A problem pattern in credit and compliance, and an approach built for it

The problem

In several segments I work in, some decisions have to meet three tests at once:

• Rule-bound. The answer is set by written rules or contract terms, full of definitions, cross-references, and amendments, often across long documents.
• Repeatable. The same inputs have to give the same answer every time.
• Explainable. Someone outside the firm, such as an examiner, auditor, agent bank, or investor, will ask how the answer was reached.

Some examples:

• Regtech: which rules apply to this bank, and which control meets each one.
• Syndicated lending: whether a borrower passes a covenant test under a credit agreement and its amendments.
• Private credit and funds: which side-letter terms a fund owes each investor, and which loans count toward a borrowing base.

Outside my segments, the same shape shows up in insurance claims against policy wording, in tax treatment, and in trade documents checked against a letter of credit.

Large language models and machine learning are powerful, and both fall short here. Language models read the documents well, and recent tests show them catching up on repeatability, but they're still weak on contract-bound reasoning and on showing how they reached an answer. Machine learning needs labeled outcomes and returns a score. Both help, and neither meets all three tests on its own.

The approach

Neuro-symbolic AI is a class of technology designed for problems like these. It combines learning (reading, recognizing patterns) with symbolic reasoning (explicit knowledge and logic that can be checked). The aim is answers that are flexible, exact, and explainable at once.

The target state is one coherent system, where learning and reasoning work together rather than side by side. That isn't available yet. A few production systems in finance are sophisticated approximations:

• Droit, bought by FIS in March 2026, encodes regulation as executable logic (A-Team Insight).
• Nammu21, now built into Finastra's Loan IQ, turns credit agreement language into computable business logic and authoritative reference data (Nammu21; Finastra).

The symbolic side is hard to build. It takes domain ontologies, curated rule and contract content, and people who know both the domain and the technology. That specialized IP is a moat for the providers who have it.

The counterweight: models keep getting better at reasoning on their own, and any structure has to be kept current as rules and contracts change.

Open question: which decisions in your work fit this pattern, and are you evaluating approaches built for it, or fitting a general model to it?