Administrative Denominator Reduction

Use this skill when a practice wants to cut what a Medicare between-visit program costs to run without cutting care, decide what to automate first, find out whether it really needs another coordinator or nurse, or stop a monitoring or alert stream from…

This skill is Labor as Code™ for the cost-of-delivery method in Getting Paid to Do the Right Thing. Labor as Code is a skill that encapsulates a person's knowledge, judgment, taste, and skill in durable, inspectable, versioned form. The know-how is the book's economic engine: the payer fixes the payment per patient-month, the practice controls the admin minutes it spends earning it, and software changes the denominator when it takes over the remembering, counting, and checking while people keep the conversations and decisions.

The method spans five chapters. "A Billing Code Doesn’t Tell You How to Run the Care" names the Care-Delivery Operating System and its nine components, and defines the Administrative Tax. "Spreadsheets Hide the Admin Minutes That Make Programs Lose Money" defines the Administrative Denominator, admin minutes per patient-month, and the two-week study that measures it (Tables 6.1–6.3). "Software Should Track the Work and People Should Make the Decisions" gives the three-way sort and Riverbend's first three automations. "Stop Building Another Inbox" gives the Signal-to-Attention test and the two monitoring numbers. "The Care Team Is an Operating Specification" gives the caseload formula. Read them for the argument and the Riverbend, Lakeshore, and Northfield cases. This skill applies the same steps to the user's own program and never reuses the cases' minutes as the user's.

The skill measures; it doesn't estimate. Estimates of admin work come in low, because the minutes nobody planned for are the minutes nobody remembers spending. At Riverbend, a careful nurse's estimate of 12 admin minutes per patient-month measured 20. The skill can help itemize an estimate, and it labels that as an estimate until a two-week log replaces it.

The skill never moves a care minute to make a number look better. The goal is to turn admin minutes into care minutes. Any scenario, caseload, or projection that lowers care minutes is wrong, and the skill throws it out.

The skill takes totals, not logs. Study logs carry study IDs and times. The user totals minutes by task code and role; only the totals come into the session. No identifiable patient data, and no pasting of records into a tool without a business associate agreement.

Before running

1. Name one program. The study, the sort, and the caseload all run per program.
2. Ask what prompted this. A hiring request, a vendor demo, a program losing money, a nurse staying late, an alert load. Write it down; Step 7 answers it.
3. Ask who touches the program. Every role: nurses, medical assistants, front desk handling consent, physicians on escalations, the administrator reconciling remittances. Leave out an outside billing company's staff; its fee goes into Practice Math separately.
4. Ask for the enrolled count for the period. Every per-patient number here divides by enrolled patients, never by served or billed.

Step 1: Map the nine components

For each component of the Care-Delivery Operating System, ask where it lives on an ordinary Tuesday: software, a spreadsheet, a person's head (name the role), or nowhere. Not where it should live, and not where the EHR vendor says it could.

1. Eligibility logic
2. Enrollment and consent
3. Care plans
4. Work queues
5. Escalation pathways
6. Documentation
7. Billing validation (the pre-claim check)
8. Audit evidence
9. Quality and outcome review

For every component in a head or nowhere, say what would break in a month when that person is out, how anyone would notice, and rank them by risk. Don't accept "in the EHR" as safe on its own; a care plan nobody updates lives in the EHR and still doesn't work. Write next to each component the one person who would notice if it stopped working. A component with no name is a failure waiting for a date.

Software is the right home for anything that is mostly memory: eligibility, queues, time capture, pre-claim checks, audit trails. People are the right home for judgment: the consent conversation, the escalation decision, whether a patient is better.

Step 2: Itemize the Administrative Tax

Ask the person who does most of the program's work to list every task a patient would not recognize as care, with rough minutes per enrolled patient-month (or per episode, for an event program such as transitional care). Don't argue with the estimates yet. Total them into monthly hours.

Keep the Administrative Tax distinct from the Workflow Tax in Doing the Right Thing Is Not a Business Model. The Workflow Tax is unbilled labor a product imposes. The Administrative Tax is created by the program itself and exists before any software arrives. A product can lower one and add the other; count both when judging software.

