been concentrated in a bundle that may contain separable work. ## 8. Healthcare role research template For every healthcare role studied, capture: ### Role identity - common title - credentials - licensure or certification - common care settings - typical reporting relationship - supervision requirements - typical scope of practice - major regulatory or payer constraints - malpractice, professional liability, indemnification, or other relevant risk-financing structure - common liability-sensitive activities ### Nominal purpose What is the professional supposedly there to accomplish? ### Actual work What consumes time in practice? Collect observed work rather than relying only on job descriptions. ### Work-unit decomposition For each unit: - trigger - inputs - evidence - action - output - decision - authority - risk - handoff - system of record - physical or digital - synchronous or asynchronous - patient-facing or internal - credential requirement - minimum necessary credential exposure - authority requirement - liability owner - indemnifying party where applicable - insurance or risk-financing instrument where applicable - potential failure claimant or harmed party ### Work origin and incentive map For each meaningful work unit ask: - why does this work exist - who requires it - who benefits from it - who pays for it - who bears its delay or friction - whether the cause is clinical, regulatory, contractual, reimbursement-driven, defensive, technical, organizational, or historical - whether automation removes the cause or only absorbs the burden ### Capability formation Ask whether performing the work develops: - judgment - pattern recognition - tacit knowledge - supervised competence - credential progression - exception recognition - future Native Alpha™ If the work is removed from humans, identify the replacement learning mechanism. ### Patient and caregiver labor Capture work performed by patients, families, and unpaid caregivers, including: - scheduling and rescheduling - records collection and transfer - insurance calls - symptom monitoring - medication tracking - transportation coordination - home-care execution - portal work - repeated history entry - escalation and advocacy Do not count staff work as eliminated if it is transferred to an unpaid person. ### Dependency, security, and systemic failure Capture: - common model and vendor dependencies - shared source systems - shared policies and prompts - correlated failure paths - adversarial threats - required fallback mode - maximum tolerable downtime - recovery owner ### Rights and knowledge capture Capture: - background know-how used - new know-how generated - worker-contract and prompt ownership - data and derived-data rights - evaluation-set ownership - portability requirements - third-party or affiliate rights - any Native Alpha™ at risk of being transferred or diluted ### Human capability requirement Why, if at all, does a human materially improve this work? ### Native Alpha™ assessment For the role and each meaningful work unit, ask: - What capability is common to anyone holding the credential? - What capability is unusual to this particular person, team, organization, or workflow position? - Does that unusual capability measurably change the outcome? - Is the apparent differentiation actually commodity work hidden inside a prestigious role? - Would automation destroy trust, learning, relationship value, reputation, proprietary insight, or another compounding advantage? - Can surrounding work be removed so the differentiated capability can be applied to more cases or harder cases? - What evidence would prove that Native Alpha™ exists rather than merely assert it? ### Candidate disposition - retain with human - AI assist - packet worker - recommendation worker - deterministic automation - AI execution under policy - delegate to different human role - eliminate - redesign upstream process ### OAT How will we know the recomposed work is better? ### Failure modes What could go wrong? ### Economic and risk effect - labor minutes saved - credentialed minutes saved - credential exposure reduced or concentrated - professional liability exposure changed - organizational liability exposure changed - risk transferred versus risk actually reduced - evidence quality for claims, audits, or incident review improved - throughput increased - delay reduced - errors reduced - revenue protected - clinician capacity released - patient access increased - Native Alpha™ Density increased - differentiated professional capacity applied to more or harder cases - commodity work removed from differentiated people - trust, learning, relationships, data, workflow position, or intellectual property strengthened - outsourcing avoided - new service made feasible ## 9. Healthcare role deep-dive agenda ### 9.1 Physicians: MD and DO Central question: > How much physician time and physician credential exposure are consumed by work that does not require physician-level judgment, authority, accountability, trust, or malpractice-bearing responsibility? Research the physician workday as a bundle of: - history acquisition - chart review - information synthesis - differential diagnosis - diagnosis - treatment planning - order preparation - order authorization - documentation - coding support - inbox management - result review - medication refill review - prior authorization - referral coordination - patient education - shared decision-making - family communication - care-team communication - quality-measure documentation - administrative forms - insurance documentation - peer-to-peer payer review - supervision - teaching - unusual exception handling Hypotheses to test: - Much of pre-encounter and post-encounter work can become structured packet work. - Physician review time may be a more useful optimization target than physician documentation time. - Some "documentation burden" is actually evidence-reconstruction burden caused by fragmented systems. - The highest-value AI worker may not be a scribe. It may be a longitudinal evidence and exception worker that prepares the physician decision surface. - Some work should move to another human role rather than to AI. - The economic value of released physician minutes should be measured against access, patient complexity handled, and reduced after-hours work, not only headcount. - Physician malpractice exposure may be unnecessarily attached to workflow steps that require physician approval but not physician preparation or execution. - Separating evidence preparation from medical decision-making may improve both throughput and post-event traceability. - Reducing physician touches is not enough; the research should measure whether physician liability exposure is reduced, merely shifted to the organization, or unchanged. - The safest recomposition may deliberately retain physician accountability while reducing the amount of physician attention required to exercise it. - Physician credentials define a permitted scope, but they do not reveal each physician's Native Alpha™. One physician may have unusual diagnostic pattern recognition, another exceptional patient trust, another rare-disease expertise, another superior procedural skill, and another unusually strong local referral knowledge. - Unbundling should make those differences visible rather than treating all physician minutes as interchangeable. - Some patient-facing work should remain physician-owned even when technically automatable if the interaction is where trust, adherence, unusual insight, reputation, or referral relationships compound. - The strongest recomposition may make the physician more important to the accepted outcome while requiring fewer physician minutes overall. Potential workers: - pre-visit evidence packet worker - care-gap worker - medication change summarizer - result triage packet worker - referral evidence worker - prior-authorization evidence worker - longitudinal patient narrative worker - coding evidence worker - documentation completeness worker - after-visit instruction drafter Human ownership likely remains strongest around diagnosis, treatment decisions, informed consent, complex risk communication, unusual exceptions, and accountability. Research must verify legal and institutional boundaries rather than assume them. ### 9.2 Registered Nurses: RN Central question: > Which parts of nursing require nursing judgment, licensed authority, accountability, presence, or liability-bearing responsibility, and which parts merely consume nursing capacity because information and coordination systems are poorly designed? Work bundle: - assessment - triage - medication administration - patient monitoring - care-plan execution - patient education - discharge preparation - documentation - care coordination - escalation - inbox or message management - referral follow-up - telephone outreach - protocol-driven tasks - handoff preparation - quality documentation - supply and logistics coordination Research focus: - distinguish assessment from data collection - distinguish triage judgment from triage packet preparation - distinguish patient education from content generation - distinguish medication administration from medication reconciliation preparation - identify where physical presence is essential - identify where relational continuity is clinically valuable - examine whether AI reduces cognitive switching and coordination load Potential workers: - triage packet worker - discharge follow-up worker - care-plan status worker - abnormal-result routing worker - medication reconciliation packet worker - patient education preparation worker - escalation surveillance worker - handoff packet worker ### 9.3 Nurse Practitioners and Physician Assistants Central question: > How does AI change delegation, supervision, and panel management when advanced practice clinicians themselves supervise distributed AI capabilities? Research: - tasks independently owned versus physician-supervised - scope differences by jurisdiction - pre-visit preparation - protocolized chronic care - inbox and follow-up burden - supervision interactions - escalation thresholds - population health work - patient education - care-gap closure The research should avoid treating NP and PA work as identical. Scope and supervision rules differ across jurisdictions and institutions. ### 9.4 Medical Assistants and Clinical Support Staff Central question: > Which high-volume coordination and preparation tasks can be standardized or automated so that MAs spend more time on patient-facing support and reliable clinical preparation? Work bundle: - rooming - vitals - intake - chart preparation - refill routing - lab tracking - referral support - forms - patient messaging - prior authorization - scheduling - preventive care outreach - supplies This role is especially important because automating MA tasks can accidentally shift work upward to nurses and physicians. Research must measure work transfer, not just local task elimination. ### 9.5 Pharmacists and Pharmacy Technicians Central question: > How can medication work be decomposed between evidence preparation, pattern detection, professional judgment, counseling, dispensing, and prescribing collaboration? Work bundle: - medication reconciliation - interaction checking - dosing review - formulary checking - refill processing - medication therapy management - patient counseling - prior authorization - adherence analysis - protocol review - inventory and dispensing controls - prescriber communication Research should distinguish between: - information retrieval - computational checking - recommendation - professional verification - patient counseling - authority to modify therapy Potential workers: - medication evidence packet worker - interaction evidence worker - formulary alternative worker - adherence-risk worker - prior-auth packet worker ### 9.6 Referral Coordinators This is a strong reference role because the work crosses systems and organizations. Work bundle: - intake - completeness checking - clinical triage support - insurance requirements - authorization - record gathering - specialist selection - scheduling - patient communication - status tracking - escalation - loop closure Research questions: - What is the true state machine of a referral? - Which events prove state changes? - What does "completed referral" actually mean? - Where do source systems disagree? - Which tasks are pure coordination taxes created by fragmented systems? - How much work disappears if Operational Truth™ becomes reliable? Potential visual: A referral should be shown as a state machine with evidence attached to every transition. Overlay current human touches in one layer, then a recomposed version showing human versus AI ownership. ### 9.7 Care Managers and Care Coordinators Work bundle: - patient segmentation - outreach - assessment - care plan management - goal tracking - social-needs coordination - referral follow-up - medication support - escalation - documentation - team communication Research focus: - relationship value versus coordination work - longitudinal context - chronic-condition knowledge - barriers and social context - persistent versus temporary memory - escalation design - trust continuity Potential hypothesis: The most valuable use of AI may be to remove tracking and preparation burden so care managers spend more time in actual human engagement. ### 9.8 Schedulers and Front Desk Work bundle: - appointment scheduling - rescheduling - cancellation handling - slot optimization - patient identification - insurance collection - eligibility verification - intake forms - reminders - routing - basic questions - payment collection - exception handling Research questions: - Which work is deterministic? - Which work requires negotiation or empathy? - Which scheduling policies are encoded nowhere except staff experience? - Can cancellation, waitlist, overbooking, and capacity rules be represented as code? - What exceptions still require human judgment? ### 9.9 Prior Authorization Staff Prior authorization may be a canonical example of labor that exists largely because of institutional friction. Research: - evidence gathering - payer rule identification - form completion - submission - status checking - missing-data resolution - appeal preparation - peer-to-peer preparation - patient communication - clinician interruption Questions: - What work can disappear through direct structured evidence exchange? - What work should AI prepare but humans own? - What payer rules can be represented as executable policy? - How should changing payer policies be monitored and versioned? - Can outcome testing measure approval rate, cycle time, clinician minutes consumed, and avoidable denials? ### 9.10 Coders, Billers, Revenue Cycle Staff Work bundle: - charge capture - coding - documentation review - claim preparation - claim submission - eligibility - denial management - payment posting - underpayment detection - appeals - patient billing - reconciliation - revenue reporting Research questions: - Which work exists because clinical documentation and billing evidence are separated? - Can coding move from retrospective interpretation toward evidence-linked suggestion? - Where must certified coding expertise remain? - What is the risk of optimizing for reimbursement at the expense of clinical truth? - Can AI produce an inspectable evidence chain for every code or claim decision? ### 9.11 Health Information Management Research: - release of information - record integrity - patient matching - coding governance - retention - amendment - privacy - audit - terminology - data quality This role can test the proposition that "data preparation" is operational work rather than a one-time analytics project. ### 9.12 Quality, Compliance, Privacy, and Risk Work bundle: - policy interpretation - evidence collection - audit preparation - monitoring - training - incident analysis - corrective actions - reporting - control testing - regulatory surveillance Questions: - Can continuous evidence replace periodic evidence hunts? - Which compliance work exists only because operational truth is weak? - Can AI map control obligations directly to source evidence? - How should policy changes update worker contracts? - Can an AI workforce make compliance more continuous rather than more bureaucratic? ### 9.13 Behavioral Health Professionals Roles may include psychiatrists, psychologists, LCSWs, LMHCs, LPCs, and others. Research should be cautious about assuming that relational work is reducible to information processing. Study: - intake - screening - scoring - documentation - treatment planning - resource matching - crisis escalation - follow-up - scheduling - patient communication - clinical relationship Questions: - Which administrative burden can be removed without degrading therapeutic alliance? - Where does patient