An AI Mission for Regulatory Reporting

A regulatory report is a sentence an institution says out loud about itself, and a named human being puts their signature under it. What almost nobody builds for is the moment, quarters later, when someone asks how that sentence came to be true.

The request usually arrives without drama. A supervisor, an internal auditor, or simply a newly arrived head of finance asks a mild-sounding question about one figure in a return that went out two or three periods ago. The team pulls the pipeline, re-runs it, and the number comes back the same, which feels like the end of the matter until the follow-up arrives: on what basis was it that number and not the other defensible one? Which extract of the source system, taken at which cut-off, with which late adjustments included, and who decided that the boundary cases went into this bucket rather than that one? At that point the reconstruction begins in earnest, and it runs through a spreadsheet on a shared drive, a mail thread with four people on it, a reconciliation tab whose formulas nobody has looked at in a year, and the memory of an analyst who has since moved teams. The number was never in doubt; the account of the number turned out to be the fragile thing, and it was fragile because nobody was ever asked to produce it as a deliverable.

The reflex is to file this under housekeeping — a documentation habit that slipped, fixable with a better folder structure and a stricter template. That reading is comfortable and wrong. What the episode reveals is a mistake about what the reporting function is for, built into the way most institutions organise the work. They build reporting programmes to produce reports, because the report is the artefact that leaves the building and therefore the thing that gets project-managed, calendared, and celebrated when it goes out on time. Everything else — the trail of inputs, versions, interpretations, exceptions and decisions that produced it — is treated as residue, the sawdust left over from making the furniture. Residue is exactly what you cannot produce on demand eighteen months later.

What is submitted is a claim, not a calculation

It helps to be precise about what a regulatory report actually is, because the arithmetic framing hides it. A thermometer reading is a measurement, true or false independent of anyone's intent; a regulatory figure is not that. It is an assertion the institution makes about itself, prepared under definitions that require interpretation, and attested by a person accountable for its fairness and accuracy to the best of their knowledge. That attestation is the load-bearing element of the whole structure. The same digits in a spreadsheet cell are an output; the same digits under a signature are a position the institution is answerable for, and the difference between those two things is not mathematical at all. It is a difference in who has staked something.

Once you see the report as a claim rather than a computation, the interesting question stops being whether the number is right. In most institutions the number usually is right, because arithmetic is the easy part and there are three layers of checking on it. The interesting question is whether the basis for the claim can be reconstructed on demand, by someone who was not there, without the cooperation of the people who were. This is not a bureaucratic nicety. An attestation is a statement about knowledge, and a person who signs is implicitly asserting that they had a reasonable basis for doing so. A basis that cannot be reconstructed is, uncomfortably, a basis that may never have existed in any durable sense — what existed was confidence, which is a feeling, and feelings do not survive a personnel change.

Reconstruction is hard because reporting figures are unusually dense in judgment for numbers that look so clinical. Somewhere in the chain, someone decided which entities fell inside a perimeter and which sat just outside it, how an unusual instrument or counterparty relationship should be classified when the definition genuinely admitted two readings, which of two systems was authoritative on a day when they disagreed, and whether a late adjustment belonged to the closed period or the open one. Each of these calls is defensible on its own; several may have been made in the last seventy-two hours before submission by people who were tired. Taken together, they are where the report actually lives — and they are precisely the layer that conventional pipelines throw away, persisting the output while the reasoning evaporates.

Lineage is the deliverable; the report is the by-product

Invert the priority and most of the difficulty dissolves. If the thing you are building is a report, then lineage is overhead — a tax paid grudgingly, after the fact, in the form of an evidence pack assembled from screenshots and recollection once the pressure is off. If the thing you are building is lineage — a durable, queryable record of every input and its version, every rule as it was applied and to what population, every interpretive choice with its author and its stated rationale, every exception with its trigger and its disposition — then the report is very nearly free. The number is simply the terminal node of that record, rendered into whatever shape the return requires. You do not assemble evidence for the figure; the figure is a summary of evidence you already hold.

You can tell which way round an institution has it by where the investment goes. Programmes built around the report invest in the close calendar, the reconciliation checklist, and the sign-off meeting, all oriented toward getting a file out of the door by a date. Programmes built around lineage invest in capture at the moment of decision, which is a less satisfying thing to put on a slide and a far better thing to own when the question comes. The distinction matters most in its failure mode: evidence assembled retroactively is written by people who already know the answer, and retroactive narration is always tidier than what actually happened, smoothing over the Tuesday when two systems disagreed and someone made a call in a corridor. A contemporaneous record is messier and more valuable, precisely because it preserves the alternative that was considered and rejected, and the reason it was rejected.

There is a second, quieter dividend. When a definition is reinterpreted, a source system is migrated, or an error is found in an upstream feed, the real question is never only what happens next period — it is what this does to the periods already submitted. Where lineage is the deliverable, that is a query: show me every figure in the last eight periods whose value depended on this rule, this feed, or this classification. Where the report was the deliverable, it is a project staffed by people reading old spreadsheets, and it takes as long as it takes.

The part machines are good at is the part humans abandon under deadline

This is where automation has genuine business, and also where it is most often mis-sold. Under deadline pressure the first casualty is always the record. Nobody stops at eleven at night to write down why they took the second reading of an ambiguous definition; they take it, they move on, and they intend to document it later. For a person, doing the work and recording the work are two separate acts, and only one has a deadline attached. For a system they can be the same act. A reasoning layer sitting alongside the pipeline — with specialist agents on extraction, reconciliation, classification and exception handling, writing observations as it goes — can hold the provenance of every extract, the version of every rule applied, the population it touched, and the full history of every exception, not because software is more diligent than people but because for software there is no second act to skip.

What such a system must never do follows directly from the nature of an attestation: it does not sign, certify, or submit anything. It cannot, in any meaningful sense, because accountability requires someone who can be asked to explain themselves and who bears consequences, and no software has that property. Its proper role is to make a good attestation possible — to place in front of the accountable human a complete and honest account of what was done, including, especially, the places where the answer was genuinely uncertain, the interpretations that were live, and the exceptions disposed of by judgment rather than rule. Human-in-the-loop here is not a safety veneer bolted onto an otherwise autonomous process; it is the architecture, and everything the machine does is in service of the moment when a person decides whether they are willing to put their name to it.

Much of what is currently sold into this space fails on exactly that distinction, which is one reason Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, alongside a broader warning about "agent washing." In reporting, agent washing has a very specific signature: a tool that produces the number faster while leaving the account of the number exactly as thin as it was before. It compresses the close and does nothing for the reconstruction, which means it optimises the by-product and ignores the deliverable. The organisations that get this right are running what the reporting on the shift toward autonomous enterprise operations describes as missions rather than tasks — and an AI Mission for regulatory reporting, in the sense StudioX uses the term, is not a mission to produce a return. It is a mission to produce and maintain the institution's account of how that return came to say what it says, with the return itself falling out at the end.

The mental model worth carrying away is that the filing is a rendering, not an asset. What the institution owns is a position — a set of interpretations, applied to a set of data, at a moment in time — and the report is one view of it, formatted for a particular recipient. An institution that can reconstruct its reasoning can survive being wrong, because it can show exactly where the reasoning turned and correct it deliberately. An institution that can only reproduce its numbers cannot even establish that it was right; it can only insist. Which suggests a different measure for the reporting function: not whether it filed on time, but how quickly it can answer, for any figure in any period, the only question that has ever really mattered — why is this number this number — without having to call anybody.