Can an AI categorize all of a company’s expenses on its own?

No. It proposes a category for recurring expenses with good confidence, but new purchases, unusual amounts and expenses with special treatment must be validated by a person. No entry should be booked without validation.

What accuracy can we expect from automatic categorization?

Published guides quote 85 to 95% correct categories, but those figures mostly come from vendors and are not independently verified. Plan for two to three months of corrections before reaching a stable level on your own data.

Do we need to change accounting software to categorize automatically?

Not necessarily. Categorization plugs into your existing bank, cards and accounting. The real prerequisites are a clean chart of accounts and twelve to twenty-four months of validated entries.

What happens when a receipt is missing?

It should be requested from the employee as soon as the line appears, not at month-end. An expense without a receipt stays pending and should not be booked definitively.

Categorizing SME expenses without losing your month-end

Categorizing company expenses: what can really be automated?

Short answer: an AI can propose an accounting category for most recurring expenses (subscriptions, fuel, rent, regular suppliers) by learning from your company’s own history. It makes more mistakes on rare expenses, unusual amounts and purchases whose nature depends on context only a person knows. The workable setup is: the AI proposes, a person validates the doubtful lines, and nothing is booked without that validation.

In many small companies, expense categorization is done in bulk at month-end by whoever has the least time for it. The bank statement sits next to a spreadsheet, receipts are spread across three mailboxes, and someone has to guess what hides behind a line like “CARD PAYMENT 14.10”. This article describes a setup that removes the chore without removing human control, with sourced figures where they exist and an opinion where they don’t.

What manual categorization costs, and why it piles up

The direct cost is data-entry time. The hidden cost is delay: an expense categorized three weeks later is one nobody remembers. The employee who made the purchase can no longer find the receipt, the line stays pending, and the month-end close gets squeezed.

Here is a worked example, with assumptions you should replace with your own. A 15-person company with 400 transactions a month (current account, cards, expense claims) that spends an average of 45 seconds on each to read it, categorize it and attach a receipt loses about 5 hours a month, or 60 hours a year. At an internal cost of 70 Swiss francs an hour, that is roughly 4,200 francs a year. These figures are an order of magnitude, not a measurement.

For context, an OpinionWay survey of French finance managers for Spendesk, quoted in a guide updated in 2026, ranks time saved as the top priority of accounting automation (71% of respondents), ahead of fewer manual tasks (52%). The survey comes from a software vendor and covers France: it points to a trend, it does not measure your case.

How automatic categorization works, in four steps

The principle is the same whatever the tool. This is the chain as we build it for the processes the studio automates:

Collection. Transactions arrive daily from the bank, company cards and expense claims. Receipts are matched by amount, date and supplier.
Proposal. For each line, the AI proposes an account, a VAT code and a cost centre, based on past entries and your chart of accounts. It reports its confidence level.
Sorting by confidence. High-confidence, low-amount lines go into a fast validation queue. The others are flagged to a person, with the reason for the doubt.
Validation and booking. A person approves, corrects or rejects. Only validated lines reach the books, and every correction feeds the next proposals.

The fourth step is what separates a usable automation from a gadget: without it, a wrong category is discovered at month-end, or worse, by your accountant.

What accuracy to expect, honestly

Published guides quote correct-categorization rates of 85 to 95%. One of them, published in October 2026, states itself that the figure comes from vendors and has not been independently verified. We treat it the same way: a plausible starting range, not a guarantee.

Three factors move that rate in practice:

A clean chart of accounts. If three accounts overlap, the AI hesitates as much as a new hire would. Simplifying the chart before launch is worth more than any tuning.
The amount of history. A supplier seen forty times is classified almost certainly. A supplier seen once is a question.
Bank label quality. A truncated or generic label gives no clue about the nature of the purchase. The receipt then becomes essential.

