Where AI actually helps in finance — and where it doesn't

AI helps most in finance where the work is high-volume, rules-based, and checkable: reconciling transactions, coding invoices, drafting first-pass commentary, and flagging the odd-looking entry for a human to look at. It helps least where the work needs judgement, accountability, or a defensible answer someone has to sign. The useful question is not "can AI do finance" but "which parts, and who checks it."

After a few years auditing regulated firms and then running a finance function, I've found the line sits in a fairly predictable place. Here's how I'd draw it.

Where does AI genuinely help?

The wins are in the repetitive middle of the month. Matching a few thousand Stripe payouts to bank lines in Xero. Suggesting a nominal code for a supplier invoice based on the last fifty like it. Reading a PDF bill and pulling out the numbers. Writing the first draft of the variance commentary so the accountant edits rather than starts from blank.

What these have in common is volume, a pattern to learn from, and an easy way to check the output. If a coded invoice is wrong, you see it. If a reconciliation doesn't balance, it tells you. The cost of a mistake is low because the mistake is visible and reversible. That's the sweet spot.

In practice this is where the hours go and where they come back. A reconciliation that took most of a day can run in the background and surface only the exceptions. The team stops doing the matching and starts doing the checking, which is a better use of a qualified person.

Where does it fall short?

It struggles wherever the answer has to be defensible. Anything that ends in a signature, a filing, or a number a board will act on needs a human who understands why it's right, not just that a model produced it.

It's also weak on the genuinely novel. A new contract type, a one-off transaction, a judgement call on revenue recognition or a provision. Models are good at the average of what they've seen and poor at the case they haven't. Finance is full of edge cases that matter precisely because they're unusual.

And it has no skin in the game. A model like Claude, ChatGPT or Gemini will give you a confident answer to a question it should have refused. It doesn't know what it doesn't know, and it won't be the one explaining the mistake to a regulator or an investor. That accountability gap is the real reason the human stays in the loop.

So how should you actually use it?

Treat it as a fast junior who never tires and never quite gets the last 5% right. You give it the volume work, you keep the review, and you never let it sign anything.

The pattern that works is narrow and checked: point it at one well-defined task, give it only the data it needs, and put a human at the decision point. That's not a limitation to apologise for. It's the same control you'd put around any junior doing the same work, and it's why the right guardrails matter more than the model you pick.

It's also worth being honest that a lot of finance pain doesn't need AI at all. Often the fix is a connector between two systems or a simple rule, and reaching for a model is the slower, riskier option. I've written separately about why a simple automation usually wins.

If you're weighing up where AI fits in your own numbers, that's exactly the kind of thing a fractional CFO or Chief AI Officer is there to help you map. The goal isn't to use AI everywhere. It's to use it in the few places it earns its keep, and to keep a qualified eye on the rest.