AI Accounting: What Is Real and What Is Marketing

AI accounting is the use of AI to do finance work: reading invoices, matching them to purchase orders, coding transactions, reconciling accounts, flagging anomalies, and drafting the reporting that closes a month. Some of that is genuinely solved and in production at scale. Some of it is a demo with a confident voice. The difference is not subtle once you know where to look, and it costs real money to get wrong.

Adoption has moved fast enough that the question is now practical rather than speculative. Wolters Kluwer’s Future Ready Accountant research put AI adoption among tax and accounting firms at 41% in 2025, up from 9% a year earlier, with 72% using it at least weekly and 35% daily. This is no longer a pilot conversation.

The rule that sorts real from oversold

There is a single test that predicts, with unusual reliability, whether a finance AI use case works in production.

Does the task have a right answer that can be checked mechanically, and does a mistake become visible quickly?

An invoice has the correct total printed on it. A purchase order either matches or it does not. A duplicate payment is detectably a duplicate. In all of those, the model’s output can be verified against something, cheaply, by another system. When it is wrong, you find out.

A cash-flow forecast has no printed right answer. Nobody can check it today, and when it turns out wrong in March you cannot tell whether the model was bad or the world moved. That does not make forecasting worthless, but it does mean the error stays invisible for a quarter, which is exactly the property you do not want in the finance function.

Gartner’s 2025 survey of finance organisations running AI found the three use cases that passed a third of respondents were knowledge management at 49%, accounts payable automation at 37%, and error and anomaly detection at 34%. That ordering is not an accident. It tracks how quickly a mistake becomes visible: an anomaly alert only ever raises a hand for a human, and an invoice carries its own answer.

What is genuinely real

Document extraction and matching. Reading an invoice, a statement or a receipt into structured data, then matching it against a purchase order and a delivery note. This is the most mature application in finance and the one with the clearest economics: IOFM benchmarking puts the cost per invoice at roughly $6.30 with no automation against about $2.68 with end-to-end automation.

Anomaly and duplicate detection. Scanning every transaction rather than a sample, and flagging what does not fit. It raises a hand rather than making a decision, which is the safest possible shape for an AI output.

Close acceleration. A joint MIT and Stanford study found AI shifted 8.5% of accountants’ time away from data entry and shortened monthly close cycles by 7.5 days, with firms supporting 55% more clients per week. Seven and a half days off a close is not a productivity anecdote; it changes when the business can act on its own numbers.

Drafting the narrative. The commentary that explains the variance, written from the actual figures. The numbers stay computed; the prose gets drafted.

What is being oversold

Autonomous judgment. The pitch is a finance function that runs itself. What exists is a system that handles routine handling well and needs a person wherever a decision carries consequence. The gap between the two is widest on exactly the numeric, high-stakes work the marketing leans on hardest.

Forecast precision. Better inputs and faster iteration genuinely improve forecasting. The claimed accuracy figures usually come from vendors measuring themselves on their own customers, and a forecast is the hardest possible thing to audit after the fact.

“It eliminates errors.” It changes the error profile. Manual entry produces scattered, random mistakes that a reviewer catches by noticing something looks odd. Automated processing produces systematic ones: when it is wrong, it is wrong the same way across thousands of records, silently and consistently. That is harder to spot and larger when found, which is an argument for sampling automated output deliberately rather than trusting it because it is machine-made.

The compliance story. Automating a control does not discharge responsibility for it. Someone still has to be able to explain what ran, on what basis, and answer for it, which is the governance layer rather than a feature you buy.

Where the money actually is

For most companies below enterprise scale, the return is concentrated in accounts payable and the close, and it is boring.

It is boring because the win is not intelligence. It is the removal of handling: the keying, the chasing, the reconciling of one system against another, the three days at month-end spent assembling what happened. That work has no judgment in it, produces nothing a client values, and consumes a disproportionate share of a finance team’s month. Removing it is worth more than any forecasting feature, and it is the part that actually works.

The second-order effect is the one worth planning for. When the close takes days instead of weeks, finance moves from reporting on the past to informing decisions in the present, which is the change worth making across the back office rather than in finance alone.

How to buy it without getting sold

Three questions cut through most vendor conversations.

What does it do when it is not sure? If the answer is that it proceeds, walk away. Anything touching the ledger needs a confidence threshold and an exception queue with an owner, which is the same design that makes any automated workflow trustworthy.

Can I see the working? Every automated entry needs a traceable basis: which document, which rule, which match. An accounting system that cannot show its working is not auditable, and auditable is not optional here.

What happens to my data and my rules if I leave? Finance processes encode how your company actually operates. If that logic is only reproducible inside one vendor’s platform, the switching cost lands on the part of the business you can least afford to have held hostage.

The honest starting move is unglamorous: take last month’s close, list every step, and mark each one as either *has a checkable right answer* or *requires a judgment*. Automate down the first list. Leave the second alone, and staff it better with the hours the first list gives back.

If you are weighing AI for your finance team, tell Daniel which process you would start with, what it runs on today, and what the business needs from it. Book 30 minutes with Daniel.

Frequently Asked Questions

What can AI actually do in accounting today?

Four things reliably: extracting data from invoices, statements and receipts into structured records; matching transactions against purchase orders and flagging duplicates; scanning every transaction for anomalies rather than sampling; and drafting the narrative commentary from computed figures. A joint MIT and Stanford study found it shifted 8.5% of accountants’ time away from data entry and shortened monthly close cycles by 7.5 days.

How do you tell a real finance AI use case from an oversold one?

Ask whether the task has a right answer that can be checked mechanically, and whether a mistake becomes visible quickly. An invoice carries its correct total; a purchase order either matches or does not. A cash-flow forecast has neither property, so an error stays invisible for a quarter. Gartner’s 2025 finance survey shows adoption following exactly that ordering.

Does AI eliminate accounting errors?

No, it changes their shape. Manual entry produces scattered random errors that a reviewer catches by noticing something looks wrong. Automated processing produces systematic ones: when it is wrong it is wrong identically across thousands of records, quietly and consistently. That is harder to detect and bigger when found, so automated output needs deliberate sampling rather than trust.

Where is the return on AI in finance concentrated?

Accounts payable and the month-end close, for most companies below enterprise scale. IOFM benchmarking puts cost per invoice at around $6.30 without automation against about $2.68 with it end to end. The win is the removal of handling, the keying and chasing and reconciling, rather than any form of intelligence, which is why it is both unglamorous and reliable.

Can AI replace an accountant or a CFO?

No, and the pitch that it can is where the marketing is thickest. What exists handles routine processing well and needs a person wherever a decision carries consequence. Automating a control also does not discharge responsibility for it: someone must still explain what ran, on what basis, and answer for it.

What should we ask a finance AI vendor?

Three things. What does it do when it is not confident, since anything touching the ledger needs a threshold and an exception queue with an owner. Can it show its working, meaning which document and which rule produced each entry, because auditable is not optional in finance. And what happens to your data and your process rules if you leave.