Our expectation, which remains a working hypothesis and not a measured result: a disappointing curve at first, improving as the system learns from your validations. That is why we suggest tracking the correction rate week after week rather than judging the tool on day one. A correction rate that falls every week is the best sign that the setup is working.

What is automated, what is validated, what stays human

Type of expense	Recommended handling	Why
Subscriptions and recurring suppliers	Automatic proposal, batch validation	Plenty of history, little ambiguity
Fuel, tolls, parking, routine meals	Automatic proposal, batch validation	Stable categories, low amounts
Purchase from a new supplier	Individual validation	No history: the category is a guess
Unusual amount from a known supplier	Individual validation with alert	Could be an error, a duplicate or a different kind of expense
Capital expenditure, durable equipment	Decision by a person	Expense versus asset is a question for your accountant
Entertainment, gifts, mixed-use spending	Decision by a person	Treatment depends on context the statement does not show
Entries tied to VAT and tax obligations	Validation by the accountant	The AI proposes a code, it does not carry the responsibility

This table is not accounting advice. Questions of tax treatment, VAT and record retention belong to your accountant or adviser: have them validate your rules before you write them into the system.

Setting it up without replacing your tools

The starting point is not new software. It is your current accounting system, your bank and your cards. Useful automatic categorization plugs into those, and does not force a migration. This is the order we follow:

Clean up the chart of accounts in use: remove duplicate accounts and fix a short list of categories.
Load the history: twelve to twenty-four months of validated entries serve as examples.
Run in parallel: for one month the AI proposes while a person categorizes as before. Compare, line by line.
Set thresholds: below which amount and confidence level a line goes to fast validation.
Decide who validates what: a named person, with a deputy, and a maximum validation delay.

If receipts are the real bottleneck, the neighbouring topic is supplier-invoice entry: we covered it in what an AI reads well and what it reads badly. The two chains combine, but they roll out better one after the other.

Common mistakes that sink the project

Trying to validate everything at once. A validation queue of 400 lines is abandoned within two weeks. High-confidence lines must be validated in batches.
Not linking correction to learning. If correcting a category changes nothing in the next proposal, the team stops correcting.
Letting the AI book without validation. Direct booking is tempting, but it moves the risk to the close.
Forgetting missing receipts. A category without a receipt is a fragile entry. The receipt should be requested from the employee as soon as the line appears.
Sending bank data to a consumer service. Statements contain personal and commercial data. Check where they are processed before choosing a tool.

When this approach is not the right choice

We advise against automating categorization in three situations. First, if you handle fewer than about sixty transactions a month: the time saved does not justify the setup, and a rule in your banking software is enough. Second, if your chart of accounts is unstable or being redesigned: stabilize it first, otherwise the AI learns rules that are about to change. Third, if nobody has time to validate: an automation without a validator is a risk, not a gain.

In the first two cases, the right investment is an hour of tidying. In the third, you first need to decide who is accountable for validation.

What a typical month looks like once it is running

This is the rhythm the setup above aims for, as an illustration and not a measured result. Every day, new transactions are proposed and high-confidence lines join the fast validation queue. Once a week, the designated person validates the batch in a few minutes and handles the flagged lines. At month-end, only the exceptions remain: missing receipts and expenses that only a person can qualify.

The important change is not the time saved at month-end, it is that the work spreads across the whole period. An expense is qualified a few days after it was incurred, while the employee still remembers its context. That is what reduces pending lines at the close.

The same logic applies to reviewers: a person who validates twenty lines every Friday keeps the context in mind, while a person who validates four hundred lines on the thirtieth does not. Short, regular validation sessions are more reliable than one long one, and they are easier to hand over when someone is away.

Where to start

The goal of a first pass is not to automate everything, it is to find out whether the volume justifies the effort. Take one month of statements. Count the lines, measure the real time spent and note the ten suppliers that come back most. If those ten suppliers account for more than half of the lines, automatic categorization is likely to save you time quickly. To see where you stand, the studio’s plans and prices show how a first automation is scoped and priced.