Traceability Is a Feature: Designing Citations for AML Recommendations

A recommendation that cannot show its sources is a guess with good manners. If you build investigation tooling, here is how to make traceability a feature, not an afterthought.

The principle

Investigation AI sits at the point an analyst opens the case. Its job is to assemble evidence and propose a next step. Our overview of investigation AI versus detection AI explains the placement. The design question is how each statement connects back to the record it came from.

Five design rules

1. One source per claim. Each sentence in a summary links to a record: a transaction, a KYC field, a prior case or a policy paragraph.
2. Show retrieval, not just output. Let the analyst open the retrieved records. This is the idea behind retrieval-grounded AI that generates evidence.
3. Mark inference. Label suggestions and assumptions differently from facts.
4. Keep the human step visible. Show who accepted, edited or rejected a recommendation, and when. FATF notes most tools still need people to use and review them.
5. Log versions. If a summary is regenerated, keep both versions and the reason.

A caution on regulation

Don’t claim the EU AI Act makes all AML AI high-risk. Annex III names credit scoring, and transaction monitoring is carved out of that credit clause. Classification turns on use, so ask your legal team, then design for explainability anyway, as the FCA’s expectations on AML AI models suggest.

Conclusion

Traceable output makes review faster, supports audit and builds the organisational memory that teams otherwise lose. We build for this at TechnoXander. If you’ve shipped citations in a regulated tool, what did analysts click first?