Why Citations Are Non-Negotiable in Financial AI

Every number a financial AI reports about a company must link to the document it came from. A plausible but unsourced figure is not a convenience, it is a liability, because in finance a wrong number that reads correctly is exactly the kind that ends up in a decision.

This is the single principle that separates a research tool you can rely on from a clever text generator you cannot. It is not about politeness or academic habit. It is about the fact that a figure you cannot trace is a figure you cannot trust, and an untrustworthy figure in a spreadsheet that drives capital is a hazard, not a help.

The problem is confidence without a source

Language models are fluent. Ask one for a company’s revenue in a given quarter and it will produce a sentence in the exact shape of a correct answer, with a number that sits in a believable range. The grammar is perfect. The figure may be a guess. That gap between how a thing sounds and whether it is true is the whole problem, and we have written before about why general chatbots invent financial numbers.

The danger here is not the obvious error. A number that is clearly absurd gets caught. The dangerous error is the plausible one: off by a small margin, drawn from the wrong period, or attributed to the parent when you asked about a subsidiary. It reads correctly, so it disarms your judgement, and it flows into a model or a memo unchecked.

A wrong number you can spot is a nuisance. A wrong number that looks right is a trap.

A citation is what closes that gap. If every figure carries a link to the specific filing, statement, and period it came from, then a plausible-but-wrong number stops being invisible. You can open the source, match the line, and confirm or reject it in seconds. Without that link, you are trusting a sentence because it sounds like an answer, which is the one thing a serious research process cannot afford.

Why finance punishes this harder than most fields

Ask an AI for a dinner recipe and a slightly-off answer costs you a pinch of salt. Financial figures are unforgiving in ways that turn a small model slip into a real mistake.

The number must be exact. Roughly right is wrong. A margin stated as one figure is a different fact from the same margin a fraction of a point away, and a model that lands near the answer has still landed on a false one.

The period must match. The same company has a different revenue for the quarter, the trailing twelve months, and the full year, and different again year on year. A figure can be genuinely correct for one period and completely wrong for the one you asked about.

The entity must match. Consolidated is not standalone. A segment is not the whole. A parent is not a subsidiary. The same word, say revenue, points to several different true numbers inside one group, and only the source tells you which one you are holding.

And the history itself can move. When a company reports a quarter, it often restates the prior-year comparable for an accounting change, a demerger, or a segment redefinition, so the version of the past you read today is not always the version that was on record when a decision was made. This is the point-in-time problem, and it matters because a number without a dated source is a number floating free of the exact context that makes it true. We go deeper on this in why point-in-time data matters.

In each of these cases, the citation is not extra colour. It is the thing that pins a figure to a single, checkable fact instead of a plausible average of many.

A citation is a receipt, not a footnote

There is a weak version of citing and a strong one, and only the strong one earns trust.

The weak version gestures at a company and a year and hopes that feels authoritative. It does not help. You still have to go and find the actual number yourself, and if you have to do that, the citation bought you nothing.

The strong version is a receipt. It names the primary document, points at the exact statement or line, states the reporting period, and lets you open the source and confirm the figure at a glance. The test is simple: can a careful person follow the link and land on the precise number, in seconds, without a treasure hunt? If yes, it is a real citation. If no, it is decoration.

A useful way to hold the distinction:

Weak reference	Strong citation
”According to the company’s filings”	Names the specific filing and date
Points at a whole document	Points at the exact statement or line
Period left vague	States the exact reporting period
Entity unstated	Specifies consolidated or standalone, parent or segment
You must go and find it	You can confirm it at a glance

This is why source-linking is really a discipline about primary documents. A number worth trusting comes from the filing itself, not from a paraphrase of a summary of a filing. The closer the citation sits to the original disclosure, the less room there is for a figure to drift on its way to you.

What source-linked answers actually protect

The habit of demanding a source protects three things that matter to anyone deploying capital.

It protects against the confident error, the plausible wrong number that would otherwise pass unnoticed. If nothing enters your work without a traceable source, the improvised figure has nowhere to hide.

It protects the audit trail. Financial decisions get questioned later, by a committee, a client, a future version of yourself asking why you believed what you believed. A chain of sourced figures lets you reconstruct the reasoning. A pile of unsourced assertions leaves you defending a number you can no longer stand behind.

It protects your time, which is the counterintuitive part. Citations look like friction, an extra step. In practice they remove far more work than they add, because the alternative to a good citation is re-verifying everything by hand, and a claim you can check in five seconds is cheaper than one you have to hunt down across a stack of filings. This is the same reason data quality beats model quality in this domain: the value is not in a fluent answer, it is in an answer you can stand on.

Altys’s stance: no number without its source

Our position is plain. In financial AI, an answer without a source is not a lighter version of a good answer, it is a different and worse thing. We would rather a system say “I cannot find that figure” than produce a confident number it cannot point to, because a blank prompts a check while a plausible fabrication prevents one.

That stance has a cost. It is harder to build a tool that binds every figure to a primary document than one that simply talks fluently about companies. It rules out the easy win of a smooth answer that happens to be unverifiable. But the whole point of research-grade AI is that a professional can trust the output enough to act on it, and trust in finance is not built on fluency. It is built on being able to check, quickly, that a number is what it claims to be. A model that reasons well over verified, sourced figures is doing real work. The same model improvising numbers it cannot cite is a hazard wearing the costume of an analyst, and understanding what it takes to genuinely read a financial statement makes clear why the source, not the sentence, is the thing that matters.

There is an old habit on any good research desk: never write down a number you cannot say where it came from. It predates AI by a long way, and it survives AI unchanged. The tools are new. The rule is not. A figure is only as good as the source you can point to, and a financial AI that forgets this is not saving you work, it is quietly handing you risk.