preference for human interaction matter independently of measurable clinical outcome? - What crisis situations should prohibit autonomous execution? - What forms of AI preparation improve clinician presence rather than intrude on it? ### 9.14 Therapists: PT, OT, SLP Work bundle: - evaluation - treatment planning - exercise or intervention selection - demonstration - observation - coaching - progress measurement - documentation - scheduling - home-program education - authorization Research questions: - Which physical observations cannot be inferred reliably from digital data? - Where can computer vision or sensors augment measurement? - Can preparation and documentation be removed while preserving clinician-patient time? - Which exercise recommendations require professional review? ### 9.15 Laboratory and Imaging Roles Roles: - medical laboratory scientists - lab technicians - radiologic technologists - sonographers - imaging support - pathologists and radiologists as distinct higher-order interpretive roles Research: - acquisition - quality control - routing - abnormal flagging - result communication - interpretation support - documentation - scheduling - protocol selection This area is useful for separating physical acquisition work from interpretation and from communication. ### 9.16 Practice Managers and Healthcare Administrators Work bundle: - staffing - schedule management - productivity - revenue monitoring - vendor management - compliance - incident follow-up - patient complaints - supply issues - hiring - reporting - meeting coordination - project management Research hypothesis: A large amount of management labor may be spent reconstructing operational state because systems do not expose Operational Truth™. Potential workers: - exception digest worker - staffing risk worker - revenue leakage worker - unresolved issue worker - operational evidence worker - meeting-to-obligation worker ### 9.17 Clinical Research Operations Roles: - investigators - clinical research coordinators - data managers - monitors - regulatory staff - medical writers - biostatistical support Work bundle: - protocol interpretation - participant screening - consent support - scheduling - source documentation - data entry - query resolution - adverse event tracking - regulatory binders - monitoring - site communication Questions: - Which work exists to reconcile duplicate representations of the same event? - Can source-to-case-report evidence chains become continuously verifiable? - Which parts require investigator judgment versus coordinator preparation? - How should AI worker credentials operate under GCP and research oversight? ### 9.18 Medical Device, Pharma, and Regulatory Roles Healthcare should extend beyond care delivery. Roles may include: - regulatory affairs - quality assurance - quality engineering - clinical affairs - pharmacovigilance - medical affairs - safety - validation - verification - complaint handling - CAPA - document control Research questions: - Can evidence gathering, traceability, document assembly, and change-impact analysis become AI workforce functions? - Which sign-offs represent actual authority and accountability? - Which reviews are meaningful versus ritual? - Can Labor as Code™ represent regulated work instructions and required evidence? - Can OAT coexist with formal validation and quality-system requirements? ## 10. Healthcare credential and liability decomposition agenda Healthcare should be the primary laboratory for studying the relationship between credentials, work decomposition, and professional liability. The central research question is: > Can healthcare work be recomposed so that each accepted outcome consumes only the professional capability, authority, and liability exposure that it actually requires? ### 10.1 Liability-bearing credentials as a discovery signal Build a healthcare map of roles in which credentials are associated with meaningful professional liability, institutional indemnification, supervision duties, or other risk-financing mechanisms. Candidate roles include: - physicians and surgeons - dentists - registered nurses and advanced practice nurses - physician assistants - pharmacists - psychologists and behavioral health professionals - physical, occupational, and speech therapists - radiologists and imaging professionals - pathologists and laboratory directors - clinical researchers and principal investigators - medical directors - healthcare executives with fiduciary or regulatory accountability - biomedical and clinical engineers where professional liability attaches For each role, distinguish: - professional liability borne individually - liability borne by the employer or health system - vicarious or supervisory liability - contractual indemnification - self-insured retention or risk-pool participation - vendor or technology liability - payer-related financial risk - regulatory penalties that are not insurance claims but still create financial exposure The point is not to build an insurance catalog. The point is to identify where consequential risk has been bundled around a credential and then investigate whether the work bundle matches the risk bundle. ### 10.2 Credential exposure per outcome Develop a metric for the amount of professional credential exposure consumed by an outcome. Candidate components: - credentialed minutes - number of credentialed touches - number of credentialed decisions - number of credentialed authorizations - number of supervised actions - number of legally consequential communications - time under active professional responsibility - severity-weighted authority exercised Possible metric: `credential exposure per accepted outcome` Examples: - physician exposure per completed visit - RN exposure per safely resolved triage episode - pharmacist exposure per completed medication reconciliation - radiologist exposure per finalized imaging episode - principal-investigator exposure per compliant research decision This metric should never be optimized without safety and outcome constraints. A lower number is useful only when the accepted outcome remains clinically, legally, ethically, and operationally sound. ### 10.3 Minimum necessary credential exposure For each work unit, ask in order: 1. Does this require professional knowledge? 2. Does it require professional judgment? 3. Does it require professional authority? 4. Does it require professional accountability? 5. Does it require the credentialed human to execute the work directly? 6. Could a machine or another human prepare the evidence? 7. Could execution be separated from authorization? 8. Could review be exception-based rather than universal? 9. What evidence is needed to prove that the professional exercised appropriate oversight? 10. What is the minimum credential exposure consistent with a safe accepted outcome? ### 10.4 Liability decomposition by work stage Study liability separately across: - information acquisition - evidence preparation - interpretation - recommendation - authorization - execution - communication - monitoring - escalation - supervision - documentation - loop closure The same credential may be required at some stages but not others. The research should determine whether legal and insurance structures recognize these distinctions or continue to treat the entire episode as a single professional act. ### 10.5 Risk transfer versus risk reduction Every recomposition proposal should explicitly answer: - Did underlying probability of harm decrease? - Did severity decrease? - Did detectability improve? - Did time to detection improve? - Did attribution improve? - Did the professional's individual exposure decrease? - Did the organization's exposure increase? - Did liability move to a vendor or technology provider? - Did contractual indemnification change who pays without changing who controls the risk? - Did the workflow create a new uninsured or ambiguous gap? This distinction is critical. A workflow is not safer merely because liability moved off the clinician. ### 10.6 Insurance and risk-financing hypotheses The research should investigate, without presuming, whether more granular evidence about work can change how risk is financed. Possible future mechanisms include: - task- or workflow-specific professional liability endorsements - lower premiums or different deductibles for demonstrably controlled workflows - organization-level coverage that explicitly includes approved AI workers - vendor indemnification linked to worker-contract scope - performance warranties - shared-risk arrangements - captive or self-insurance models informed by work-unit evidence - cyber, technology E&O, product liability, and malpractice coverage coordination - coverage conditioned on OAT performance, evidence retention, or auditability These are research hypotheses, not recommendations. The research must involve qualified insurance, legal, actuarial, and clinical experts before drawing conclusions about insurability or premium effects. ### 10.7 Liability Density as an opportunity signal Explore a heuristic for identifying high-value unbundling targets. Candidate model: `Unbundling Opportunity = Professional Scarcity × Liability Density × Work Decomposability × Observability × Controllability` Where: - Professional Scarcity estimates the cost and availability of the human capability. - Liability Density estimates how much legal or financial consequence is concentrated around the role or workflow. - Work Decomposability estimates whether meaningful stages can be separated. - Observability estimates whether Operational Truth™ can establish what happened. - Controllability estimates whether worker contracts, policy, review gates, and Outcomes Acceptance Testing™ can constrain the work. The heuristic should be treated as a research instrument, not a mathematical truth. ### 10.8 Cross-role comparison study Compare workflows that vary across the dimensions above. Examples: - surgeon performing a technically difficult procedure: high liability, lower decomposability of the core physical act - physician reconstructing a longitudinal record before a treatment decision: high professional cost, high liability, high decomposability of preparation - RN collecting routine information before triage: moderate-to-high credential exposure, potentially high decomposability - pharmacist reviewing a machine-prepared medication reconciliation: high-value professional review with potentially low preparation burden - prior-authorization staff gathering evidence for physician attestation: high administrative burden with a narrow professional authorization point The research should seek workflows where the professional's scarce capability is concentrated into a small number of consequential decisions while surrounding work can be safely separated. ## 11. Cross-role healthcare research questions The role studies should feed several larger questions. ### 11.1 Work transfer When a task is automated in one role, where does the remaining work go? A local efficiency gain may create work elsewhere. Measure: - upward delegation - downward delegation - patient self-service burden - exception volume - reviewer burden - IT support burden - compliance burden ### 11.2 Review burden AI assistance can create a new class of meaningless work: reviewing machine output that is usually correct but cannot safely be ignored. Research: - universal review - risk-based review - confidence-based review - exception-only review - sample audit - retrospective audit Key question: > When does human-in-the-loop become human-as-rubber-stamp? ### 11.3 Coordination tax How much healthcare labor exists because systems, organizations, payers, and professionals cannot reliably share state? Examples: - referral status calls - fax follow-up - missing record pursuit - repeated patient history collection - authorization status checks - scheduling callbacks - duplicate documentation This may be one of the largest categories of meaningless work. ### 11.4 Evidence reconstruction How much professional time is spent reconstructing the evidence necessary to make a decision? Potential examples: - chart review - medication history reconstruction - prior authorization packet preparation - referral completeness review - longitudinal timeline creation - audit preparation Hypothesis: > AIgreatest near-term value in consequential work may be shrinking the evidence-reconstruction surface before a human decision. ### 11.5 Exception economics Automation often handles the common case but leaves humans with only difficult exceptions. Research whether this: - raises human skill requirements - increases cognitive intensity - reduces opportunities for junior staff to learn - changes staffing ratios - creates fatigue from continuous exception handling - requires new training or job design This is a potential counterweight to the "top of capability" thesis. ### 11.6 Patient and caregiver labor Healthcare contains a large invisible workforce outside payroll. Patients and caregivers routinely perform: - appointment coordination - transportation planning - medication administration and tracking - home monitoring - symptom interpretation and escalation - insurance navigation - records retrieval - form completion - portal management - specialist coordination - post-discharge care Research should measure unpaid minutes per accepted outcome and identify which institutional burdens have been exported to families. Key principle: > Work is not eliminated when it disappears from payroll and reappears in the patient's kitchen. ### 11.7 Incentive-created work Map work generated by payer-provider conflict, reimbursement design, defensive practice, regulatory interpretation, vendor contracting, departmental incentives, and other institutional structures. Key questions: - Which work would vanish if the underlying incentive changed? - Which party has little incentive to eliminate the burden because another party bears the cost? - Does AI reduce friction in a way that entrenches the underlying low-value process? - Could lower execution cost cause a payer, regulator, or organization to demand more of the work? ### 11.8 Capability formation and apprenticeship Study how professionals acquire expertise when AI removes routine exposure. Measure: - supervised cases completed - variety of cases encountered - time to independent competence - error-recognition ability - judgment calibration - exposure to rare but important exceptions - confidence versus actual competence Research whether deliberate simulation, shadow execution, sampled human performance, or structured rotations can replace learning previously obtained through routine work. ### 11.9 Common-mode healthcare failure Map failures that can affect many patients simultaneously because the same machine worker, source data, policy, or vendor is reused at scale. Examples: - a bad formulary rule propagated across every prior authorization - an incorrect clinical threshold used by every triage worker - a stale source system feeding every packet worker - a compromised integration altering thousands of records - a model update changing behavior across an entire health system Healthcare safety research should compare ordinary case-level error with correlated population-level failure. ## 12. Labor as Code™ research program Labor as Code™ should be treated as a serious systems discipline, not a metaphor. ### 12.1 Research objective Determine whether operational work can be represented with enough precision that it can be: - inspected - versioned - assigned - simulated - tested - audited - executed - improved - governed - compared ### 12.2 Candidate primitives Research a minimal set of primitives: - Worker - Work Unit - Trigger - State - Evidence - Input - Output - Decision - Action - Authority - Credential - Policy - Risk - Review Gate - Escalation - Memory - Handoff - Outcome - Acceptance Test - Incident - Incentive - Obligation - Dependency - Threat - Fallback Mode - Learning Requirement - Rights - Version ### 12.3 Work Unit schema ```yaml work_unit: id: name: purpose: outcome: work_origin: incentive_or_obligation: trigger: preconditions: required_evidence: inputs: source_priority: decision_type: actions: outputs: authority_required: credentials_required: risk_profile: human_capability_required: handoff_from: handoff_to: systems_touched: dependencies: common_mode_failure_modes: adversarial_threats: fallback_mode: physical_presence_required: time_constraint: failure_modes: escalation: capability_formation_value: unpaid_patient_or_caregiver_work: rights_profile: acceptance_test: ``` ### 12.4 Work graph Research whether a workflow is better represented as a graph than a sequence. Nodes: - work units - decisions - states - evidence events Edges: - handoffs - dependencies - escalation - state transition - evidence provenance The graph should allow analysis