AI in accounting — what it does, what it doesn't, and where to start

Artificial intelligence in accounting means software that reads unstructured financial information, proposes the accounting treatment, and learns from the corrections you make — instead of waiting for a person to type it in or for a developer to write a rule for every case.

That is the whole idea. Everything else is detail about which tasks it reaches, how far you let it go without a human, and what your team does with the hours it gives back.

This guide is the map. Each section links to a deeper piece where there is one.

The three things AI changes

Most of the confusion around AI and accounting comes from treating it as one capability. It is three, and they arrive at different speeds.

Reading. Extracting figures from an invoice, a bank statement, a supplier statement, a marketplace settlement report. This is the most mature of the three. It works on formats nobody standardised, which is precisely where rule-based automation always broke.

Deciding. Proposing which account a transaction belongs to, which invoice a payment settles, whether two records are the same thing. This is where the real time goes in most finance functions, and it is where machine learning beats a rulebook — not because it is cleverer, but because it generalises to cases nobody anticipated.

Explaining. Answering "why is this month's margin down" in a sentence, from your actual data. Newest of the three, and the one that changes what a finance meeting feels like.

A system that does the first but not the second still leaves you doing the thinking. A system that does the second without showing its work will not survive an audit. You want all three, and you want to be able to see into each.

What it actually takes off your desk

In a typical small or mid-sized finance function, the recurring work looks like this:

• Typing supplier invoices into the ledger
• Matching payments to invoices and chasing the ones that don't match
• Reconciling the bank, then reconciling the marketplaces and gateways
• Coding transactions to the right account
• Assembling the month-end pack
• Answering "can you check this figure" from three different people

The first four are substantially automatable today. The fifth becomes assembly rather than construction. The sixth changes character entirely — the question gets asked of the system instead of the person.

What remains is judgement: the treatment of an unusual transaction, the estimate that needs a view, the conversation with the auditor, the decision about whether the numbers mean what they appear to mean. That is not a consolation prize. It is the part of the job that was always being squeezed.

We go through this in detail in what AI actually does in accounting.

What it does not do

Three honest limits, because vendors are quiet about them.

It does not know your business. It knows patterns in your data. The first time you buy something you have never bought before, it will guess, and the guess will be based on the words on the invoice, not on your intent. You correct it once and it holds. But there is no version of this where you never look.

It does not remove accountability. If an automated entry is wrong, the answer to "who posted this" cannot be "the software". Someone owns the process, someone reviewed the exceptions, and the trail has to show both. This is worth settling before you automate, not after — see who is accountable for an automated entry.

It does not fix bad data. If your chart of accounts has grown to four hundred lines with six ways to book the same cost, automation will apply that mess consistently and at speed. Data quality is the real constraint, and it is almost always the thing that decides whether a project works.

More on the boundary in where AI in accounting still falls short.

The accuracy question, answered properly

"How accurate is it" is the first question everyone asks, and the honest answer has two halves.

On extraction and matching, a well-implemented system handles the large majority of routine transactions without a person touching them. We design toward 98% of reconciliation lines matched automatically — deliberately not 100%, because the last stretch is made of genuine ambiguity that a person should look at. A vendor promising 100% is either not counting the exceptions or hiding them.

But the number that matters more is different: of the transactions it got wrong, how many did you find out about? A system that is 99% accurate and silent about the 1% is worse than one that is 96% accurate and hands you a clean exception queue. That is why confidence scores matter and why the exception queue is a control, not a failure list.

What changes for the people

The fear is redundancy. The reality, in most finance functions, is a change of shape.

Data entry stops being how a junior learns the business — which is a genuine problem, because that is how a generation of accountants absorbed what a normal transaction looks like. Practices that handle this well replace it deliberately: juniors review exceptions instead of typing entries, which exposes them to the odd cases faster than typing the ordinary ones ever did.

Senior people spend less time assembling and more time being asked questions they can now answer in the meeting. Controllers move from producing the pack to interrogating it.