For an event program, check the first task hardest. For transitional care, "finding out the patient was discharged" decides whether any of the rest happens, and a discharge the practice never learns about never shows up as a denial or as admin minutes.

Step 3: Run the two-week admin-minutes study

Give the user the method exactly as the book states it:

1. Pick one program and one two-week period without holidays.
2. List every person who touches the program; give each a log sheet.
3. Give every enrolled patient a study ID, a number and never a name. Keep the key apart from the logs.
4. Print the task codes at the top of every sheet.
5. When a person starts a program task, they write the start time, the study ID, and the task code.
6. When they finish, they write the minutes, rounded to the nearest minute. A voicemail is one line, even at one minute.
7. An interruption ends the line; start a new line on return.
8. At the end of each day, each person spends five minutes filling gaps.
9. At the end, total minutes by task code and by care or admin.
10. Multiply the two-week totals by 2.17 to get a month, and divide by the number of enrolled patients.

The task codes:

The care-or-admin test is one question: if this minute disappeared, would the patient notice? That puts writing the note on the admin side: necessary, and the first minute worth shrinking. Whether a minute counts toward a billing threshold is a separate compliance question for the compliance adviser.

When the totals come back, compute admin and care minutes per enrolled patient-month and show the arithmetic so the user can redo it by hand. Rank the admin codes by monthly minutes. Compare each code with the Step 2 estimate, line by line. Report two more things if the data allows: the admin minutes absorbed by enrolled patients never reached, and which person pays most of the tax.

If the study hasn't been run, stop here for any hiring or buying recommendation and say so: the correct next step is the study.

Step 4: Sort every task into three bins

List the program's tasks with their measured minutes and a care or admin label. Put each task in one bin:

• Machine does it: remember, count, schedule, route, match, check completeness, flag missing data, stop an incomplete claim.
• Machine assists: summarize, draft, classify, suggest, with a person reading, correcting, and owning every output before it reaches a patient, a chart, or a claim.
• Human decides: diagnosis, treatment changes, whether a pattern matters, the consent conversation, anything a license or judgment is for, and every escalation.

Settle each task with the book's three questions, in order:

1. Is the task remembering, counting, or checking against a rule? Then a machine should do it.
2. If the machine got it wrong, would a person catch the error before it reached a patient, a chart, or a claim? Then it can be machine-assisted.
3. Does the task need a license, a judgment about this patient, or a conversation the patient has to trust? Then a human decides. A yes here wins over the first two answers.

Flag every task whose bin depends on how the software is designed. The test is what happens to the patients the software doesn't show you. A list of everyone who owes the program something, sorted by how overdue, is scheduling. A risk score that quietly drops the bottom of the panel is a clinical judgment nobody licensed made. Any clinical judgment in "machine does it" is misplaced, and any care minute moved into a machine bin is wrong.

Then project: machine-does tasks fall to the minutes a person spends on exceptions, machine-assists tasks fall to review time, care minutes stay where they are. Label the result "projected." A sort ranks targets; it doesn't predict what real software does with real patients. Riverbend projected 10 admin minutes and measured 14 after its changes.

Name both automation failures when the user's sort shows them. Irresponsible automation is a human-decides task in a machine bin. Under-automation is a machine-does task in a human bin, such as a nurse counting minutes or re-keying readings. The fix for both is the sort.

Step 5: Apply the Signal-to-Attention test to any data stream

Skip this step if the program has no device readings, portal messages, or AI alerts reaching clinicians. Otherwise, ask one question of each item type: does this item need a human decision today?

• If yes, it goes to the right person with context and a deadline.
• If no, the machine handles it (a reminder, a receipt), summarizes it (a month of in-range readings becomes one line), or holds it (a single high reading waits for the next one).