of: - bottlenecks - unnecessary human touches - duplicated work - fragile handoffs - unclear ownership - missing evidence - high-risk concentration - credential bottlenecks - latency - automation candidates ### 12.5 Git as a model for work versioning Explore whether operational work can borrow useful concepts from software configuration management: - version - diff - pull request - review - approval - release - rollback - test - audit history Do not assume the analogy is universally useful. Research question: > Which aspects of work actually benefit from software-like version control, and which become bureaucratic if represented too rigidly? ## 13. AI Workforce™ research program ### 13.1 Definition An AI Workforce™ is the intentionally designed system by which human and machine workers jointly produce real-world outcomes. It is not synonymous with AI agents. ### 13.2 Workforce composition model For every process, represent: - human workers - AI workers - deterministic automation - software applications - source systems - data stores - external parties - physical devices - governance roles ### 13.3 Human roles in an AI Workforce™ Humans may act as: - originator - subject-matter expert - operator - reviewer - supervisor - accountable owner - exception handler - approver - auditor - trainer - policy owner - incident responder - recipient of output ### 13.4 Machine roles Machines may act as: - observer - retriever - summarizer - classifier - packet preparer - recommender - drafter - router - scheduler - monitor - executor - verifier - auditor - simulator ### 13.5 Management of the AI Workforce™ Research how a manager should supervise machine workers. Possible management artifacts: - worker roster - capability matrix - authority matrix - incident history - current versions - OAT performance - escalation rate - human override rate - drift indicators - cost per outcome - latency per outcome - evidence completeness - recertification status - dependency concentration - correlated-failure exposure - adversarial-test status - fallback readiness - recovery-test status - capability-formation impact - rights and portability status Key question: > Is managing AI workers closer to managing people, software services, regulated devices, or some hybrid of all three? ## 14. Operational Truth™ research program ### 14.1 Core question Can an organization continuously know what is actually happening well enough to safely delegate work to machines? ### 14.2 State taxonomy For each important object, distinguish: - Intended state - Reported state - Recorded state - Inferred state - Observed state - Verified state Objects may include: - patient - referral - appointment - authorization - claim - care plan - medication - result - task - worker - incident ### 14.3 Evidence supply chain Every operational claim should be traceable: Claim -> derived evidence -> source event -> source system -> actor/device -> timestamp Research: - authoritative timestamp - conflict handling - missing-event detection - stale evidence - source reliability - data lineage - human-entered versus machine-observed evidence ### 14.4 Operational Truth™ as labor reduction Hypothesis: A meaningful percentage of administrative labor exists because organizations cannot trust their own operational state. Examples: - calling to confirm status - checking multiple screens - manually reconciling lists - creating spreadsheets - holding status meetings - emailing for updates - duplicating documentation If Operational Truth™ is reliable, this work may disappear rather than merely be automated. ## 15. Data readiness research program Do not create a generic "clean your data" workstream. Tie every data requirement to a worker contract. For each worker ask: - What evidence does it need? - Which source is authoritative? - What freshness is required? - What completeness is required? - What conflicts are tolerable? - What missingness forces escalation? - What permissions apply? - What provenance must be preserved? - What data should never enter model context? - What data may be retained as memory? Develop a Worker Data Readiness Score rather than an enterprise-wide AI readiness score. Potential dimensions: - availability - completeness - correctness - timeliness - authority - conflict rate - provenance - accessibility - semantic clarity - workflow alignment ## 16. Credential, authority, liability, and accountability research ### 16.1 Human credential model Map: Credential -> capability -> legal scope -> institutional privilege -> supervision -> authority -> accountability -> liability -> risk-financing instrument The model should distinguish what the credential permits, what the organization permits, what the professional personally owns, what the organization owns, and what an insurer or other risk-financing arrangement covers. ### 16.2 Liability-bearing credential heuristic Treat meaningful liability instruments as a research signal. A credential associated with malpractice, professional liability, errors-and-omissions coverage, bonding, indemnification, or similar financial protection may identify a role in which consequential authority has been bundled together with surrounding work. This does not establish that the work is automatable. Score each candidate role on: - scarcity of professional capability - liability density - work decomposability - evidence observability - controllability - degree of physical or tacit work - regulatory indivisibility - customer or patient expectation of direct human presence ### 16.3 AI credential model Map: Worker version -> tested capability -> approved evidence -> allowed action -> review gate -> escalation -> OAT performance -> auditability -> accountable owner ### 16.4 Authority is not capability A worker may be technically capable of an action but not authorized to perform it. This distinction must be explicit. Example: An AI worker may accurately recommend a medication change but have no authority to prescribe, enter, or communicate that change independently. ### 16.5 Accountability cannot be hand-waved Every consequential worker contract should identify: - operational owner - clinical or professional owner where applicable - technology owner - policy owner - incident owner - liability owner - indemnifying party - insurance or risk-financing mechanism where relevant Research whether a single accountable human must exist for every machine worker or whether organizational accountability can sometimes be sufficient. Also research where responsibility must remain indivisible even when execution is decomposed. A design that creates ambiguity about who owns the failure should be considered worse, not better. ## 17. Risk, liability, and progressive authority Current credentialing, licensure, and insurance systems were largely designed around human roles and organizations. Unbundling can expose where operational risk, legal accountability, and financial liability actually sit. The research should go further by distinguishing risk creation, risk control, risk ownership, risk transfer, and risk financing. ### 17.1 Progressive authority model Authority should be earned through evidence. Possible factors: - OAT pass rate - severity-weighted error rate - confidence calibration - conflict handling - escalation appropriateness - performance across populations - performance across sites - version stability - incident history - reviewer agreement - override rate ### 17.2 Automatic reduction of authority Research whether workers should lose authority automatically when: - source data quality drops - a model or tool version changes - OAT performance degrades - a new patient population is introduced - workflow conditions change - policy changes - incident severity threshold is crossed ### 17.3 Blast radius A highly accurate worker may still be dangerous if one error affects thousands of cases. Risk should include scale. ### 17.4 Credential exposure and liability surface Every workflow should estimate its liability surface, including: - number and type of credentialed decisions - number of credentialed touches - severity of authority exercised - number of handoffs between accountable parties - ambiguity of responsibility - exposure to failure-to-act or failure-to-follow-up risk - reliance on machine-generated evidence - ability to reconstruct who knew what and when - availability of contemporaneous evidence - number of parties with indemnification obligations Research whether decomposed workflows reduce or increase this surface. ### 17.5 Risk packaging Explore whether explicitly decomposed work can support more appropriate packaging of risk. Possible patterns: - human professional retains final decision liability while machine workers carry no independent authority - organization assumes workflow liability while professionals operate within approved decision boundaries - vendor assumes specified technology-performance obligations through contract or insurance - shared-risk structures allocate responsibility according to worker contract and failure mode - certain low-risk work units are removed entirely from the professional liability envelope The research should treat these as legal and commercial design questions requiring counsel and insurance expertise. It should not infer enforceability, coverage, or insurability from technical architecture. ### 17.6 Common-mode and dependency concentration risk For every material worker, determine: - models shared with other workers - shared prompts, policies, and code - source systems shared across workflows - common integration points - cloud and vendor concentration - common credentials or secrets - shared human approvers - shared upstream data transformations Develop a dependency-concentration view that answers: > If this dependency fails or behaves incorrectly, how many outcomes can be affected before detection and containment? ### 17.7 Adversarial security model The AI Workforce™ should assume that some inputs are actively hostile rather than merely erroneous. Threats to research include: - prompt injection in records or documents - malicious instructions hidden in external content - poisoned source data - fraudulent or spoofed identities - compromised accounts - manipulated clinical or operational evidence - malicious insiders - tool misuse - excessive permissions - data exfiltration - policy or prompt tampering - model or dependency supply-chain compromise AI worker credentials should eventually include evidence that the worker has been tested against relevant adversarial conditions. ### 17.8 Graceful degradation and continuity A mature AI Workforce™ must answer: > What happens when the AI is unavailable, untrusted, degraded, or disconnected? Research: - manual fallback procedures - reduced-authority operating modes - local caching of critical policies and evidence - priority queues during outages - human staffing required for degraded operation - maximum tolerable downtime - recovery order - reconciliation after service restoration - continuity drills - dependency substitution A system that is economically attractive only when every AI dependency is healthy may be too brittle for consequential work. ## 18. Outcomes Acceptance Testing™ research OAT should be a major pillar of the research because it converts AI discussions from demos to evidence. ### 18.1 Test the outcome, not the artifact Examples: Weak: - Note generated. - Referral message sent. - Prior authorization submitted. - Patient classified. Better: - Clinician accepted the note with no material correction and required evidence was present. - Referral reached the correct specialist and the patient was scheduled within the defined window. - Prior authorization was approved or correctly escalated with no avoidable missing evidence. - High-risk patient was identified within the required time with acceptable false-negative rate. ### 18.2 Acceptance test layers 1. Unit-level work test 2. Handoff test 3. End-to-end workflow test 4. Safety test 5. Outcome test 6. Economic test 7. Human experience test 8. Equity or population performance test where relevant 9. Security and adversarial test 10. Resilience and degraded-mode test ### 18.3 OAT is necessary but not causal proof Outcomes Acceptance Testing™ answers whether the accepted outcome occurred under defined conditions. Research studies must separately determine whether the recomposed workflow caused any observed improvement. ### 18.4 Continuous testing A worker that passed once is not permanently safe. Research: - continuous sampling - sentinel cases - synthetic cases - real-world audit - drift detection - version-triggered recertification - adverse-event feedback ## 19. Economic research The research should avoid vague "productivity" claims. Measure specific economics. ### 19.1 Human capability minutes and credential exposure For each process measure: - total human minutes - credentialed human minutes - credentialed touches - credentialed decisions - credentialed authorizations - severity-weighted authority exercised - credential exposure per accepted outcome - high-scarcity human minutes - review minutes - exception minutes - coordination minutes - documentation minutes - evidence-reconstruction minutes ### 19.2 Cost per accepted outcome Possible denominator: - completed referral - correctly closed care gap - approved authorization - completed visit - resolved patient message - clean claim - resolved denial - successful transition of care ### 19.3 Capacity released Measure what people actually do with released time. If ten physician hours are "saved" but appointment capacity does not increase, burnout does not fall, quality does not improve, and no other useful work occurs, the economic value may be overstated. ### 19.4 New economically feasible work Research work that was previously not done because human preparation cost exceeded the value. Examples: - longitudinal evidence review for every complex patient - proactive care-gap surveillance - continuous denial pattern analysis - personalized patient education - continuous compliance evidence gathering - small-practice quality analytics This may be more strategically important than simple cost takeout. ### 19.5 Liability and risk-financing economics Track separately: - malpractice or professional liability claims frequency - claims severity - incident frequency and severity - deductible or self-insured-retention utilization - professional liability premium changes where observable - organizational risk-financing costs - indemnification costs - vendor insurance requirements - legal defense cost - time spent on incident reconstruction - evidence completeness after adverse events The research should avoid attributing insurance changes to AI without controlling for specialty, jurisdiction, claims history, organization size, coverage terms, and broader market conditions. The central question is whether decomposition changes the underlying risk and the quality of evidence available to underwrite or defend that risk. ### 19.6 Transition economics Track the cost of moving from the bundled workflow to the recomposed workflow: - duplicate operation during shadow mode - migration engineering - integration work - retraining - temporary productivity loss - change-management effort - policy and legal review - worker recertification - data remediation - rollback and incident cost - contract changes - legacy-system retirement A theoretically superior future-state workflow may be commercially inferior when transition cost, risk, or time is included. ### 19.7 Unpaid labor economics Include patient and caregiver time in the economic model where relevant. Candidate measures: - unpaid minutes per accepted outcome - number of calls, forms, portals, and handoffs required of the patient or caregiver - transportation and coordination burden - uncompensated home monitoring or administrative work - workdays or wages lost because of avoidable healthcare coordination This does not require assigning a simplistic dollar value to every family interaction. It requires refusing to treat unpaid labor as free. ## 20. Human experience research Zero Meaningless Work should be falsifiable. Measure whether workers experience: - less administrative burden - fewer interruptions - less context switching - more time in professional work - more autonomy - more meaningful patient/customer interaction - more cognitive intensity - more exception fatigue - loss of learning opportunities - increased surveillance - reduced trust - reduced professional identity - reduced learning opportunities - slower competency development - loss of mentoring or apprenticeship - overreliance on machine-prepared evidence A system that eliminates clerical work but turns every clinician into a high-stakes exception handler may not be an improvement. A system that improves today's throughput while weakening