An AI Mission for Banking: Credit Memo Drafting

A credit memo's job was never to describe a borrower well. It was to make a bank's reasoning inspectable months later by someone who was not in the room — and that is a very different craft from writing.

On a Wednesday morning in a credit committee, someone stops on page nine of a twenty-two page memo and asks where a working capital figure came from. The memo is well written; the industry section reads cleanly, the risk discussion is organized, the recommendation is clear. But the analyst who prepared it has to think before answering, because the number came out of a spread built from an interim trial balance that arrived by email in the spring, adjusted for a reclassification a relationship manager mentioned on a call, and cross-checked against a certificate that was itself a month stale. Answering takes several minutes and ends, honestly, in a partial reconstruction from memory. Nobody ever doubted the prose. What wobbled was the provenance underneath it.

That small moment is the entire subject. A credit memo is not really a document, in the way a brochure or a term sheet is a document. It is the artifact by which an institution proves that it thought — the record that some group of people looked at a specific set of facts, on a specific date, and reasoned from them to a conclusion they were willing to sign. Its readers are not only the committee that approves it: the annual reviewer reads it to see what has changed, a portfolio manager reads it to understand a concentration, and if the relationship deteriorates a workout officer reads it to find out what the bank believed and when it believed it. None of them are reading for style. All of them are reading for evidence, and for the ability to follow any given assertion back to the thing that supports it.

Fluency was always the cheap part of a memo

Capable language models have been experienced in banking, initially, as a writing breakthrough, and it is worth being blunt about how little that is worth. Producing several thousand words of plausible narrative about a borrower — the company's history, its market position, the shape of its cash flows, a competent summary of what could go wrong — is now nearly free and nearly instantaneous. It is also the part of memo production that was never genuinely scarce; banks have had competent narrative prose about borrowers for as long as they have had junior analysts and a template. What remains scarce is the chain running from each individual claim back to something a third person can independently verify.

Anyone who has actually assembled a memo knows where the hours go, and it is not into sentences. They go into reconciling sources that disagree with each other: audited statements, management-prepared interims, a covenant compliance certificate, a field examination, an aging schedule, notes from a site visit, and the bank's own exposure records across affiliates and guarantors, each carrying its own as-of date, its own preparer, and its own quiet adjustments. The same underlying concept will appear in four places with four different values, and the craft is not writing the sentence that contains the number. The craft is deciding which of the four belongs in the sentence, understanding why the others differ, and leaving behind a trail that lets the next reader see the choice rather than merely inherit it.

This is exactly where generative fluency does its most subtle damage, because a model asked to describe a borrower will smooth over precisely the seams a careful analyst would expose. It will write that margins improved without noting that the improvement compares management-prepared interim figures against an audited prior year, or summarize a concentration without recording that the underlying schedule came from the borrower and was never independently confirmed. This is not hallucination in the dramatic sense that gets discussed at conferences; nothing is fabricated. It is something quieter and more corrosive — the unearned assertion, a statement that is probably true and is presented in a register indistinguishable from the claims in the same memo that were actually verified. A memo where earned and unearned claims look identical has lost the property that made it useful.

Assembly is machine work; judgment has to keep a name on it

If you take a memo apart, you find two substances bound together by habit rather than necessity. The first is evidence: the spreads and trend tables, the covenant calculations, the comparison of terms against internal policy, the aggregation of related exposures, the reconciliation of the same fact across four documents, the identification of what is missing or stale. This work is enormous, largely bounded, and almost entirely invisible in the finished product. It is also where the overwhelming majority of an analyst's week actually disappears.

The second substance is judgment: what the evidence means, which of the identified risks the institution considers material and which it is content to accept, and why. That is authorship in the real sense, and it should stay authorship, because it is the part the institution is accountable for and the part that carries a person's name. No plausible improvement in machine reasoning changes the fact that a credit decision is a commitment, and a commitment needs an author who can be asked to explain it.

Most of what gets built for this problem gets the division exactly backwards. It automates the writing, because writing is the visible output and therefore the thing that looks impressive in a demonstration, and it leaves the analyst doing the gathering and reconciling that consumed the week in the first place. The correct split runs the other way: the machine should own the assembly — retrieving documents, extracting figures, reconciling disagreements, attaching to every number a record of the file, the page, the date, and the preparer it came from — while the human owns interpretation on top of an evidence base that has already been built and cited. In the vocabulary StudioX uses for this kind of work, that means specialist agents drawing on enterprise knowledge and producing observations traceable by construction, with human-in-the-loop checkpoints placed where a conclusion is being drawn rather than sprinkled over every mechanical step. The intent is not to make the analyst a faster writer, but to hand them a fully sourced evidence base and let them spend their attention on the judgment they are paid for.

This distinction matters commercially, not just philosophically, because the market is full of systems that produce the artifact without producing the substrate. Gartner has predicted that more than forty percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear value, and what it calls "agent washing" — older tooling relabeled without the underlying capability changing. A system that drafts a credit memo and cannot show you, per assertion, what it drew from and when, is a textbook instance of that pattern. It has automated the visible thing and left the expensive thing exactly where it was, while creating a new liability nobody asked for.

An untraceable memo is worse than having no memo at all

That last claim deserves defending, because it sounds like an exaggeration and is not. Reviewers calibrate their skepticism to effort, and for a century fluent, confident credit prose was a reasonable proxy for diligence, because producing it was expensive — you could not write a coherent twenty-page narrative about a borrower without having read the file. That correlation was doing a great deal of quiet work in every credit process built on top of it. Break it by making fluency free, and the proxy does not merely stop working; it inverts, because the most polished memo in the stack may now be the one where the least verification occurred, and it will consume the committee's trust at exactly the rate its polish suggests it deserves.

The cost of that inversion is deferred, which is what makes it dangerous, and it arrives at the least convenient moment. It surfaces in an annual review, when someone tries to determine what has actually changed and discovers the baseline was never anchored to anything checkable, and it surfaces at a workout, when the institution most needs a defensible account of what it knew and when it knew it and finds instead a well-written narrative with no floor beneath it. An absent memo is at least an honest absence; the gap is visible and everyone treats the file accordingly. A fluent, untraceable memo makes the same gap invisible, which is strictly worse, because invisible gaps do not get compensated for.

This has a direct consequence for how such a system should be built, and it is a sequencing point rather than a feature point. Traceability cannot be retrofitted onto generated prose, because a system that writes first and cites afterwards produces citations for the language rather than language derived from the evidence, and those two come apart in precisely the cases that matter. The evidence layer has to come first, with its lineage intact, and the prose constrained so it cannot assert more than the evidence supports. The most important capability such a system can have is the ability to decline — to mark a section as unsupported, to flag a figure as stale, to leave a hole where a source should have been rather than smooth it over. A drafting system that never produces an awkward gap is not being careful; it is being fluent, which is the failure mode, not the goal. This is the distinction that the emerging body of work on the autonomous enterprise keeps returning to across every function it examines — that the value of autonomy lies in what it can demonstrate, not in what it can produce.

The reframing worth carrying away is that the interesting question was never whether a machine can write a credit memo. It can, easily, and that capability is close to worthless on its own. The question is what fraction of the memo's assertions carry a live pointer back to a document, a date, and a preparer, and how quickly that pointer can be followed when a committee member asks about page nine. Measured that way, a memo stops being a piece of writing and becomes something closer to a chain of custody, with the prose serving as packaging rather than product. The institutions that make that shift will find the machine superbly suited to the enormous, invisible assembly nobody ever saw, and constitutionally unsuited to the judgment that always had to have a name attached to it — which is, plainly, the division of labour the memo was invented to record in the first place.