Bookkeeping is a natural target for AI. It is repetitive, structured and full of rules, and most small business owners would happily hand it over. Some of it can be handed over today. Some of it cannot, and the difference matters because errors in accounts do not announce themselves: they surface at the year end, during a VAT review or when the accountant asks why a number looks odd.
What AI does well
➠ Suggesting categories for bank transactions, based on the
description and your past choices.
➠ Matching payments to invoices, including part payments and
slightly different amounts.
➠ Reading data from invoices and receipts, such as the
supplier, date, totals and VAT.
➠ Flagging oddities: duplicate payments, unusual amounts, a
supplier invoice with no matching order.
➠ Drafting payment reminders and supplier queries.
➠ Explaining a report in plain language, which helps the
people who dislike reading spreadsheets.
➠ Preparing a clean list of questions for the accountant.
These are tasks where a mistake is visible, easy to correct and low in consequence, and where the tool saves real time.
Where it is unreliable
Tax treatment of unusual transactions is the main one. A model will state a VAT rate, a depreciation period or a deduction rule with complete confidence, and it may be applying a rule from another country or from a year ago. Anything that changed recently is especially exposed. Judgement calls are the other: whether an expense is personal or business, how to treat a deposit, what to do with a refund that spans two years. These depend on facts and on your accountant's view, not on pattern matching.
Consistency is a quieter problem. The same supplier may be categorised one way in March and another way in October unless you fix a rule, and inconsistent categories make the figures harder to read and to defend.
The accountant's role does not go away
In France the title expert-comptable is protected, and the accountant carries professional responsibility for the figures they prepare or certify. An AI tool cannot take that on. What it can do is reduce the cost of the mess you hand over: well-categorised transactions, matched invoices and a short list of exceptions mean fewer hours at the accountant's rate. Ask your accountant how they want data delivered, which categories they use and which items they always want to see.
Guardrails that keep it safe
Review before posting. For the first few months a person approves every categorisation. After that, review a sample each month, say one in ten.
Build a rules list from your accountant's guidance and apply it consistently, covering things like fuel, meals or software subscriptions.
Keep the source documents attached to every entry. An audit needs the invoice, not a model's reading of it.
Reconcile the bank every month and investigate differences straight away.
Never accept a suggested VAT rate or tax treatment without checking it.
Use read-only connections to bank and accounting tools where you can.
Lock each month once it has been reviewed, so figures do not change silently.
A weekly routine
Twenty minutes is enough. Import the week's transactions, let the tool suggest categories and matches, approve or correct them and look at the list of exceptions. Once a month, send the accountant a summary of what changed and any open questions. Done weekly, the work stays small.
Why e-invoicing helps
Since 1 September 2026 French businesses must be able to receive structured electronic invoices, and the obligation to issue them follows in phases, from large and mid-size companies now to small ones in September 2027. Structured data is easier to read than a PDF, which makes extraction more reliable and cuts re-keying. It does not remove the need to check, but it does remove one source of error.
Protecting the data
Bank data and supplier invoices are commercial information and, for sole traders, personal data. Check that any tool you connect has a data processing agreement, lists its sub-processors and lets you delete what it has stored. Where it relies on an external model provider, find out whether your content is used for training.
Where this leaves you
Mirage Cloud connects to Qonto and Pennylane and includes an accounting agent scoped to French accounting, VAT and tax questions. Its own acceptable use policy says output must be verified before use, which is the right way to treat it: a fast assistant that prepares the work and explains it, with a person and an accountant making the decisions.
Top comments (0)