tomorrow's professional capability may also be a bad trade. ## 21. Patient, caregiver, and customer experience and labor research Healthcare workflow optimization should not make patients perform the work instead. Measure both experience and labor: - unpaid patient or caregiver minutes per outcome - number of contacts required - repeated information requests - time to resolution - access - continuity - clarity - trust - ability to reach a human - burden placed on caregivers - forms, portals, and records work performed outside the organization - travel and coordination burden - ability to delegate to a family member or advocate safely - failure recovery A useful rule to test: > Do not call work "eliminated" if it was merely pushed onto the patient. ## 22. Organizational unbundling The research should eventually move beyond roles. Departments are also bundles. Examples: - revenue cycle - quality - compliance - care management - referral management - front desk - utilization management - IT support Research questions: - Which department boundaries exist because coordination used to be expensive? - Which could become capability pools rather than fixed departments? - Which management layers primarily aggregate and report status? - Which outsourcing contracts are bundles that can now be decomposed? - Which vendor categories exist because software had to package a complete workflow rather than expose narrow capabilities? - Which departments exist partly because reimbursement, contracting, compliance, or budget incentives reward the boundary itself? - Which manager or department benefits from preserving work that another group pays to perform? - Which activities should disappear through incentive redesign rather than become more efficiently automated? This connects naturally to Zero Neo without making software elimination the primary thesis. ## 23. Software unbundling Many SaaS categories bundle: - database - forms - workflow - rules - reporting - messaging - permissions - integration - domain-specific configuration Research whether AI plus existing Microsoft 365, Google Workspace, EHR, ERP, CRM, and other systems can recompose these capabilities without another standalone application. Questions: - When is new software actually required? - Which category-specific applications mainly encode work that could be represented directly as Labor as Code™? - Where does domain software provide indispensable safety, transaction integrity, regulatory evidence, or network connectivity? - When does "Zero Neo" become reckless? ## 24. Physical work The research should not become digital-only. Many real-world jobs contain: - physical movement - observation - manipulation - environment-specific judgment - face-to-face interaction - safety checks Research the boundary among: - AI planning - human execution - robotic execution - sensor observation - computer vision - human verification Healthcare examples: - patient transport - specimen collection - medication administration - wound care - imaging acquisition - room turnover - inventory handling A useful thesis to test: > AI may remove the informational and coordination overhead around physical work even when the physical work remains entirely human. ## 25. Evidence strategy This research should separate doctrine from evidence. ### 25.1 Evidence tiers Tier 1: Primary authoritative sources - statutes - regulations - licensing boards - CMS - HHS - ONC - FDA - state agencies - payer manuals - professional standards - official scope-of-practice documents - contracts and coverage forms where authorized - insurance policy language and underwriting requirements where relevant - intellectual-property agreements and invention-assignment terms where rights are material Tier 2: Direct operational evidence - time-motion studies - workflow logs - EHR audit logs - staffing data - claims - scheduling data - actual work samples - interviews - observation Tier 3: Peer-reviewed research - workforce burden - burnout - workflow - automation - clinical decision support - human factors - safety Tier 4: Vendor or industry evidence Useful for hypotheses but not sufficient for consequential claims. Tier 5: Anecdote and expert judgment Useful for discovering questions, not proving them. ### 25.2 Evidence ledger Every material research claim should eventually have: - claim - evidence - source - source tier - date - jurisdiction - population - confidence - contradictory evidence - implications - unresolved questions ## 26. Research methods Use multiple methods. ### Desk research - regulation - workforce standards - reimbursement - professional scope - academic literature - industry reports ### Workflow ethnography Observe real work. Capture: - interruptions - workarounds - shadow spreadsheets - copy/paste - phone calls - status chasing - duplicate documentation - undocumented tacit decisions ### Process mining Where logs exist, reconstruct actual workflow rather than intended workflow. ### Interviews Interview: - frontline workers - supervisors - patients - compliance - risk - finance - IT - executives Focus on actual work and exceptions, not opinions about AI. ### Artifact analysis Study: - forms - templates - inboxes - spreadsheets - checklists - SOPs - policies - reports - handoff documents - audit evidence ### Controlled prototypes Build narrow packet workers before broad agents. ### Incentive mapping For important work units, map the payer, imposer, beneficiary, risk bearer, delay bearer, and party with authority to remove or change the work. ### Transition and shadow-mode studies Run current and proposed workflows in parallel long enough to measure hidden work, migration burden, exception differences, and cutover risk. ### Capability-formation studies Measure how professionals learn before and after automation, including exposure to routine cases, exceptions, supervision, feedback, and time to independent competence. ### Red-team and resilience testing Test adversarial inputs, compromised evidence, dependency outages, degraded data, model changes, fallback modes, recovery, and correlated failure. ### Causal evaluation Use study designs appropriate to the workflow to separate true intervention effects from secular trends, staffing changes, payer changes, case mix, or other confounders. ### Rights diligence Trace background know-how, data rights, employee and contractor obligations, vendor terms, output ownership, portability, and foreground intellectual property before assuming that encoded capability is commercially controllable. ### OAT Test actual outcomes. ## 27. Research hypotheses to test H1. The job title is usually a poor unit for AI automation design. H2. Work-unit decomposition produces safer and more useful AI systems than role-level agent design. H3. A large fraction of high-cost professional time is spent on evidence reconstruction, coordination, and documentation rather than the professional judgment implied by the credential. H4. Reliable Operational Truth™ eliminates some work entirely rather than merely automating it. H5. Packet workers create a safer path to useful deployment than immediately autonomous execution workers in consequential settings. H6. Task-level AI credentials are more useful than vendor or model-level trust claims. H7. The highest-value productivity measure is often scarce human capability released per accepted outcome, not headcount removed. H8. Human review eventually becomes its own source of meaningless work unless review is risk-based and evidence-driven. H9. AI can increase the effective supply of scarce professional capability even if headcount remains unchanged. H10. Some current software categories are artifacts of historical coordination costs and can be reduced or eliminated when work itself becomes programmable. H11. Some current organizational structures are artifacts of human coordination costs and can be recomposed. H12. Without explicit outcome tests, AI Workforce™ projects systematically overstate success. H13. Data readiness should be evaluated per worker contract, not as a generic organizational maturity score. H14. AI deployment can worsen work if it pushes humans toward continuous high-intensity exception handling. H15. The best work redesign may move tasks to different humans, not to AI. H16. Liability-bearing credentials are a useful discovery signal for high-value unbundling opportunities when the surrounding work is decomposable, observable, and controllable. H17. Credential exposure per accepted outcome is a more useful design metric than credentialed minutes alone in consequential workflows. H18. Explicit separation of preparation, recommendation, authorization, execution, and supervision can reduce unnecessary professional exposure without reducing professional accountability. H19. Work-unit evidence and Operational Truth™ can improve post-event attribution and reduce ambiguity about who knew what, who decided what, and what controls operated. H20. Better risk decomposition can reduce underlying professional risk in some workflows rather than merely transferring liability to another party. H21. In sufficiently mature workflows, granular performance evidence may support more precise underwriting, indemnification, or risk-financing structures than broad role-level assumptions. H22. Some workflows will prove legally or operationally indivisible even when technically decomposable, and those should remain bundled. H23. Job titles and credentials systematically conceal variation in Native Alpha™ among people who nominally occupy the same role. H24. Removing commodity work from differentiated professionals increases Native Alpha™ Density and can improve outcomes even when total headcount does not change. H25. Some technically automatable work should remain human because it is a source of trust, learning, relationship compounding, proprietary insight, reputation, or distribution advantage. H26. Unbundling can falsify supposed differentiation: some capabilities treated as professional or organizational advantages will prove commodity once isolated and measured. H27. AI Workforce™ creates more durable value when it amplifies Native Alpha™ than when it merely lowers labor cost. H28. Recomposed systems that increase Native Alpha™ Density while reducing unnecessary credential exposure will outperform systems optimized only for automation rate or headcount reduction. H29. A meaningful share of apparently meaningless work persists because the party imposing it does not bear its cost. H30. Automating institutionally imposed work without changing the underlying incentive can entrench the work and may increase its volume. H31. Transition cost and brownfield coexistence can reverse the economics of an otherwise superior recomposed workflow. H32. Some routine professional work has positive option value because it creates the pattern recognition and tacit knowledge required for future expert performance. H33. AI Workforce™ scale increases the importance of correlated failure and dependency concentration even when case-level error rates fall. H34. Adversarial robustness is a distinct capability that should be tested and credentialed separately from ordinary task accuracy. H35. Graceful degradation and recoverability materially affect the safety, financeability, and operational acceptability of consequential AI Workforce™ systems. H36. Outcomes Acceptance Testing™ is necessary but insufficient for causal claims about the impact of a recomposed workflow. H37. Healthcare organizations systematically undercount patient and caregiver labor when claiming administrative work has been reduced. H38. Clear rights to encoded know-how, worker contracts, evaluation assets, and derived operating data are required for Native Alpha™ to compound commercially. H39. Encoding differentiated human expertise can destroy Native Alpha™ when rights, access, or vendor terms allow the capability to diffuse to parties that did not previously possess it. H40. Recomposed systems with low dependency concentration, tested fallback modes, and clean rights will be more durable than systems optimized only for short-term cost and throughput. ## 28. Falsification criteria The research should actively look for evidence that challenges the doctrine. The thesis is weakened if: - role-level agents consistently outperform decomposed worker systems without increased risk - decomposition overhead exceeds its operational benefit - workers spend more time reviewing AI than they previously spent performing the task - released time is not converted into useful capacity - patient burden increases - exception complexity causes burnout - task-level credentialing becomes too expensive to maintain - model or workflow change invalidates contracts too frequently - Operational Truth™ cannot be achieved economically - organizations cannot maintain the required evidence - regulation requires role-level human execution even where capability is technically automatable - AI creates more coordination surfaces than it removes - unbundling increases ambiguity about professional or organizational responsibility - professional liability exposure is merely shifted without reducing underlying risk - insurers and legal regimes cannot use task-level evidence in a meaningful way - additional vendors and AI systems create more indemnification and insurance complexity than the workflow savings justify - credential-exposure minimization produces poorer training, safety, trust, or outcomes - the underlying incentive continues to generate the work faster than automation can remove the burden - transition and coexistence costs erase the expected economic benefit - removal of routine work materially slows capability formation or weakens professional judgment - patient or caregiver labor increases even while internal labor falls - correlated failures create more severe population-level harm than the prior heterogeneous human error pattern - adversarial manipulation cannot be contained within an acceptable blast radius - the operation cannot degrade safely when AI, model, data, or cloud dependencies fail - apparent outcome improvement disappears under a credible causal design - rights to encoded know-how, data, workflows, or evaluation assets are too ambiguous to support durable commercialization - vendor terms or affiliate structures allow the organization’s Native Alpha™ to leak or become hostage to a third party Native Alpha™ hypotheses should be rejected or revised when: - the supposedly unusual capability does not measurably change outcomes - peers with ordinary access to the same tools and information reproduce the result easily - removing the human does not diminish trust, quality, learning, relationships, access, reputation, or strategic position - Native Alpha™ Density rises as measured but accepted outcomes do not improve, suggesting the metric is poorly specified - preserving a supposedly differentiated activity creates more cost, delay, risk, or friction than value - the advantage does not compound and remains dependent on one person's undocumented intuition without a strategic reason to preserve that dependence ## 29. Canonical artifacts the research should produce The research program should eventually maintain machine-readable and human-readable versions of: 1. Work Unit Schema 2. Worker Contract Schema 3. Human Capability Requirement Model 4. Human Credential Map 5. AI Worker Credential 6. Authority Matrix 7. Risk Profile 8. Source Priority Policy 9. Data Readiness Profile 10. Operational Truth™ State Model 11. Evidence Supply Chain 12. Review Gate Policy 13. Escalation Policy 14. Memory Policy 15. Outcomes Acceptance Test 16. Incident Record 17. Worker Performance Record 18. Workforce Composition Map 19. Work Graph 20. Recomposition Proposal 21. Economic Impact Model 22. Human Experience Scorecard 23. Patient Experience Scorecard 24. Recertification Record 25. Credential Exposure Profile 26. Liability Decomposition Map 27. Risk Ownership Matrix 28. Indemnification Map 29. Insurance and Risk-Financing Profile 30. Risk Transfer vs Risk Reduction Assessment 31. Liability-Sensitive Workflow Map 32. Native Alpha™ Capability Map 33. Native Alpha™ Density Scorecard 34. Commodity vs Differentiated Work Map 35. Native Alpha™ Compounding Map 36. Preserve / Amplify / Encode / Reject Decision Record 37. Work Origin and Incentive Map 38. Burden Payer / Burden Imposer Matrix 39. Transition and Cutover Plan 40. Shadow-Mode Comparison Record 41. Capability Formation Map 42. Learning Exposure Plan 43. Patient and Caregiver Labor Map 44. Dependency and Common-Mode Failure Map 45. Adversarial Threat Model 46. Graceful Degradation and Continuity Plan 47. Recovery Drill Record 48. Causal Evaluation Plan 49. Distributional Effects Assessment 50. Rights Provenance Record 51. Background / Foreground IP Map 52. Knowledge Capture and Licensing