What AI actually does in accounting, task by task

"AI does your accounting" is a sentence that means nothing until you name the tasks. Accounting is perhaps forty distinct activities, and artificial intelligence reaches them very unevenly — near-total on some, barely at all on others.

Here is the task-by-task version, with an honest completion percentage for each.

Reading documents into data — mostly solved

You receive a supplier invoice as a PDF. Something has to turn it into a vendor, a date, a net amount, a tax amount, line items and a reference.

Older systems did this with templates: one per supplier, configured by hand, broken the moment the supplier changed their layout. Modern extraction reads the document the way a person does, from position and context rather than fixed coordinates, so it handles a supplier it has never seen.

Where it gets to: the fields come out right on the overwhelming majority of documents. Handwriting, photographs taken at an angle, and invoices where the tax treatment is stated in prose rather than a field are where it stumbles.

What is left for a person: confirming the ones the system flags, and the first invoice from each genuinely unusual supplier.

Deciding which account it belongs to — good, and improves

Coding is the task that quietly consumes finance teams. A payment to a hardware store might be maintenance, might be a fixed asset, might be stock, depending on what was bought and why.

A model that has seen your history proposes the account and tells you how sure it is. Correct it, and the correction sticks — not as a hard rule but as evidence, so the same supplier with a different description can still be treated differently.

Where it gets to: high on recurring spend, weaker on one-off purchases and anything where the correct answer depends on intent rather than description.

What is left for a person: the genuinely novel purchase, and periodic review that the coding still reflects how you want costs analysed.

Matching payments to invoices — high, with a catch

Simple case: one payment, one invoice, same amount. Solved for decades.

Real case: one payment covering nine invoices, less a credit note, less a deduction the customer applied without telling you, arriving four days late. That is where matching engines earn their keep, and where AI outperforms rules — it can propose the combination that adds up and show you the arithmetic.

Where it gets to: most of the volume, most of the time.

The catch: the residue is not noise. An unmatched payment usually means something real happened — a dispute, a deduction, an error. The value is not only in the matches but in a clean, short list of what genuinely needs a human. That is the exception queue as a control.

Reconciling the bank — high

Bank feeds arrive structured, which makes this easier than document work. The intelligence is in matching feed lines to ledger entries where references are missing or wrong, and in recognising the recurring items that should post themselves.

Where it gets to: routine reconciliation runs largely unattended between reviews.

What is left for a person: the genuinely unidentified receipt, and the judgement call on anything stale.

Covered fully in automating bank reconciliation with AI.

Reconciling sales channels — high, and this is where the money hides

If you sell through marketplaces or take card payments, the gap between what you sold and what landed in the bank is filled with commissions, transaction fees, shipping subsidies, promotional deductions and adjustments — each with its own timing.

This is arithmetic-heavy, high-volume, and almost impossible to do properly by hand at any scale. It is also where unrecovered money accumulates quietly.

Where it gets to: a well-built reconciliation accounts for every line of a settlement, whether or not it matched automatically.

What is left for a person: deciding what to do about a deduction you disagree with. See disputing a marketplace deduction.

Preparing the month-end pack — partial, and improving fast

The assembly is automatable: pulling balances, running comparatives, producing the schedules, flagging what moved unusually.

The judgement is not: whether an accrual is still appropriate, whether a provision needs revisiting, whether an unusual variance is a real change or a posting error.

Where it gets to: the pack builds itself; the review does not.

What is left for a person: everything that involves a view. More in closing the books with AI assistance.

Answering questions about the numbers — newly possible

Ask "why is gross margin down two points" and get an answer built from the transactions, with the ability to click into what it is claiming.

This is the genuinely new capability, and the one that changes meetings. It also carries the sharpest risk: a confident, fluent, wrong answer is far more dangerous than a blank screen. Every figure has to be traceable back to the records it came from, which is why explainability is not a nice-to-have.

Where it gets to: reliable on questions that are arithmetic over your data.

What is left for a person: anything requiring context that is not in the system — that a customer was lost, that a price rise was deliberate.

Judgement, estimates and treatment — barely at all

Whether revenue should be recognised now or over time. Whether a receivable is impaired. How to treat something the standards did not anticipate. Whether the numbers, taken together, mean the business is healthy.

Where it gets to: it can surface the inputs and flag what changed. It does not make the call.

What is left for a person: the call. This is the durable core of the profession, and it is a larger share of the job than it was.

The pattern

Look down the list and the rule is visible: the more a task is about turning one representation of a fact into another, the more completely it automates. The more it is about deciding what a fact means, the less.

That is a useful test when someone claims their product automates something. Ask which of the two it is doing.

Common questions

Which accounting tasks can AI do today?

Document extraction, transaction coding, payment-to-invoice matching, bank reconciliation, sales-channel and settlement reconciliation, and assembly of routine reporting are all substantially automatable now. Tasks involving accounting judgement — revenue recognition treatment, impairment assessments, estimates and provisions — are not, though AI can gather and present the inputs those decisions rest on.

Does AI post entries to the ledger by itself?

It can, and in most well-designed setups it does for high-confidence routine transactions, while anything below a confidence threshold is held for review instead of posted. The important design question is not whether software posts entries but whether every posted entry is attributable and reversible, so that the record shows what was posted, on what basis, and who was accountable for the process.

Can AI replace a bookkeeper?

It replaces most of the mechanical work a bookkeeper does — entry, matching, filing — but not the judgement about unusual transactions, the relationship with the business owner, or accountability for the result. In practice bookkeepers who adopt it take on more clients rather than fewer, because the constraint on a bookkeeping practice has always been processing capacity rather than demand.

What is the hardest accounting task for AI?

Anything where the correct treatment depends on intent or context that does not appear in the document. A purchase from a hardware supplier could be repairs, a capital item or stock, and the invoice frequently does not say which — the answer lives in why the purchase was made, which is information the system does not have until someone tells it.

Related: AI in accounting · where AI in accounting still falls short · how accurate is AI in accounting