Rules that don't bend:

• An item reporting new symptoms always goes to a person, whatever its action rate.
• Missing data is not a clinical decision. It goes to a coordinator, never to a physician.
• Clinicians write every threshold and sign it. The skill proposes the shape of a rule and leaves every clinical number blank. Replace any population default with a blank for the patient's personal range.
• Every tier has an owner and a deadline. An escalation with no deadline is a message.
• Nothing reaches a physician as raw device data; escalations come through a nurse with a summary and a specific question.

Compute the two numbers from the user's counts: alerts per clinician-hour (items reaching a nurse or physician ÷ hours they spent on the program) and percent of alerts that led to an action (a call that changed something, a treatment change, an appointment, or an escalation ÷ all items reaching a person). A click on "reviewed" is not an action. The fix for a low action rate is a better trigger or a better cohort, never faster clearing. If a clinician can't name a reading that would change what they do for a patient, ask whether that patient belongs in monitoring.

Step 6: Size the caseload from minutes

Only after the study:

caseload per coordinator = monthly work minutes ÷ (care minutes + admin minutes)

Ask for the coordinator's real monthly work hours after time off, meetings, and training. If the user doesn't know, the book reports about 130 hours (7,800 minutes) in most practices the author has seen; label that as the book's observation, not the user's number. Recompute at three lower admin-minute values and show how many coordinators a target enrollment needs at each. Care minutes stay fixed. A caseload that works only by cutting care means the program doesn't work at that staffing level.

Check physician minutes the same way if the user has them. If the escalation rate is well above what the user's own clinicians expect (the book's illustrative figure is about 8 percent of enrolled patients a month), check the escalation rule before checking the physicians; coordinators with no written rule escalate everything.

Step 7: Conclude with targets, destinations, and the hiring answer

Close with a short recommendation, not a table dump:

• The three largest machine-does or machine-assists admin tasks, as the first automation targets, in order, with measured minutes. For a solo or small practice the book's first target is the worklist, because memory fails first. For a small group it's the same time capture and pre-claim check at every site, so sites count the same way. For a large organization, sort the tasks before automating a central department's workflow; automating an expensive structure makes it a slightly cheaper expensive structure.
• What to try before buying anything. Check whether the EHR or existing contract already includes an unconfigured module. Automate only what someone has watched a person do by hand. Test any tool against a month the practice already reconciled by hand.
• What not to automate yet, with the reason: anything that hides patients from the list, outreach drafts nobody has read, ambient notes before the practice has checked drafts against real calls.
• Where the freed minutes go. A saved minute has no cash value until someone decides where it goes. Ask the partners to decide in advance (for example, offering the program to eligible patients never asked). If the minutes are simply absorbed, say plainly that the program's economics won't change.
• The hiring answer. At the measured admin minutes, does the caseload formula support a hire? The book's rule: measure admin minutes before hiring anyone, fix the most repeated admin tasks, then rerun the formula and hire against that number. A new person added to a program at twenty admin minutes per patient-month mostly buys more spreadsheet.

Where the choice is genuinely open (two targets of similar size, one software and one process; or a design-dependent task where the vendor's answer isn't known), lay out both with their reasoning and what would decide between them. Where no study has been run, say the evidence can't support a hiring or buying decision and name the two-week study as the step that would.

Checks before calling the work done

• Every per-patient number divides by enrolled patients, billed or not.
• Care minutes are identical in every scenario and caseload.
• Estimates and measurements are labeled as such, and projections say "projected."
• No clinical number was set by the skill. Every threshold is a blank for a clinician.
• Every component and every tier has a named role.
• The conclusion names three targets, where freed minutes go, and a hiring answer or the measurement it waits on.

What this skill does not do

It doesn't recommend a vendor or product. It doesn't set clinical thresholds, decide escalations, or decide which patients can wait. It doesn't decide whether a minute counts toward a billing threshold; that is a compliance question. It doesn't borrow Riverbend's 20, 14, or 10 admin minutes, or the book's $0.75 a minute, as the user's numbers. And it doesn't call a program efficient because care minutes fell. When the minutes haven't been measured, the correct output is "run the two-week study first," not an automation plan built on a guess.