Plan 53. Portability and Vendor Dependency Profile These artifacts should become candidates for GitHub schemas and SpecKit templates. ## 30. Suggested repository structure ```text /unbundling-work README.md /doctrine core-thesis.md zero-meaningless-work.md principles.md /research research-plan.md hypotheses.md evidence-ledger.md falsification.md /incentives work-origin-map.md burden-payer-imposer-matrix.md institutional-friction.md /transition migration-plan.md shadow-mode.md cutover-rollback.md capability-formation.md learning-exposure.md /resilience dependency-map.md common-mode-failure.md adversarial-threat-model.md graceful-degradation.md recovery-drills.md /evaluation causal-evaluation-plan.md distributional-effects.md patient-caregiver-labor.md /rights rights-provenance.md background-foreground-ip.md knowledge-capture.md vendor-portability.md /native-alpha capability-map.md density-scorecard.md commodity-vs-differentiated-work.md compounding-map.md preserve-amplify-encode-reject.md /schemas work-unit.schema.yaml worker-contract.schema.yaml ai-worker-credential.schema.yaml oat.schema.yaml risk-profile.schema.yaml authority-matrix.schema.yaml operational-truth.schema.yaml credential-exposure.schema.yaml liability-decomposition.schema.yaml risk-ownership-matrix.schema.yaml risk-financing-profile.schema.yaml work-origin.schema.yaml dependency-map.schema.yaml threat-model.schema.yaml continuity-plan.schema.yaml rights-provenance.schema.yaml patient-caregiver-labor.schema.yaml /healthcare /roles physician.md registered-nurse.md nurse-practitioner.md physician-assistant.md medical-assistant.md pharmacist.md referral-coordinator.md care-manager.md scheduler-front-desk.md prior-authorization.md coding-billing.md health-information-management.md quality-compliance.md behavioral-health.md physical-therapy.md laboratory-imaging.md practice-management.md clinical-research.md medtech-pharma-regulatory.md /liability credential-exposure.md liability-bearing-credentials.md liability-decomposition.md risk-transfer-vs-risk-reduction.md insurance-risk-financing-hypotheses.md /workflows referral-management.md prior-authorization.md transitions-of-care.md medication-management.md revenue-cycle.md patient-messaging.md /experiments experiment-template.md packet-worker-template.md authority-progression-template.md /evidence sources.md claims/ /visuals visual-backlog.md /specs /audits ``` ## 31. How GitHub specs and AI harnesses should use this research This research plan should not itself become a monolithic implementation spec. Instead it should constrain downstream specs. Every implementation spec should identify: - research claim being tested - target work unit - current human workflow - work-origin and incentive hypothesis - party imposing, paying for, benefiting from, and bearing delay from the work - current evidence sources - current outcome - proposed worker contract - human capability retained - capability-formation value of work being removed - replacement learning pathway where required - patient or caregiver labor currently involved and expected change - claimed Native Alpha™ and evidence supporting or falsifying it - commodity versus differentiated work classification - expected effect on Native Alpha™ Density - compounding mechanism, if any - machine capability introduced - authority level - risk profile - dependency map and common-mode failure analysis - adversarial threat model - fallback and graceful-degradation mode - recovery and continuity requirements - minimum necessary credential exposure - current liability owner - proposed liability owner - risk transferred versus risk actually reduced - indemnification or insurance dependencies where relevant - review gate - Operational Truth™ dependencies - OAT - causal evaluation plan for any claimed impact - rights provenance and ownership of encoded know-how - background and foreground IP treatment - vendor/model portability requirements - transition, shadow-mode, cutover, and rollback plan - expected economic effect including transition cost - expected human experience effect - rollback criteria The harness should be instructed not to invent legal scope, clinical authority, payer rules, or regulatory requirements. These must be sourced and represented as evidence. ## 32. Visual narrative backlog These are narrative suggestions for future whiteboard-style visuals. They are intentionally conceptual so a designer or image-generation harness can later turn them into consistent diagrams. ### Visual 1: The wrong target Left side: A large red target labeled "Zero People" with workers disappearing. Right side: A better target labeled "Zero Meaningless Work" with a human surrounded by removed paperwork, status chasing, duplicate data entry, and evidence gathering. Caption: "Do not optimize away people. Optimize away unnecessary consumption of human capability." ### Visual 2: The bundled job Draw a large box labeled "Registered Nurse." Inside it, many smaller boxes: assessment, vitals, medication administration, education, documentation, coordination, follow-up, triage, escalation. Then show the large box breaking apart into work units. Caption: "The job title is a container. The work is inside." ### Visual 3: Credential as a trust and risk bundle Draw a credential badge with six layers: Knowledge Skill Authority Accountability Liability Insurance / risk financing Then show individual work units connecting only to the layers they truly require. Some require knowledge but not authority. Some require authorization but not direct execution. A small subset consumes the full professional bundle. Caption: "Do not ask whether AI can replace the credential. Ask which work actually consumes the capability, authority, accountability, and liability envelope." ### Visual 4: From title to worker pack Flow: "AI Care Coordinator" with question marks becomes: Care Gap Worker Referral Status Worker Outreach Draft Worker Escalation Worker Documentation Worker Each worker has a small contract card. Caption: "Not one vague agent. Several narrow workers with explicit contracts." ### Visual 5: Progressive authority ladder Seven steps: Observe -> Summarize -> Packet -> Recommend -> Draft -> Execute with approval -> Execute within policy Add risk and evidence increasing requirements upward. Caption: "Authority is earned through evidence." ### Visual 6: Recomposition Center: Physician. Around the physician: pre-visit packet worker, medication worker, care-gap worker, documentation worker, referral worker. Show physician retaining decision and accountability arrows. Caption: "The future is recomposition, not role imitation." ### Visual 7: Operational Truth™ Three columns: Reported Recorded Verified Use a referral example. Reported: "Complete" Recorded: "Referral sent" Verified: "Patient scheduled with specialist" Caption: "Data is not truth. State must be proven." ### Visual 8: Evidence supply chain Patient event -> source system -> timestamped evidence -> worker input -> worker output -> human decision -> outcome. Every arrow traceable. Caption: "Every consequential output should be able to show its work." ### Visual 9: Risk lives in the work Large job box with mixed red, yellow, and green task blocks. Break it apart. Low-risk preparation moves to AI. Medium-risk recommendations have review gates. High-risk decisions remain with accountable humans. Caption: "Unbundle the work and the risk becomes visible." ### Visual 10: Outcomes Acceptance Testing™ Bad test: "Automation ran." Good test: "Correct patient received correct follow-up within 48 hours and unresolved exceptions were escalated." Caption: "Execution is not acceptance." ### Visual 11: Human capability budget Represent a clinician day as 100 tokens of scarce human capability. Current: 60 tokens coordination/documentation/evidence reconstruction 40 tokens professional work Recomposed: 20 tokens review/exceptions 80 tokens professional work Caption: "Human capability is a scarce resource. Spend it deliberately." ### Visual 12: Coordination tax Show EHR, payer, specialist, patient, pharmacy, scheduling system as disconnected islands. Humans are drawn as bridges carrying messages between islands. Then show an evidence and workflow layer connecting the islands, with humans removed from routine bridge work. Caption: "Many people are not doing the work. They are carrying state between systems." ### Visual 13: Work does not disappear if it moves to the patient Before: Staff scheduling burden. Bad automation: Patient fights portal, phone tree, and duplicate forms. Good recomposition: System resolves routine work and human appears for exceptions. Caption: "Do not count work as eliminated when you merely transfer it." ### Visual 14: The unbundling and recomposition loop Observe -> Unbundle -> Classify -> Encode -> Allocate -> Recompose -> Test -> Verify -> Improve Put Operational Truth™ under the entire loop as the evidence foundation. Caption: "Unbundling is not the destination. Better outcomes are." ### Visual 15: Labor as Code™ stack Bottom to top: Operational Truth™ Evidence and Data Work Units Worker Contracts Credentials, Authority, and Liability Human + AI Workforce™ Outcomes Acceptance Testing™ Caption: "Labor becomes programmable only when the work is explicit." ### Visual 16: The professional liability bundle Start with one large box labeled "MD Work + Malpractice Exposure." Inside it place many different activities: chart review, history gathering, evidence synthesis, diagnosis, consent, order authorization, documentation, follow-up, result review, and care coordination. Then unbundle those activities into separate work units. Show only a subset connecting to the full physician judgment and liability envelope. Caption: "Professional liability is bundled around a role. Risk may actually live in specific decisions and failures." ### Visual 17: Minimum necessary credential exposure Draw a funnel: All work -> work requiring professional knowledge -> work requiring judgment -> work requiring authority -> work requiring direct professional execution Show the credentialed professional concentrated at the narrow end rather than spread across every step. Caption: "Use the minimum credential exposure required for a safe accepted outcome." ### Visual 18: Risk transfer is not risk reduction Three side-by-side diagrams. 1. Before: clinician bears broad exposure. 2. Bad recomposition: exposure arrows simply move to hospital or vendor while underlying failure probability stays the same. 3. Good recomposition: evidence, controls, narrower authority, and OAT reduce the underlying failure surface before risk is allocated. Caption: "Moving who pays is not the same as making the work safer." ### Visual 19: Liability Density opportunity map Use a two-axis map. Horizontal axis: work decomposability, low to high. Vertical axis: liability density, low to high. Place example healthcare activities: - complex surgery: very high liability, lower core-act decomposability - physician evidence reconstruction: high liability context, high preparation decomposability - prior-authorization evidence gathering: moderate liability context, high decomposability - routine clerical data entry: low liability, high decomposability Highlight the high-liability / high-decomposability quadrant as a priority research zone. Caption: "High liability alone is not the opportunity. High liability plus decomposable work is." ### Visual 20: Credential Exposure down, Native Alpha™ Density up Draw two opposing vertical arrows around the recomposition engine. Left arrow downward: Commodity work, routine coordination, generic evidence gathering, unnecessary credential touches, broad liability exposure. Right arrow upward: Differentiated judgment, trusted relationships, unusual expertise, proprietary learning, difficult decisions, strategic capability. Center: Unbundling Work -> Labor as Code™ -> AI Workforce™ Caption: "Reduce unnecessary credential exposure. Increase Native Alpha™ Density." ### Visual 21: Job title hides Native Alpha™ Start with five identical boxes labeled "Cardiologist." Open each box to reveal a different pattern of unusual capability: complex diagnostics, patient trust, electrophysiology expertise, referral-network knowledge, procedural skill. Around each, show common commodity work that can be removed or standardized. Caption: "The credential tells us what someone may do. It does not tell us what they do unusually well." ### Visual 22: Preserve, amplify, encode, reject Place a candidate differentiated capability in the center. Route it through four questions: - Does it materially change outcomes? - Is it reproducible by ordinary peers or tools? - Does human participation create strategic value? - Does it compound? End in four destinations: Preserve / Amplify / Encode / Reject as false differentiation. Caption: "Native Alpha™ is a hypothesis to test, not a compliment." ### Visual 23: AI Workforce™ as leverage, not substitution Before: One specialist surrounded by evidence gathering, documentation, status chasing, and routine preparation, serving a small number of difficult cases. After: AI workers and automation handle commodity preparation while the same specialist applies unusual judgment across many more difficult cases. Caption: "The professional becomes more important to the outcome while doing less commodity work." ### Visual 24: The compounding test Show two branches after automation. Branch A: Cost reduction -> copied by competitors -> advantage disappears. Branch B: Released capability -> better outcomes -> better data / relationships / workflow position / know-how -> stronger future capability -> compounding loop. Caption: "Efficiency is useful. Compounding Native Alpha™ is strategic." ### Visual 25: Why the work exists Draw one work unit in the center, such as prior authorization evidence gathering. Around it, show possible work-creation mechanisms: Clinical necessity / Safety control / Regulation / Reimbursement / Contract / Defensive practice / Fragmented systems / Vendor limitation / Habit / Cost shifting. Caption: "Before automating the burden, identify what creates the burden." ### Visual 26: Automate the burden versus remove the cause Two paths from the same painful workflow. Path A: Bad rule -> AI performs the bad rule faster -> burden remains. Path B: Bad rule -> incentive or contract redesigned -> work disappears. Caption: "Automation is not elimination when the institution keeps manufacturing the work." ### Visual 27: The apprenticeship trap Left: Junior professional performs a mix of routine and difficult cases and gradually develops pattern recognition. Middle: AI absorbs all routine cases. Right fork: Bad future: junior sees only confusing exceptions and never develops calibration. Good future: simulation, sampled routine cases, supervised execution, deliberate practice, and graded authority create a new learning path. Caption: "Zero Meaningless Work must not become Zero Apprenticeship." ### Visual 28: Heterogeneous error versus correlated error Left: 100 human workers make scattered, different mistakes. Right: 100 AI workers share one bad policy and make the same mistake at once. Overlay blast-radius and detection-time arrows. Caption: "Scale changes the shape of risk." ### Visual 29: The invisible healthcare workforce Place the patient and caregiver in the center of a home. Around them show scheduling, portal work, medication tracking, transportation, records chasing, insurance calls, home monitoring, and escalation. Then connect those tasks back to healthcare organizations that often treat this labor as external or free. Caption: "Work does not stop being work because nobody puts it on payroll." ### Visual 30: Brownfield migration Show two operating systems side by side: Current bundled workflow -> shadow mode -> parallel validation -> limited cutover -> expanded authority -> full recomposition. Underneath, show rollback available at every stage. Caption: "A good future-state design can still fail during migration." ### Visual 31: Native Alpha™ and rights Show a professional's unusual judgment flowing into worker contracts, evaluation cases, policies, data, and operating know-how. Then fork: Controlled path: clean ownership -> durable license -> portable implementation -> compounding Native Alpha™. Leaky path: unclear employee/contractor rights -> vendor rights -> model lock-in -> capability diffuses away. Caption: "Encoding Native Alpha™ creates an asset only if the rights are clean." ### Visual 32: Graceful degradation Normal mode: AI workers execute within policy. Degraded mode: AI loses trusted data or model access -> authority automatically drops -> humans receive prioritized packets -> noncritical work queues. Recovery: Systems return -> reconciliation -> OAT checks -> authority restored. Caption: "Consequential work needs