Regulatory reporting.
It reconciles.

Regulatory reporting teams rarely struggle with the arithmetic. They struggle with the fact that a figure in a submission has to be reconciled to a source system that was not designed to produce it, that definitions differ between the regulation and the general ledger, and that when a number moves between periods somebody has to explain why in terms a supervisor accepts.

The recurring work is therefore: chase the break, find the explanation, evidence it, and write the commentary. All of that is document and system archaeology, and it is done under a deadline every quarter by people who could be doing analysis instead.

An AI system is well suited to the archaeology and badly suited to the arithmetic. That division is the whole architecture, and inverting it — letting a model compute a reportable figure — is the failure mode we design against explicitly.

Every AI-produced element is labelled as such, with its sources attached, and is reviewed by a named person before it forms part of a submission. That review is the control, and pretending it can be removed is how a reporting system becomes a finding.

The lineage runs in both directions. From any figure you can reach the source records; from any source change you can find the figures it affects. Building this is unglamorous data engineering and it is what makes the AI layer useful rather than decorative — an agent that cannot show its working produces text nobody is allowed to rely on.

And the model version is pinned and recorded per reporting period. If a supervisor asks in five years how a particular explanation was produced, the answer has to be reproducible, which means the system that produced it has to be reconstructable.