We cover this properly in AI accountants — what the job becomes and will AI replace accountants.

Where to start

The failure pattern is starting everywhere. The pattern that works is starting with one process that is high-volume, low-judgement, and currently painful.

For most businesses that is either supplier invoice capture or reconciliation of one sales channel. Both are repetitive, both have an obvious right answer you can check against, and both produce a visible result within weeks.

Do not start with the month-end close. It touches everything, judgement is dense, and if it goes wrong it goes wrong in front of the board.

The sequence that tends to hold:

1. Pick one process and write down how it works today, including the exceptions
2. Fix the data that process depends on — the chart of accounts, the supplier master
3. Run it in parallel with the old way for one full cycle
4. Compare, correct, and only then switch off the old way
5. Add the second process once the first is genuinely being used, not merely installed

Starting with AI in accounting walks through this, and choosing the first accounting process to automate helps with the pick.

What to ask a vendor

Short list, and the answers are revealing:

• Show me an exception. Not a success — an exception, and what the reviewer sees.
• What happens to a transaction the system has never seen before?
• Can I see why it coded this to that account?
• What does your audit trail record, and can my auditor read it without your help?
• What happens to my data if I leave?

How to evaluate AI accounting software expands this, and red flags in an AI accounting demo covers what a rehearsed demo hides.

What it looks like in your kind of business

The general shape is the same everywhere; the hard part is not. A retailer with six outlets, an online seller across three marketplaces and a trading company waiting six weeks for landed cost each have a different first problem.

• A multi-outlet retailer — six streams of takings that have to be attributed to an outlet and a day
• An online seller — four defensible figures for the same sale, and four reconciliations that have to complete
• A trading company — margin that cannot be calculated until landed cost finishes arriving
• A wholesale distributor — receivables and rebates, both usually accounted for badly
• A group of companies — a fifth close that nobody has systematised
• A franchise network — numbers produced by businesses you do not control
• Selling in two countries — currency movement mistaken for performance
• A seasonal business — the worst processing quality in the most important month
• High volume, low value — where every control becomes uneconomic
• A founder with no finance team — no second pair of eyes

Common questions

What is AI accounting?

AI accounting is the use of machine learning and language models to read financial documents, propose accounting treatment, and reconcile transactions with limited human input, learning from corrections instead of requiring a rule to be written for every case. It differs from older accounting automation in that it generalises to formats and situations nobody configured in advance, which is why it copes with things like supplier invoices and marketplace settlement reports that arrive in hundreds of different shapes.

Will AI replace accountants?

It is replacing specific tasks rather than the role, and the tasks going first are the ones with the least judgement in them — data entry, matching, first-pass coding. What grows in their place is review, exception handling, interpretation and advice. The realistic risk to an individual accountant is not being replaced by software but being outcompeted by an accountant who uses it, in the same way spreadsheets did not remove accountants but did remove those who refused to learn them.

How accurate is AI in accounting?

On routine extraction and matching, a well-implemented system handles the large majority of transactions without human intervention, and a sensible design target is around 98% of reconciliation lines matched automatically, leaving genuine ambiguity for a person. Accuracy on its own is a misleading measure, though — what matters as much is whether the system tells you which transactions it was unsure about, because undetected errors cost far more than flagged ones.

Is AI accounting safe for compliance and audit?

It can be, provided every automated action is attributable and reversible, meaning the record shows what was posted, on what basis, which rule or model version produced it, who reviewed it, and what the original document said. Auditors are generally comfortable with automated processing where the trail is complete and the controls around exceptions are documented; they object to systems that produce entries nobody can explain after the fact.

Where should a business start with AI in accounting?

Start with one high-volume, low-judgement process that is currently painful — usually supplier invoice capture or reconciling a single sales channel — and run it alongside the existing method for one full cycle before switching over. Starting with the month-end close is the common mistake, because it touches every other process and failure is highly visible.

Related: what AI actually does in accounting · where AI in accounting still falls short · how to evaluate AI accounting software