a safe answer for when AI is unavailable." ## 33. Initial healthcare experiments The first experiments should be narrow, high-frequency, evidence-rich, and easy to evaluate. ### Experiment A: Referral status packet worker Goal: Prepare verified referral status and next-action packets. No autonomous patient contact initially. Measure: - coordinator minutes per referral - stale referral age - false status rate - escalation accuracy - time to closure - human edits ### Experiment B: Pre-visit physician evidence packet Goal: Reduce physician evidence reconstruction before complex visits. Measure: - physician prep minutes - missing critical evidence - packet correction rate - visit throughput - clinician trust - after-hours work ### Experiment C: Prior authorization evidence packet Goal: Assemble required evidence and identify missing elements before submission. Measure: - staff minutes - clinician interruptions - first-pass approval - cycle time - avoidable denial rate ### Experiment D: RN triage packet Goal: Structure patient message, relevant history, medication context, risk flags, and recommended protocol path for nurse review. Measure: - nurse review time - escalation sensitivity - false reassurance - response latency - override rate ### Experiment E: Care-gap packet Goal: Identify overdue care and prepare evidence-linked action recommendations. Measure: - precision - recall - staff review minutes - closed gaps - inappropriate outreach avoided ### Experiment F: Revenue-cycle denial evidence worker Goal: Classify denial, retrieve relevant evidence, and prepare an appeal packet. Measure: - staff minutes - appeal quality - overturn rate - cycle time - revenue recovered ### Experiment G: Credential Exposure and Liability Map Goal: Select one high-liability healthcare workflow and decompose every professional touch, authority point, liability assumption, indemnification relationship, and evidence artifact. Candidate workflows: - abnormal test result follow-up - medication change and refill approval - referral loop closure - triage escalation - prior authorization requiring physician attestation Measure: - credentialed touches per accepted outcome - credentialed minutes per accepted outcome - professional decisions versus preparation steps - current liability owner by stage - ambiguity in responsibility - evidence available after simulated failure - proposed minimum necessary credential exposure - risk transferred versus risk actually reduced Output: A before-and-after Liability Decomposition Map and Credential Exposure Profile. ### Experiment H: Insurability and underwriting interview study Goal: Test whether insurers, brokers, captive managers, health-system risk leaders, malpractice counsel, and actuaries consider work-unit evidence useful for risk assessment. Questions: - What information currently drives professional liability underwriting? - Which workflow controls materially influence perceived risk? - Could OAT performance, evidence provenance, incident traceability, or worker-contract boundaries ever influence underwriting? - What would make AI-enabled workflow decomposition increase rather than decrease premiums? - Which liability exposures remain legally inseparable from the licensed professional? Output: A research-backed boundary between technically attractive risk decomposition and commercially or legally usable risk decomposition. ### Experiment I: Physician Native Alpha™ decomposition study Goal: Test whether unbundling reveals meaningful differentiation among professionals who share the same credential and nominal role. Method: Select several physicians in the same specialty or care setting. Map identical categories of work, then separately identify where outcomes appear to benefit from person-specific expertise, relationships, pattern recognition, procedural skill, local context, judgment, or trust. Questions: - Which capabilities are common professional competence versus person-specific differentiation? - Which claims of unusual capability survive peer comparison and outcome evidence? - Which commodity activities consume the most time around differentiated work? - Which human interactions should remain direct because they create trust or learning? - How much more differentiated work becomes possible after surrounding commodity work is removed? - Does Native Alpha™ Density increase without unacceptable increases in cognitive intensity or exception fatigue? Measure: - commodity minutes removed - credential exposure per accepted outcome - differentiated professional minutes per accepted outcome - difficult cases handled - outcome differences where measurable - patient trust and continuity where relevant - learning, referral, data, or workflow compounding signals Output: A Native Alpha™ Capability Map, Commodity vs Differentiated Work Map, and before-and-after Native Alpha™ Density Scorecard. ### Experiment J: Prior authorization incentive and work-origin map Goal: Determine which prior-authorization work exists because of genuine clinical risk control versus payer-provider incentive conflict, documentation structure, contract terms, or information asymmetry. Measure: - work units by origin - payer versus provider minutes - clinician minutes imposed by external requirements - denial prevention value - work that could disappear through better evidence exchange - work that remains even after automation because the underlying incentive is unchanged Output: A Work Origin and Incentive Map showing which burden should be automated and which should be eliminated at the institutional source. ### Experiment K: Shadow-to-cutover workflow migration Goal: Measure the real transition cost of moving one high-volume workflow from human execution to a recomposed AI Workforce™. Method: - baseline current operation - shadow AI operation with no authority - compare outputs and hidden work - introduce limited packet use - progress authority only after OAT evidence - conduct planned rollback drill Measure: - duplicate labor during transition - integration cost - staff training time - exception discovery - cutover incidents - rollback performance - time to net positive economics ### Experiment L: Capability formation study Goal: Determine whether removing routine work weakens the development of future professional expertise. Candidate populations: - residents or fellows - new RNs - pharmacy residents or technicians progressing in responsibility - junior revenue-cycle or coding staff - junior regulatory or quality professionals Measure: - case exposure - time to competence - error recognition - calibration - exception performance - need for supervision - learner confidence versus measured competence Output: A Capability Formation Map and replacement Learning Exposure Plan. ### Experiment M: Correlated failure and resilience drill Goal: Test a worker pack against a deliberately introduced common-mode failure. Scenarios: - stale source data - incorrect shared policy - model behavior change - integration outage - compromised external document - cloud or vendor unavailability Measure: - cases affected before detection - time to authority reduction - maximum blast radius - fallback success - recovery time - reconciliation errors ### Experiment N: Patient and caregiver labor study Goal: Measure the unpaid work surrounding one common healthcare journey before and after recomposition. Candidate journeys: - specialty referral - post-discharge follow-up - chronic medication management - diagnostic workup - prior authorization appeal Measure: - unpaid minutes - calls - portal sessions - forms - records transfers - transportation coordination - missed work - repeated explanations - caregiver involvement Output: A Patient and Caregiver Labor Map that is included in the economic effect rather than treated only as experience feedback. ### Experiment O: Rights and knowledge-capture pilot Goal: Take one high-value professional workflow and trace what intellectual assets are created when the work is encoded. Map: - background professional know-how - organizational policies - preexisting intellectual property - worker contracts - prompts - evaluation cases - evidence structures - derived operating data - newly created know-how - vendor and model-provider terms - employee and contractor obligations Output: A Rights Provenance Record, Background / Foreground IP Map, and Knowledge Capture and Licensing Plan. ## 34. Research phases ### Phase 1: Doctrine and ontology Deliver: - core thesis - terminology - Work Unit schema - Worker Contract schema - risk model - AI Worker Credential - OAT model - Operational Truth™ model - work-origin and incentive taxonomy - dependency and systemic-risk model - rights-provenance model ### Phase 2: Healthcare work, credential, liability, and Native Alpha™ atlas Create detailed role and workflow maps, including claims of differentiated human and organizational capability that can later be tested. Include work-origin, patient/caregiver labor, capability-formation, dependency, and rights views rather than only paid staff activity. Priority: - physician - RN - medical assistant - referral coordinator - prior authorization - pharmacist - revenue cycle - practice manager ### Phase 3: Evidence, credential exposure, Native Alpha™ testing, and measurement Gather: - time studies - workflow logs - source-system evidence - policies - regulations - role boundaries - cost data - evidence for or against claimed Native Alpha™ - measures of commodity versus differentiated work - trust, relationship, workflow-position, learning, data, and compounding effects - institutional incentives that create or preserve work - unpaid patient and caregiver labor - dependency concentration and correlated failure surfaces - ownership and rights evidence ### Phase 4: Packet-worker prototypes and shadow operation Build narrow workers with no autonomous high-risk action. Run them in shadow mode against the current operation to expose hidden work, transition cost, disagreement, dependency risk, and fallback requirements before cutover. ### Phase 5: OAT and authority progression Test: - reliability - review burden - exceptions - safety - economic effect - human experience - adversarial robustness - degraded-mode operation - correlated failure - capability-formation effect - patient and caregiver labor ### Phase 6: Recomposition, Native Alpha™ amplification, liability, and risk-financing studies Redesign complete workflows, not isolated tasks. Test whether the recomposed system increases Native Alpha™ Density and compounding advantage rather than merely lowering labor cost. Include incentive redesign, transition economics, systemic risk, rights control, causal evidence, and risk-financing implications. ### Phase 7: Generalization Test framework in other consequential industries: - government - legal - accounting - insurance - financial services - manufacturing - construction - regulated software - cybersecurity ## 35. Metrics for the research area Do not measure success by number of AI workers. Measure: ### Work - human touches per outcome - human minutes per outcome - credentialed minutes per outcome - credentialed touches per outcome - credentialed decisions per outcome - credentialed authorizations per outcome - credential exposure per accepted outcome - coordination minutes - documentation minutes - evidence-reconstruction minutes - review minutes - exception rate ### Outcome - completion - correctness - latency - safety - quality - customer/patient outcome - loop closure ### Economics - cost per accepted outcome - capacity released - revenue gained or protected - avoidable cost - outsourcing reduced - new work made feasible ### Risk and liability - severity-weighted professional authority exercised - liability-bearing steps per outcome - responsibility ambiguity count - failure-to-follow-up exposure - incident frequency and severity - evidence completeness after incidents - professional claims frequency and severity where measurable - risk transferred versus risk actually reduced - insurance, indemnification, or self-insurance cost where measurable ### Trust - human override rate - human correction rate - worker suspension rate - OAT pass rate - source conflict rate - evidence completeness - auditability ### Strategic differentiation - Native Alpha™ Density: share of scarce human or organizational capacity applied where differentiated capability materially changes outcomes - Native Alpha™ minutes per accepted outcome where measurable - commodity work removed from differentiated professionals - number of accepted outcomes reached with the same differentiated capability - difficult or unusual cases made feasible by released capability - trust, relationship, referral, reputation, or distribution effects - proprietary data or learning generated by the recomposed workflow - workflow position strengthened - intellectual property or operating know-how created - evidence that the advantage compounds rather than merely reduces cost ### Human - top-of-capability time - burnout indicators - interruption rate - context switching - after-hours work - job satisfaction - exception fatigue ### Patient/caregiver/customer - unpaid labor minutes per accepted outcome - effort - delay - repeated questions - access - ability to reach a human - trust - continuity ### Work origin and incentives - share of work units intrinsic to the outcome versus institutionally imposed - minutes created by payer, regulator, contract, vendor, or internal-policy requirements - burden imposed on one party but paid by another - work eliminated at source versus merely automated - volume growth after automation of an externally imposed requirement ### Transition - shadow-mode duration - duplicate labor during migration - migration cost - retraining time - cutover defects - rollback frequency and success - time to net positive economics - legacy systems or procedures actually retired ### Capability formation - supervised cases per learner - routine versus exception case mix - time to demonstrated competence - calibration accuracy - pattern-recognition performance - mentoring time - retained learning exposure ### Systemic risk, security, and resilience - maximum correlated blast radius - dependency concentration - number of outcomes sharing a common critical dependency - adversarial-test pass rate - time to detect common-mode failure - time to reduce authority after degradation - degraded-mode throughput - recovery time - reconciliation error after recovery - continuity drill pass rate ### Causal evidence and distributional effects - baseline-adjusted effect size - confidence interval or credible uncertainty range where appropriate - case-mix adjusted outcome difference - site and population heterogeneity - patient and caregiver unpaid minutes per outcome - work transferred across roles or outside payroll - subgroup harms or benefits hidden by aggregate averages ### Rights and ownership - work units with clear rights provenance - percentage of critical capability dependent on third-party terms - foreground intellectual property clearly owned or licensed - background intellectual property dependencies documented - portability from model or vendor - unresolved employee, contractor, affiliate, customer, or vendor rights - encoded Native Alpha™ under durable organizational control ## 36. Counterarguments that deserve serious treatment ### "Meaningless work" is subjective Correct. The research must classify work by function and consequence rather than dislike or prestige. ### AI can create more work than it removes Correct. Measure review, exceptions, monitoring, maintenance, incidents, and work transfer. ### Humans learn by doing lower-level work Correct. Unbundling can destroy apprenticeship pathways. Research how future professionals acquire tacit knowledge if AI handles routine cases. ### Top-of-license work can be exhausting Correct. A day consisting only of the hardest cases may be worse than a mixed workload. ### Some inefficiency creates safety Correct. Redundancy, independent verification, and deliberate friction may be protective. ### Patients may want humans Correct. Human presence can have independent value that is not reducible to clinical throughput. ### Regulation may lag capability Correct. Technical feasibility does not grant legal authority. ### AI errors scale Correct. Automation can increase blast radius even while reducing error rate. ### Standardization can erase useful local judgment Correct. Worker contracts should encode escalation and local policy rather than force false uniformity. ### Unbundling can create liability gaps Correct. Separating work among clinicians, organizations, vendors, and AI workers can make responsibility less clear if the contracts, evidence, and accountability model are weak. A recomposed workflow should be rejected if nobody clearly