I study one claim: AI can now assemble the dashboard and the digest people used to build by hand, which adds information faster than people can decide. It helps only if it changes what people decide, start, stop or measure. This pillar asks whether a Decisions & NBA Board does better: a short queue of decisions with evidence, options, an owner and a check on the result.
I have written the concept, a draft specification (0056-decisions-nba-board) and a milestone paper with eight hypotheses and a 30-day protocol. I have not run it in a live organization.
I own the products that would benefit from this idea
I own Intellectual Frontiers, which owns the Opsfolio products, and Netspective Communications, which builds and operates them. A Board that runs on operational truth is good for those businesses, so I want it to work. Read every claim here with that in mind. A test by a team that does not own these products counts for more than anything I write.
The record already holds the parts of this idea
The harness has a note on operational truth, a note on decision cards in place of alerts, a note on the dashboard as a symptom, and a pillar on dashboards that decide. Each covers one piece. None covers the whole loop from evidence to decision to authorized action to measured result, and none says when the right answer is to do nothing. The new work here is the whole loop. It cites the earlier records and does not redefine them.
The terms of the method are defined once, in the firm’s specification
The public ontology holds the classes a Board uses: goal, outcome metric, evidence record, decision candidate, next best action, commitment, approved action, acceptance evidence, observed result and learning record. A recommendation is never recorded as a decision, because only a named person’s choice is a decision. The specification is a draft, and no one but me has used the terms.
A smaller queue with a closed loop beats a better briefing
I think a Board of at most five decision candidates, each with cited evidence, a do-nothing option, a named owner and a dated check on the result, will shorten decision time and improve target outcomes more than an AI-prepared dashboard or digest of the same sources, with no rise in errors. I think the effect comes from the closed loop and the suppression rules, not from the model.
-
Teams on the Board show no shorter decision time or no better target outcome than teams on the AI-prepared dashboard or digest over the same period.
-
Board items are adopted at a rate no higher than the dashboard or digest’s recommendations.
-
The Board’s gains disappear when the reporting-only teams get the same owner, baseline and review cadence.
Starting with a routine message to named readers gives a shorter, better-aimed Board
I think a team should start by sending the Board as a routine email, one message for each named person or distribution group, before it builds a shared view. A message has a size, so the team must choose the one to five items worth a reader’s attention. A message has a recipient, so the team must ask what that reader can decide, and a finance lead, a product lead and an engineering lead each get their own decisions and next best actions. I have seen this force the choice in my own companies. I have not measured it, and the paper’s seventh hypothesis tests it against a shared view.
-
Teams that start with messages carry no fewer cards per message, send no more audience-specific cards and see no more decisions after each message than teams that start with a shared view.
-
Decisions after a message are no more frequent than opens and clicks alone would predict.
Evidence stays grounded when it is derived by the most deterministic tier that can do the job
I think too many people ask a language model for the data itself, and that makes the evidence less reproducible. I think a team should derive each figure with a query where a query can, add explicit rules and a typed-decision model such as Jev where queries run out, and use a language model only for what none of those can do. Each evidence record names its tier and version. I have seen language-model-first evidence fail in my own companies and have not measured it, and the paper’s eighth hypothesis tests it.
-
Boards that start from a language model show the same share of exactly recomputable figures, the same error rate and the same reversals as Boards that derive evidence by query first, at lower cost.
-
A typed-decision model’s stated confidence does not match outcomes once the team compares them.
I do not know whether the effect survives outside a willing team
I do not know whether the effect survives outside a team that already has clean data and a willing owner. I do not know how much it costs to keep operational truth current. I do not know whether decision makers stay accountable or begin to defer to a ranked list. The pilot in the milestone paper is built to find out. What would settle each question is evidence from a permission-cleared team: logged owner effort, the decision log, and outcome records against baseline. A null result will be published.