Credit decisions live in the exceptions

Every lender has a credit policy: leverage limits, minimum coverage ratios, collateral requirements, concentration caps and approval authorities. Encoding the policy itself is not difficult. Most of it already sits in a document and some of it in a rules engine.

The hard part is that a meaningful share of committee time goes to cases that do not fit the policy. A borrower exceeds a leverage limit but has a sponsor with a strong record. A covenant is breached for reasons the relationship team considers temporary. A concentration cap is reached in a sector the bank wants to grow. These cases are where judgment, and risk, concentrate.

When a committee approves an exception it is making several judgments at once, most of which never reach the approval record in a structured form.

Ask an experienced credit officer why a deal was approved and the answer often begins with a reference to an earlier case: we did something similar for a borrower in the same sector two years ago, and here is what we learned. That chain of reasoning is the bank's real credit culture.

It is also fragile. It lives in the memory of a few senior people and in approval memos written for a committee, not for retrieval. When those people move on, the bank keeps its policy but loses the judgment that made the policy workable.

Generic AI makes the problem more visible, not less. A model can summarise the policy and the file. It cannot know that the committee has repeatedly accepted a particular mitigant for a particular type of sponsor, or that it stopped accepting it after a specific loss.

The alternative is to capture exceptions as structured knowledge rather than narrative. Each approved exception links to the rule it departs from, the evidence relied on, the mitigants accepted, the authority exercised and the cases it was compared with. Over time this becomes a precedent base that both people and AI can query.

With that base in place, a new request can be analysed against the institution's own history. The question becomes not only whether the deal meets policy, but how this bank has treated similar departures, what conditions it attached and how those loans performed.

Start with one portfolio and one recurring exception type, such as leverage above policy for sponsor-backed borrowers. Collect the last two to three years of approved and declined requests, extract the reasoning with short reviews from the credit officers involved, and test whether the captured knowledge reproduces their decisions on held-out cases. The result is a working precedent base and a clear view of where policy and practice have drifted apart.