owns the failure. ### Lower professional involvement does not necessarily lower malpractice risk Correct. Some forms of professional review are protective. Insurance pricing is influenced by many factors outside the workflow. The research should measure underlying risk and claims evidence before making any premium claim. ### More vendors can mean more insurance complexity Correct. Technology E&O, cyber, product liability, malpractice, contractual indemnity, and organizational coverage can overlap or leave gaps. A technically elegant workflow may be commercially unattractive if its risk allocation is unfinanceable or opaque. ### Native Alpha™ can become self-serving mythology Correct. People and organizations routinely overestimate how differentiated their work is. Native Alpha™ should be treated as a hypothesis that must survive decomposition and measurement, not a compliment. If ordinary people with ordinary tools and ordinary access can reproduce the capability, it is probably not Native Alpha™. ### Automation can accidentally destroy Native Alpha™ Correct. A workflow can become cheaper while weakening patient trust, professional learning, referral relationships, tacit knowledge, proprietary data generation, reputation, or strategic workflow position. Cost reduction that destroys a compounding advantage can be negative value. ### AI can entrench bad incentives Correct. If a payer, regulator, department, or other party can impose work while somebody else bears the cost, cheaper automation may make it easier to demand more of the work. The research should distinguish automating institutional friction from removing its cause. ### Removing routine work can weaken the future expert pipeline Correct. Some apparently low-level work is part of apprenticeship. The framework needs an explicit replacement learning path before removing work that creates judgment, pattern recognition, or credential progression. ### AI creates monoculture risk Correct. Shared models, prompts, policies, and data can turn small case-level errors into synchronized population-level failures. Case-level accuracy is not enough. ### Security is part of capability, not an IT afterthought Correct. A worker that performs accurately on benign inputs but follows malicious instructions or trusts poisoned evidence is not competent for consequential work. ### A workflow can pass OAT and still have an unproven business case Correct. Outcomes Acceptance Testing™ verifies accepted outcomes. Causal evaluation is still needed before attributing improvements to the new workflow. ### Encoded expertise can leak the advantage Correct. Turning tacit Native Alpha™ into prompts, worker contracts, evaluation cases, or vendor-hosted workflows can make the capability more transferable. Clean rights and deliberate knowledge governance are necessary before assuming encoding creates a durable asset. ### The best future state may still be too expensive to reach Correct. Brownfield migration, parallel operation, retraining, policy changes, integrations, and retirement of legacy systems can dominate the economics. Transition cost belongs in the decision, not in an implementation footnote. ## 37. Principles for commercialization The research should produce commercially useful insight without becoming vendor-driven. Potential product or service wedges should be evaluated using: 1. Clear work unit. 2. High-frequency burden. 3. Trusted evidence sources. 4. Narrow authority. 5. Measurable outcome. 6. Obvious buyer. 7. Economic consequence. 8. Fast packet-worker deployment. 9. Low integration burden. 10. Expansion path based on demonstrated performance. 11. Clear liability owner. 12. No hidden transfer of risk to the customer, patient, professional, or affiliate. 13. Insurance and indemnification structure that can survive independent legal and investor scrutiny. 14. Evidence that the product reduces or controls risk rather than merely changing who pays for failure. 15. Clear statement of whose Native Alpha™ is being preserved, amplified, created, or compounded. 16. Evidence that the proposed system does more than automate commodity work that competitors can reproduce with the same models and tools. 17. Clear explanation of why the work exists and whether the product removes the cause or merely absorbs the burden. 18. Transition cost and cutover path that a real customer can tolerate. 19. Capability-formation plan where automation removes meaningful apprenticeship or professional learning. 20. Patient and caregiver labor effect included in the value case where relevant. 21. Common-mode failure, adversarial security, dependency concentration, and graceful-degradation plan. 22. Causal evidence proportionate to the importance of the claimed outcome improvement. 23. Clean rights to the worker contracts, evaluation assets, data products, encoded know-how, and foreground intellectual property needed to commercialize the capability. 24. Portability and vendor terms that do not make a strategically important Native Alpha™ dependent on one model or provider without deliberate justification. Do not assume that intellectual novelty creates demand. Do not assume that a patent, credential, proprietary model, workflow, or dataset creates Native Alpha™ merely because it is unusual. The relevant question is whether it changes a real commercial or operating decision and whether the advantage can compound. Demand proof should include scarce commitments such as: - payment - implementation time - data access - workflow access - executive sponsorship - recurring use - willingness to change operating procedure ## 38. Native Alpha™ as the recomposition principle Native Alpha™ should be embedded in Unbundling Work as the answer to a question that ordinary process optimization does not ask: > Once the work has been unbundled, what should we preserve, concentrate, and amplify rather than automate away? Without this question, Unbundling Work can collapse into conventional efficiency engineering: decompose the job, automate the cheap pieces, reduce labor cost, and declare success. The stronger objective is: > Unbundle commodity work so scarce, differentiated capability can be concentrated where it creates disproportionate value. ### 38.1 Credentials and job titles hide differentiation A credential defines some combination of education, demonstrated competence, legal authority, scope, trust, and accountability. It does not show what a particular person is unusually good at. Two physicians with the same specialty and license may have very different Native Alpha™. One may possess exceptional diagnostic pattern recognition. Another may have unusual trust with a patient population. Another may have rare procedural skill, unusually deep disease knowledge, better judgment about when not to intervene, or a uniquely valuable referral network. The same distinction applies to nurses, pharmacists, attorneys, engineers, researchers, salespeople, operators, and executives. The research should therefore distinguish: - credentialed capability: what the person is qualified or authorized to do - ordinary professional capability: what a competent peer can reasonably do - Native Alpha™: what this particular person, team, or organization can know, see, decide, or do unusually well in a way that matters to outcomes A major purpose of unbundling is to expose those differences. ### 38.2 Minimum credential exposure and maximum useful differentiation The liability work introduces one optimization objective: > Use the minimum necessary credential, authority, and liability exposure required to safely produce the accepted outcome. Native Alpha™ introduces the complementary objective: > Apply differentiated capability wherever it materially improves or compounds the outcome. These should be evaluated together rather than confused. A work unit may require physician authority but gain no additional value from a particular physician's unusual capability. Another work unit may not legally require a physician at all, yet a particular physician's trust relationship or unusual clinical judgment may materially improve the result. This leads to the paired design objective: > Credential Exposure ↓ / Native Alpha™ Density ↑ The goal is to consume as little credentialed liability capacity as necessary while applying as much differentiated capability as is useful. ### 38.3 Native Alpha™ Density For research purposes, Native Alpha™ Density should mean the proportion of scarce human or organizational capacity applied to work where unusual capability materially changes the accepted outcome or creates a compounding strategic advantage. It is not intended initially as a universal numerical score. The research should determine where it can be measured credibly and where qualitative evidence is more appropriate. Illustrative physician example: Before recomposition: - 100 units of physician capacity - 60 consumed by evidence reconstruction, documentation, coordination, routine preparation, and status chasing - 20 used for professional work that competent peers could perform similarly - 20 applied where that physician's unusual judgment, relationship, or expertise materially changes the result After recomposition: - total physician capacity consumed may fall - commodity work may move to AI workers, deterministic automation, or other humans - credentialed approval may be concentrated only where legally or clinically necessary - the physician may apply unusual judgment to more cases, harder cases, or higher-value interactions The physician can become more important to the system while spending fewer total minutes inside it. That is fundamentally different from a Zero People strategy. ### 38.4 Some automatable work should remain human Technical feasibility should not determine disposition by itself. A human interaction may be strategically valuable because it: - creates or protects trust - generates unusual contextual information - teaches the professional something that improves future judgment - deepens a patient, customer, referral, or partner relationship - strengthens reputation - produces proprietary data - creates distribution advantage - develops tacit knowledge - generates new intellectual property or operating know-how - places the organization more deeply into a consequential workflow If automation removes those effects, it may reduce cost while destroying Native Alpha™. The research must therefore ask not only: > Can AI do this? But also: > What strategic capability disappears if the human no longer does this? ### 38.5 Unbundling should falsify false differentiation Native Alpha™ should not become a way to protect prestigious work from scrutiny. Some work will look differentiated only because it is performed by an expensive or credentialed person. Once decomposed, the capability may prove routine, learnable, transferable, or reproducible by ordinary tools. That is a useful result. The research should actively seek to falsify Native Alpha™ claims. Candidate tests: - Do outcomes materially improve when this specific person or organization performs the work? - Can competent peers reproduce the performance? - Can standard AI and ordinary data reproduce it? - Does the advantage depend on privileged access, unusual insight, relationships, data, workflow position, intellectual property, or accumulated know-how? - Does the capability create downstream compounding advantage? - Would a customer, patient, partner, investor, or buyer allocate scarce resources because of it? If the answer is no, the capability may be valuable but it is probably not Native Alpha™. ### 38.6 Native Alpha™ at the organizational level Organizations are also bundles. A health system may operate thousands of workflows while possessing Native Alpha™ in only a few areas, such as: - unusual clinical expertise - trusted community relationships - proprietary longitudinal data - superior patient acquisition or retention - better payer or referral relationships - unusual operational know-how - stronger evidence supply chains - proprietary intellectual property - privileged workflow position - unusually effective risk management - distinctive distribution Unbundling Work should identify which activities strengthen those advantages and which activities are generic organizational baggage. This creates a useful strategic question: > If this work is not a source of Native Alpha™, why are we consuming differentiated people, bespoke software, management attention, or proprietary infrastructure to perform it? That question connects naturally to Zero Neo and to build-versus-buy decisions without making either doctrine universal. ### 38.7 AI Workforce™ as a Native Alpha™ amplifier The strongest AI Workforce™ should not merely substitute for labor. It should increase the effective supply of differentiated capability. Examples: - a rare-disease expert reviews five times as many high-quality evidence packets without five times as much evidence-gathering work - a trusted physician spends more time in difficult shared decisions and less time reconstructing records - a high-performing nurse applies triage judgment to exceptions while packet workers prepare routine evidence - a regulatory expert reviews consequential interpretations while AI workers maintain evidence matrices and traceability - an operator with unusual workflow knowledge supervises encoded work that can now be deployed across many sites The Native Alpha™ existed before the AI Workforce™. Unbundling makes it visible. Labor as Code™ makes the surrounding work explicit. AI Workforce™ gives it leverage. Operational Truth™ supplies evidence. Outcomes Acceptance Testing™ verifies that the leverage improves real outcomes. ### 38.8 Compounding is the final test Cost reduction alone is easy to copy. The most strategically interesting recompositions should cause the advantage to compound through one or more of: - proprietary data - better decision history - workflow position - trust - customer or patient relationships - distribution - reputation - operating know-how - intellectual property - faster learning - better underwriting or risk evidence - future products or services This is where Unbundling Work becomes more than an efficiency framework. It becomes a method for reallocating commodity work away from scarce capability and then deliberately strengthening the capability that actually differentiates the person or organization. ### 38.9 Native Alpha™ must survive knowledge capture Encoding unusual human or organizational capability can strengthen Native Alpha™ when it creates reusable worker contracts, better evidence, evaluation assets, proprietary data, or operating know-how under durable control. It can also weaken Native Alpha™ when: - an external vendor receives broad rights to the encoded capability - employee or contractor ownership is unclear - the organization cannot move the capability away from a model provider - customer or affiliate rights are not defined - trade-secret controls are lost - the capability becomes easily reproducible by competitors The research should therefore treat rights provenance as part of the compounding test. ### 38.10 Native Alpha™ depends on future capability formation A professional's unusual judgment usually emerged from accumulated experience. If recomposition removes the routine work from which future experts learn, the organization can consume today's Native Alpha™ while failing to regenerate tomorrow's. Every major recomposition should ask: > What learning loop created this differentiated capability, and will the new system preserve or replace that loop? Native Alpha™ Density should not be increased by starving the future capability pipeline. ### 38.11 Canonical Native Alpha™ research questions For every major recomposition, ask: 1. What work is commodity? 2. What work genuinely requires a credential or authority? 3. What work materially benefits from this particular person's or organization's unusual capability? 4. What apparent differentiation disappears under decomposition? 5. What technically automatable work should remain human because it creates strategic value? 6. What surrounding work can be removed so differentiated capability reaches more cases or harder cases? 7. Does the recomposition increase Native Alpha™ Density? 8. Does it reduce unnecessary credential exposure? 