Fin AI: A Real Finance Task, Done With AI

By 2026, Gartner expects 90% of finance functions to have deployed at least one AI-enabled tool — but its own research also found fewer than 10% of those functions expect it to cut headcount. That gap is the honest starting point: most of what counts as "fin ai" right now is not an autonomous finance department, it is a chat tool doing a narrow, checkable task faster than a person would, while a person still owns the result.

Gartner's own 2025 survey of 183 CFOs and senior finance leaders found usage had basically plateaued — 59% reported using AI in their finance function, barely up from 58% the year before, after jumping from 37% in 2023. The easy wins got taken fast. What's left is slower, because it runs into data quality, governance and trust problems that a chat window alone does not solve.

What the task actually looks like

In practice, "using AI in finance" usually means one of a short list: turning a transaction export into categorised spend, drafting the first pass of variance commentary against a budget, summarising a long vendor contract or statement, or converting a rough forecast into a written narrative a non-finance reader can follow. None of those is the tool making a decision. All of them are the tool doing a first draft that a person checks before it goes anywhere near a report.

The request needs the same discipline as any other job you hand to a chat tool: the actual numbers, not a description of them, a stated scope, and a defined shape for the answer. Writing a prompt that works on the first try covers the general version, and budget ai walks through the same discipline applied to a personal or small-business budget. A finance task adds one more requirement: say what you will not paste in. Client names, account numbers and anything a chat tool's own terms say it may retain should not be in the box, whatever the task.

A worked example

Say a small business has 18 card transactions for the month and wants them sorted into a chart-of-accounts-style summary before the bookkeeper reconciles it properly. Weak request: "categorise my expenses." Better: "Here are 18 transactions — date, vendor, amount. Sort each into one of these five categories: Software, Travel, Office supplies, Marketing, Meals. Show a running total per category. Flag any single transaction over £300 and any vendor that appears more than twice, so I can check those by hand. Do not guess a category if the vendor name doesn't make it obvious — say 'unclear' instead."

A reasonable answer sorts the clear ones cleanly — a monthly SaaS subscription is obviously Software, a train ticket is obviously Travel — and actually returns some rows marked unclear, because a vendor name like a generic-sounding limited company tells the model nothing about what was bought. That "unclear" bucket is the useful part of the answer, not a failure of it: it is the tool telling you where it is guessing, if you asked it to.

The same pattern holds for the other common task: turning a budget-versus-actual table into a written variance comment. Hand over the two columns and the category names, ask for one sentence per category that states the direction and size of the gap and nothing else, and the draft is genuinely useful — it is the blank-page problem solved, not the analysis done. What it cannot do is tell you why marketing spend ran 40% over plan unless you also tell it; a model reading two numbers has no access to the fact that a campaign got moved up a quarter; it only has the numbers you gave it, so a plausible-sounding reason it invents on its own is a guess dressed as an explanation.

Where it goes wrong

If you don't ask for an unclear bucket, you don't get a refusal — you get a confident guess that reads exactly like a correct categorisation. Producing fluent, plausible output regardless of whether the underlying claim is actually right is a documented property of how these models work, and a finance summary is a bad place to discover that property by accident, because a miscategorised expense doesn't look wrong until someone reconciles it against the bank statement weeks later.

There's a narrower failure specific to regulated finance work. FINRA's Regulatory Notice 24-09 is explicit that existing rules don't pause for generative AI — a firm using a chat tool to draft client communications or supervise correspondence is still bound by the same rules on accuracy and recordkeeping it always was, and still has to show it checked the output, not just that it generated one. The CFA Institute's own framework on ethics and AI in investment management makes the same point from the practitioner's side: the professional judgement and the confidentiality obligation don't transfer to the tool just because the tool produced the draft.

None of that is a reason to avoid the tool. It's a reason to treat its output the way you'd treat a junior analyst's first draft — useful, fast, and never the version that goes out the door unchecked. What AI is actually bad at covers the same pattern outside finance specifically, and it is the same pattern every time: confident fluency is not the same signal as correctness.

Checks before you trust the output

1. 01Reconcile the category totals against the actual statement total yourself — a split that looks tidy can still be wrong by one misfiled line.
2. 02Treat every row the model marked unclear, or didn't mark at all when it should have, as a to-do, not a result — that list is often more useful than the sorted rows.
3. 03Never paste in anything a client would not expect to see leave the building — account numbers, case details, anything covered by a confidentiality obligation. If the task needs that data, it needs a tool with an actual data agreement, not a general chat window.
4. 04Keep a note of what you checked and when, the same habit NIST's AI Risk Management Framework recommends for any AI-assisted output that feeds a decision — in a regulated firm that note may be the thing a supervisor actually asks for.
5. 05If a number looks off, start a new conversation with the corrected input rather than arguing with the old one — pushing back in the same thread gives the model something to agree its way out of. Checking an AI answer when you are not the expert covers why that habit works.

The tool sorts and drafts fast. The reconciliation, the confidentiality call, and the final sign-off are still the job.

What to do Monday

Pick one recurring task you already do by hand — sorting a transaction export, drafting the first line of a variance comment, summarising a long statement — and run it through a chat tool with an explicit unclear bucket built into the request. Check the output against the source before you use a word of it. If that goes well, how to use AI as a payroll specialist and excel accounting software cover the same discipline applied to two specific finance roles, and are worth reading next.

Coursium teaches exactly this kind of practical judgement — what a task is actually safe to hand an AI tool, and what to check before you trust what comes back. Stay ahead of AI and learn the tools on your phone.