In finance, AI has to show its work.

An answer is not an explanation

On the Tuesday of close week, a controller opens a dashboard and reads:

It's probably true. It's also unusable. Which customers? Compared with when? Driven by what? Worth how much? And what is anyone supposed to do before Thursday?

Answering those questions takes the rest of the morning and three other screens, at which point the alert has cost more time than it saved.

This is the ordinary experience of AI in finance right now. The output is fluent, plausible, and difficult to act on, because acting on it means being able to defend it to someone else.

The bar in finance isn't accuracy

For most software, a wrong answer is an inconvenience. In finance, an answer can feed a cash decision, a collections priority, a provision, a journal — something that ends up in management reporting and eventually in front of an auditor.

The person acting on it owns it.

So accuracy isn't the standard, or at least not the only one. A number that happens to be right but can't be traced is difficult to use, because the controller can't sign their name to it.

What finance needs is an output that can be followed back to the financial evidence that produced it.

That's a lower claim than most AI vendors make — and a more demanding one to build.

What an explanation has to contain

Explainability in finance has almost nothing to do with model internals. No controller has ever wanted to see attention weights. It means four things, in order.

• What changed. Past-due exposure is up 18% week on week: a number, a direction, and a comparison basis.
• What's driving it. Three customers account for 72% of the increase. Not a summary of the whole book — the specific part that moved.
• Why it matters. Two of those three represent a meaningful share of the cash expected this month, and one has a dispute open since the 14th. This is the step almost everything skips — and the one that turns a fact into a decision.
• What to do about it. Work those two accounts before the next collection cycle, and get the dispute in front of someone who can settle it.

Each of those should be traceable. The 18% should lead to the accounts behind it. The 72% should lead to the three customers. The dispute should lead to the deduction and the invoice it sits against.

An explanation is only as good as the context behind it

Here's the constraint nobody selling AI in finance mentions: a model can only explain what it can see.

An assistant that has been given the collections queue can explain collections. It will tell you what is overdue, by how much, and for how long — and it may be completely right.

What it cannot tell you is whether some of that overdue balance has already been paid and is sitting unapplied, whether a reconciliation difference is distorting the receivables position, whether concentration is quietly increasing, or whether any of this is going to show up as a close exception.

It won't flag those as gaps. It will answer confidently within its boundary, a finance data silo seen from the model's side, and the boundary is invisible to the reader.

That's why explainability isn't something you can bolt onto an AI model at the end. The context has to exist underneath it first — which is exactly what a connected finance operation provides.

What explainability does not give you

It's worth being straight about the limit, because most writing on this subject isn't.

An evidence trail does not make a conclusion correct.

A system can show you exactly which three customers drove an increase and still draw the wrong inference, because the dispute is about to be resolved, or because the same thing happened last September for a reason that isn't in any system.

What traceability gives you is the ability to check.

It moves the work from trusting an output to reviewing one — the same relationship a controller already has with a preparer: not blind acceptance, not doing it yourself again, but a reviewable piece of work with the support attached.

That's a more modest promise than “AI you can trust.” It's also the version that survives contact with an audit.

What this looks like in practice

This is the principle Finni is built on.

It doesn't return a conclusion and ask to be believed. It surfaces the change, the accounts behind it, the financial impact and a recommended action, with the supporting figures available to inspect.

It also doesn't post anything on its own. It prepares, recommends and shows the evidence, and a person decides.

In finance, that isn't a limitation to be engineered away later. It's the correct division of labour.

The test

Before acting on any AI output in finance, ask five questions:

• What changed?
• What caused it?
• What evidence supports that?
• What does it affect?
• What should I do next?

Most tools can answer the first. Some can answer the second. The ones worth using can answer all five — and let you check every one of them.