9. Does it strengthen data, relationships, workflow position, trust, distribution, intellectual property, or operating know-how? 10. Does the advantage compound, or is it merely a one-time labor saving? ## 39. Incentives and Work Creation research program ### 39.1 Core premise Work is produced by institutions as well as by customer or clinical need. A useful work-design system must distinguish work that is intrinsic to producing the outcome from work created by incentives, contracts, regulation, information asymmetry, fragmented ownership, defensive behavior, or historical structure. ### 39.2 Work-origin taxonomy Every major work unit should be classified initially as one or more of: - outcome-intrinsic work - safety-control work - legal or regulatory work - reimbursement work - payer-utilization work - contractual work - defensive documentation - coordination work caused by fragmentation - information-reconstruction work - management or reporting work - vendor-imposed work - policy-imposed work - cost-shifting work - market-power work - obsolete or historical work The classification can change as evidence improves. ### 39.3 Burden payer versus burden imposer For every institutional work unit identify: - who requires the work - who executes it - who pays for it - who benefits from it - who bears the delay - who bears the failure risk - who has authority to remove or change it This can expose situations in which one party rationally creates work because another party absorbs most of the cost. ### 39.4 Automate, redesign, renegotiate, or eliminate The disposition framework should therefore expand beyond automate versus human. Candidate dispositions: - retain because intrinsic - automate because necessary but mechanical - AI-assist because preparation is burdensome - redesign the information exchange - change policy - change contract - change reimbursement mechanism - shift accountability - consolidate organizations or handoffs - eliminate the underlying requirement The best AI intervention may be evidence that the work should disappear entirely. ### 39.5 Healthcare proving grounds High-value targets include: - prior authorization - referral documentation - quality reporting - coding and billing evidence - defensive clinical documentation - utilization review - repeated eligibility verification - duplicate consent and intake - credentialing and enrollment - regulatory evidence assembly The research should avoid assuming that every burden is irrational. Some institutional friction is deliberately protective. The objective is to identify whether the control is proportionate and whether the burden is borne by the right party. ## 40. Transition, Learning, and Capability Formation research program ### 40.1 Brownfield reality Most consequential organizations cannot replace a workflow in one cutover. They must transition while continuing to serve patients, customers, regulators, and partners. Research should treat migration itself as a work system. ### 40.2 Transition states Candidate progression: 1. Observe current work. 2. Replay historical cases. 3. Run machine worker in shadow mode. 4. Compare human and machine outputs. 5. Use machine packets with human execution unchanged. 6. Move narrow actions behind explicit approval. 7. Move narrow low-risk actions under policy. 8. Retire duplicated manual work only after evidence supports it. 9. Maintain tested rollback and degraded modes. ### 40.3 Transition cost Measure: - parallel-run labor - integration and data work - policy changes - training - temporary throughput loss - staff anxiety and resistance - exception discovery - contract changes - legal and compliance review - legacy retirement - rollback effort The research should calculate time to net positive value, not merely steady-state savings. ### 40.4 Capability formation Create a Capability Formation Map for roles where learning depends on repeated work. For every work category ask: - Does doing this work build useful capability? - What capability? - At what career stage? - How much repetition is useful before marginal learning falls? - Which feedback is necessary? - Can simulation substitute? - Should humans perform a sample even if machines could do all cases? - Which routine cases are prerequisites for handling exceptions safely? ### 40.5 New apprenticeship models Potential mechanisms: - simulation using historical cases - synthetic cases - deliberate practice - shadow decision-making before seeing the machine recommendation - sampled human execution - graded authority - structured review of machine errors - rotation through exception categories - supervised adversarial cases - competency testing tied to real outcomes The goal is not to preserve grunt work ceremonially. It is to preserve the learning that matters. ## 41. Systemic Risk, Security, and Resilience research program ### 41.1 From individual error to system error As AI Workforce™ systems scale, risk can become more correlated. The relevant question is no longer only: > How often does this worker make a mistake? It is also: > How many outcomes can the same mistake affect before the system detects, contains, and recovers from it? ### 41.2 Dependency concentration Map dependencies across: - models - model providers - cloud providers - identity systems - prompts and policies - shared tools - source systems - integration platforms - external knowledge sources - humans with unique approval authority - secrets and credentials Develop concentration thresholds where practical. ### 41.3 Common-mode failure Test failures such as: - wrong shared rule - stale shared data - incorrect model update - broken integration - common prompt flaw - policy misconfiguration - compromised identity provider - vendor outage - corrupted upstream transformation ### 41.4 Adversarial security Treat resistance to manipulation as a worker capability. Research controls for: - prompt injection - poisoned documents - spoofed evidence - malicious insiders - fraudulent users or providers - excessive tool authority - secret leakage - unauthorized action - policy tampering - compromised dependencies ### 41.5 Graceful degradation Every consequential worker pack should define: - normal mode - reduced-authority mode - read-only mode - manual fallback mode - safe-stop conditions - queued-work behavior - priority rules - recovery order - reconciliation requirements ### 41.6 Resilience as commercial infrastructure Customers, insurers, investors, boards, regulators, and partners may care less about a demo's peak performance than whether the operation remains safe during inevitable failure. Resilience should therefore be treated as part of product quality and financeability, not only IT operations. ## 42. Causal Evidence and Distributional Effects research program ### 42.1 Outcome verification versus causal inference Operational Truth™ and Outcomes Acceptance Testing™ can show what happened and whether it met acceptance criteria. Research must separately estimate what would have happened without the intervention. ### 42.2 Causal study design Choose methods proportionate to the decision: - simple baseline for low-stakes operational questions - matched comparison for moderate claims - phased deployment across sites or teams - interrupted time series - difference-in-differences - randomized evaluation where appropriate - prospective trials for consequential clinical claims The objective is not academic purity. It is enough rigor to avoid attributing unrelated improvement to AI. ### 42.3 Distributional effects Aggregate improvement can hide who gained and who absorbed the burden. Analyze effects across: - professional roles - sites - patient populations - languages - digital-access levels - complexity bands - payers - caregivers - high-risk versus routine cases - new versus experienced workers ### 42.4 Patient and caregiver labor as an outcome Patient and caregiver labor should become an explicit dependent variable. A recomposed workflow that saves 20 staff minutes by adding 40 minutes of portal, phone, transportation, record-gathering, or home-management work to a family has not achieved Zero Meaningless Work. ### 42.5 Work transfer ledger For every experiment maintain a ledger showing where work moved: - removed - automated - moved to another employee - moved upward to a credentialed professional - moved downward to lower-cost staff - moved to a vendor - moved to the patient - moved to a caregiver - moved into future review or exception handling This should prevent local optimization from masquerading as system improvement. ## 43. Rights, Knowledge Capture, and Ownership research program ### 43.1 Why rights matter Unbundling and Labor as Code™ can turn tacit capability into explicit artifacts. Those artifacts may include: - worker contracts - work-unit ontologies - prompts - tool configurations - policies - evidence schemas - evaluation cases - acceptance tests - incident libraries - derived data - operating procedures - workflow graphs - risk models - proprietary training material These may become valuable intellectual assets only if ownership and usable rights are clear. ### 43.2 Background versus foreground capability For each project distinguish: - preexisting individual know-how - preexisting company know-how - licensed affiliate or parent background intellectual property - customer-provided materials - third-party content - newly created foreground intellectual property - newly created operational data - newly created evaluation assets - improvements to preexisting assets ### 43.3 Native Alpha™ capture without extraction Encoding a professional's or portfolio company's Native Alpha™ should not become a mechanism for inappropriate affiliate extraction. Foreground intellectual property developed and paid for by a portfolio company should normally belong to that company. Background intellectual property needed from a parent or affiliate should be identified explicitly and licensed on durable, financeable terms. Related-party rights structures should survive scrutiny from an independent investor, board member, acquirer, customer, licensee, insurer, or limited partner. ### 43.4 Vendor and model-provider rights Research: - whether prompts and outputs are retained - whether data is used for training - whether derived artifacts are portable - whether the organization can reproduce the worker on another model - whether vendor terms grant unexpected licenses - whether audit evidence remains available after termination - whether the vendor can change material dependencies unilaterally ### 43.5 Knowledge capture policy Not every piece of Native Alpha™ should necessarily be encoded. Possible dispositions: - keep tacit and human-held - document internally as trade secret - encode in worker contract - patent where appropriate - publish strategically - license narrowly - share with portfolio company - keep model-agnostic for portability The rights decision should follow the commercial purpose, not a reflex to warehouse intellectual property. ### 43.6 Rights provenance as a deployment gate For strategically important workers, unresolved rights should block broad deployment just as unresolved safety or liability questions would. The research should develop a machine-readable Rights Provenance Record so downstream GitHub specs and harnesses know which assets can be used, modified, shared, trained on, licensed, or commercialized. ## 44. Relationship to other research areas ### Reimagining with AI Unbundling Work supplies a concrete method for asking which forms of work become possible when historical labor constraints are removed. ### Zero Neo Once work is decomposed and encoded, some standalone software categories may no longer be necessary. Existing platforms may be sufficient coordination surfaces. ### AI Workforce™ AI Workforce™ is the recomposed operating model produced by the research. ### Labor as Code™ Labor as Code™ is the machine-readable and governable representation of the work. ### Native Alpha™ Native Alpha™ determines whether unusual understanding of work produces a defensible capability, product wedge, or commercial advantage. One potentially unusual source of Native Alpha™ is the ability to see liability structures as information about work design. Professional insurance, indemnification, bonding, and similar instruments can reveal where society already recognizes consequential failure. The opportunity is not to arbitrage insurance. It is to identify where risk-bearing professional bundles contain separable work that can be made more observable, controllable, and economically useful. These relationships should be explicit, but Unbundling Work should stand on its own. ## 45. Recommended initial research pillars ### Pillar 1: Unbundling Work, Credentials, and Liability Study how roles, credentials, workflows, organizations, authority, and liability can be decomposed into work units, evidence, controls, accountability, and outcomes. Primary healthcare question: > Which parts of a healthcare professional's workload genuinely require that professional's credential, judgment, authority, accountability, presence, or liability-bearing responsibility? Secondary question: > What is the minimum necessary credential exposure required to safely produce the accepted outcome? ### Pillar 2: Labor as Code™ Develop the representation, schemas, contracts, tests, and versioning disciplines required to make work inspectable and programmable. Primary question: > Can real-world work be represented precisely enough to be safely assigned across humans and machines? ### Pillar 3: AI Workforce™, Native Alpha™, and Recomposition Study how decomposed work is allocated and recomposed into effective human-machine operating systems that concentrate differentiated capability rather than merely substitute for labor. Primary questions: > What is the best division of labor when people are treated as scarce capability rather than expensive headcount? > How do we reduce unnecessary credential exposure while increasing Native Alpha™ Density and leverage? ### Pillar 4: Operational Truth™, Risk, Liability, and Outcomes Establish the evidence, trust, authority, liability, risk-financing, and verification layers needed for consequential work. Primary question: > How can an organization prove that its human-machine workforce is operating on reality and producing acceptable outcomes? "Zero Meaningless Work" should sit above these pillars as the governing ambition. Five cross-cutting disciplines should apply to every pillar: 1. Incentives and Work Creation: explain why the work exists and whether the cause should be removed rather than automated. 2. Transition and Capability Formation: design brownfield migration while preserving the learning loops that create future expertise. 3. Systemic Risk, Security, and Resilience: manage correlated failure, adversarial behavior, dependency concentration, fallback, and recovery. 4. Causal Evidence and Distributional Effects: prove impact and identify where work and burden move, including to patients and caregivers. 5. Rights, Knowledge Capture, and Ownership: ensure encoded capability, data, and Native Alpha™ remain under clean and financeable control. ## 46. Final framing The research should remain aggressively practical. Do not begin by asking whether AI will replace a profession. Begin by observing the work. Ask why the work exists, who imposes it, who pays for it, and who benefits from it. Unbundle it. Find the evidence. Identify where authority, accountability, and liability actually live. Treat liability-bearing credentials as a signal to inspect the bundle, not as proof that the bundle must remain intact. Determine where human capability materially changes the outcome. Separate ordinary professional competence from genuine Native Alpha™. Test that differentiation rather than assuming it from title, credential, compensation, or prestige. Use the minimum credential exposure consistent with a safe accepted outcome while concentrating Native Alpha™ where it materially improves the outcome. Determine whether the supposedly routine work is also part of how future experts learn. Preserve or replace the learning loop before removing it. Encode the rest clearly enough to assign and test. Establish clean rights to the knowledge, data, workflows, and intellectual assets being encoded. Start machines with narrow authority. Measure real outcomes. Increase authority only when evidence supports it. Stress the system for correlated failure, hostile inputs, outages, degraded data, and recovery. Use causal evidence before claiming the new design produced the improvement. Watch for work that merely moves elsewhere, especially onto patients and caregivers. Then recompose the operation around the best available combination of humans, AI, software, and machines. The research area can therefore be summarized in one sequence: > Observe the work. Understand why it exists. Unbundle it. Understand the risk and incentives. Find the Native Alpha™. Preserve the learning that creates future capability. Encode the commodity work under clean rights. Recompose around differentiated capability. Test the outcome. Stress the system. Prove the effect. Compound what works. The desired future is not Zero People. It is a world in which people are no longer routinely consumed by work that never needed a person in the first place.