In short: Judgment layers will work first where decisions are frequent, similar and graded quickly, and where someone will later have to defend them. Map your decision classes on those axes before choosing where to build. Everywhere else, judgment is a practice before it is a product.
Key points
- AI is shaping decisions in screening alerts, financial-crime investigation, loan early warnings, credit memos, support and code review.
- Figures such as one bank drafting a large share of each credit memo with agentic tools are vendor-reported.
- Four conditions: volume of similar decisions, expert judgment at exceptions, timely outcomes, and someone who must defend the decision.
- Build the learning layer only in the high-volume, short-lag, defended quadrant.
- For low-volume strategic decisions, use a disciplined record with AI drafting and surfacing precedent.
Look at where AI is genuinely shaping decisions in 2026, rather than drafting documents about them. Screening and compliance alerts, where data providers report AI cutting false-positive alerts sharply in testing. Financial-crime investigation, where one vendor's agentic workforce is used by more than 650 financial institutions. Early-warning signals on loan books. Credit memo preparation, where one large bank is reported to draft a large share of each memo with agentic tools. Support resolutions. Code review.
These are vendor-reported figures, and they should be treated as such. But the pattern in where they appear is consistent.
Four conditions
Volume of similar decisions. Enough cases of the same kind that exceptions recur. At tens of thousands of comparable cases a year, rare variations cluster and patterns surface in weeks.
Expert judgment at the exceptions. The rulebook handles most cases, and a skilled person decides the rest. That person's decisions are what the system learns from.
Outcomes that arrive soon enough. A false positive is known in days, a credit outcome in months. Both are short enough to close the loop.
Someone who must defend the decision. A regulator, a budget holder, an investment committee. This is what makes the record worth paying for, not just possible.
Map before you build
Take your organisation's decision classes and place each on two axes: annual volume of comparable decisions, and time until the outcome is known. Mark the ones where someone must later defend the call.
High volume and short lag, with a defender: that is where a learning system can be built and evidenced now.
High volume and long lag: possible, with care, using leading indicators as interim outcomes. Choose those indicators deliberately, because a system will learn to optimise whatever proxy you give it.
Low volume and long lag: strategic decisions, partnerships, market entries. Patterns will not form. What works here is a disciplined record that a person reads, with AI drafting, checking and surfacing precedent rather than learning.
Below the floor is a different object
The mistake is to treat the low-volume case as a smaller version of the same product and to promise it the same learning. It is a different object, with different value: not that the system gets smarter, but that the organisation's own reasoning survives and is argued with before the next call.
In practice
- List your top twenty decision classes and estimate annual volume and outcome lag for each.
- Mark the classes where someone must later defend the decision.
- Build the learning layer only in the high-volume, short-lag, defended quadrant.
- For the rest, deploy the record and the review discipline, and say clearly that it is a practice.
What would prove this wrong. A judgment system showing measurable learning in low-volume, long-lag, heterogeneous decision classes, under independent evaluation. Then the volume floor is lower than it looks.
Common questions
Where is AI most useful for business decision-making?
Where decisions are frequent, similar and graded quickly, and someone must later defend them, such as screening and compliance alerts, financial-crime investigation, early-warning signals on loan books, and credit memo preparation.
Can AI learn from strategic decisions like partnerships or market entries?
Patterns will not form in low-volume, long-lag decisions. What works there is a disciplined record that a person reads, with AI drafting, checking and surfacing precedent rather than learning.
How should a company decide where to build an AI judgment system?
List your top twenty decision classes, estimate annual volume and outcome lag for each, and mark where someone must defend the decision. Build the learning layer only in the high-volume, short-lag, defended quadrant.
What is the risk of using leading indicators to train AI on long-lag decisions?
A system will learn to optimise whatever proxy you give it, so choose the leading indicators deliberately.