The most useful question about AI in finance is not where it could go. It is where it already works without anyone having to trust it on faith. Those are not the same place, and the gap between them is where most adoption decisions go wrong.
The marketing points one way. It promises forecasting that anticipates demand, models that read the market, a finance function that finally becomes strategic. The reliable wins point somewhere far less photogenic: reconciliation, the monthly close, pulling figures off invoices, assembling the binder for an audit. The work nobody puts on a conference slide.
Finance leaders who adopt from the back office forward, rather than from the strategy deck backward, tend to get more value and far fewer unpleasant surprises. The reason is structural, and it is worth being precise about.
Where AI is reliable, and where it is genuinely risky
The line that matters is not how clever the task looks. It is whether the output can be checked.
AI is dependable on bounded work with a verifiable answer. A bank line either matches a ledger entry or it does not. An invoice has a total that either reconciles to a purchase order or flags for review. A lease has a start date that is on the page or absent. These are tasks with a right answer that a person, or another system, can confirm in seconds. When the model is wrong, the error surfaces immediately and cheaply.
The risk lives at the other end, in work where a confident wrong answer is expensive and hard to spot. A revenue forecast that is plausibly off by eight per cent does not announce itself. A cash-flow projection built on a subtly misread assumption looks exactly like a good one until the quarter closes. Large language models are fluent by design, which means they produce wrong answers with the same composure as right ones. In bounded reconciliation that fluency is harmless because you check the result. In forecasting and strategy it is a liability, because the output is an argument, and a wrong argument that reads well is the most expensive kind.
This is why the unglamorous middle is the right place to start. It is not that the close is more important than strategy. It is that the close is checkable, repetitive and high-volume, which is precisely the shape of work where current models earn their keep.
Why anything that touches the ledger needs a human in the loop
There is a tempting next step once reconciliation works: let the system post the entries it is confident about. Resist it, or at least gate it hard.
The distinction is between drafting and committing. Having AI propose a match, suggest a coding, draft a journal entry or surface the twelve transactions that do not reconcile is enormously useful and carries little downside, because a person approves before anything moves. The moment the system writes to the ledger unsupervised, the economics invert. You have automated the creation of errors that now require detection, and detecting an error someone else made is far slower than reviewing one in front of you.
Keep the human on the commit step for anything that affects the books, controls or anything an auditor will later test. Let the machine do the gathering, the matching and the drafting, which is where most of the hours actually sit. This is not timidity. It is the same separation of duties that finance has enforced for a century, applied to a new kind of worker that happens to be a model.
The audit trail is not optional
Every AI step that touches financial data has to be reconstructable after the fact. If a tool suggested a match or a classification, you need a record of what it suggested, on what input, which version of the model, and who approved it. Without that, you have not automated a control. You have removed one.
This matters for two audiences. Auditors will ask how a figure was derived, and an answer that amounts to the software did not survive testing. Regulators, depending on jurisdiction and filing obligations, increasingly expect that automated decisions affecting financial statements can be explained and evidenced. A black box that produces numbers you cannot trace is a finding waiting to happen.
The practical implication is a procurement question, not just a technical one. Before a tool goes near close or reporting, ask what it logs, whether those logs are exportable, and whether they capture the input, the suggestion and the human sign-off as a linked record. If the honest answer is that it does not, that tool belongs in analysis or drafting, well away from anything that feeds the financial statements.
Smaller firms are catching up with off-the-shelf tools
A few years ago, this kind of capability meant a data science team and a custom build, which put it out of reach for most mid-sized firms and nearly all small ones. That has changed, and the change favours the smaller firm more than the large one.
The reconciliation, extraction and close features now arrive inside the platforms finance teams already run. The ledgers, the close-management products and the accounts-payable tools have folded machine reading and matching into the workflow, so a ten-person practice can adopt the same document extraction a listed company uses without hiring a single engineer. For teams weighing what to bring in, curated rundowns such as this guide to AI tools for accountants are more useful than another vendor demo, because they compare what the tools actually do against the work a finance team needs done.
The lesson from the firms doing this well is narrow scope. Pick one painful, bounded, high-volume process, namely bank reconciliation or invoice coding, prove the time saving and the control, then move to the next. The teams that struggle are the ones that bought a platform to optimise everything and ended up with a project instead of a result.
The skill that finance staff actually need now
The reflexive worry is that AI replaces junior finance roles, because so much of the early-career work is exactly the bounded, repetitive matching the tools are good at. Some of that work will compress. The conclusion that follows is not that the people become surplus.
The work shifts from producing the reconciliation to reviewing it. That is a harder skill, not an easier one. Reviewing AI output well demands knowing what a wrong answer looks like, where the model tends to fail, which exceptions deserve scrutiny and which are noise. It rewards judgement and scepticism over speed and stamina, and it asks finance staff to become editors and controllers of machine output rather than its manual substitute.
That is a more interesting job than the one it replaces, but it does not arrive for free. It needs training, deliberate review standards and managers who treat checking AI output as real work with its own discipline, rather than a rubber stamp on the way to month-end.
The honest summary is that AI is genuinely changing finance, just not at the front of the deck where the pitches concentrate. The first real gains are in the back office, in the checkable middle of the function, and the firms capturing them are the unflashy ones automating reconciliation and document extraction while keeping a person on every entry that reaches the ledger. The org chart will catch up to the technology eventually. There is no prize for forcing it to move before the controls, the audit trail and the people are ready.
