In short: At 150 decisions a year, all different, no model will learn your judgment, and pooling across companies does not rescue it. What still works is narrower and still valuable: your own prior reasoning, recorded properly, plus public base rates, read and argued with by a person.
Key points
- Low volume, high heterogeneity and slow, noisy outcomes defeat pattern learning; waiting longer does not fix it.
- Pooling fifteen companies' decisions gives 2,250 records, but the comparable sample within any segment may be in single digits.
- Pooling raises consent and ownership problems, and in narrow markets de-identification is not anonymity.
- At this scale AI should draft and check decision records, not learn from them.
- Run a 30-minute weekly decision review, a quarterly falsifier sweep and an annual 'what do we keep getting wrong' read.
The enterprise designs that work rest on volume. Enough similar cases that exceptions recur and patterns form. A large financial institution can run tens of thousands of comparable credit or compliance decisions a year through the same workflow. At that density, rare variations cluster, and a system can learn from the experts who handle them.
A founder-led company, or a small specialist team inside a large one, has the opposite profile.
Different physics
Perhaps 150 material commercial decisions a year. A pricing exception. A partnership choice. The sequence of a launch. Whether to hire a first salesperson. They share almost nothing with each other. Each one is expensive. One wrong call can be the company, or the year.
And the outcomes are slow and noisy. A partnership is graded in eighteen months, and the grade is confounded by everything else that happened in those eighteen months.
That combination, low volume, high heterogeneity, slow and noisy outcomes, defeats pattern learning. It is not a matter of waiting longer.
Why pooling does not fix it
The obvious response is to pool: fifteen companies times 150 decisions is 2,250 records. It is not 2,250 samples. Within any segment comparable enough to learn from, same stage, same model, same market, same kind of decision, the effective sample may be in single digits.
Pooling also raises problems that matter more than statistics. Did each client consent to their decisions informing someone else's? Who owns a pattern derived from several clients? What happens when it helps a competitor? And in narrow markets, a pattern assembled from three accounts can be recognisable without a single name attached. De-identification is not anonymity when the market has twenty players.
What does work at this scale
Your own prior reasoning. The single most transferable thing at low volume is the company's own reasoning on its nearest comparable decision. That needs one organisation, no consent architecture, and a record kept with discipline: evidence, alternatives, a falsifier, the outcome.
Public base rates. Where a decision has a public reference class, use it. Forecasting research has long shown that starting from the outside view, how similar efforts usually turn out, beats starting from the specifics of your own plan. Some domains have rich public outcome data: launches, listings, funding rounds, programme results. Base rates are a weaker claim than patterns, and a more honest one.
AI as drafter and checker, not learner. At this scale the model's job is to draft the entry from a call or a memo, propose the alternatives nobody named, suggest a falsifier, flag contradictions with earlier entries, and surface the nearest prior case before the next meeting. A person reads it and argues with it.
A rhythm that fits
A 30-minute weekly decision review for anything material. A quarterly sweep of every falsifier whose date has passed. An annual read of the whole record by the leadership team, asking one question: what did we keep getting wrong?
That is not a scaled-down enterprise brain. It is the one layer of one that works at any size, because it does not depend on a model learning from repetition.
In practice
- Record every material decision with evidence, alternatives, falsifier and outcome. Aim for under five minutes each.
- For each decision class, find one public reference class and its base rate.
- Before each material decision, read the nearest prior case aloud.
- Run the quarterly falsifier sweep and the annual "what do we keep getting wrong" review.
What would prove this wrong. Pooled cross-company models producing demonstrably better decisions for low-volume, heterogeneous classes than a company's own disciplined record plus public base rates, under independent evaluation.
Common questions
Can AI learn a small company's decision-making judgment?
Not at around 150 material decisions a year that share almost nothing with each other. Low volume, high heterogeneity and slow, noisy outcomes defeat pattern learning.
Does pooling decision data across companies help train a model?
No. Within any segment comparable enough to learn from, the effective sample may be in single digits, and pooling raises consent, ownership and recognisability problems.
What should a small team use instead of a learned decision model?
Its own prior reasoning on the nearest comparable decision, recorded with evidence, alternatives, a falsifier and the outcome, plus public base rates for the decision's reference class.
What role should AI play in decision records for a small company?
Drafter and checker: draft the entry, propose alternatives nobody named, suggest a falsifier, flag contradictions with earlier entries and surface the nearest prior case. A person reads it and argues with it.