Automatic categorisation of bank transactions: how reliable

Quick answer. Automatic categorisation assigns an accounting account to each bank transaction, using rules and, increasingly, machine learning. It is reliable on recurring operations, but gets it wrong on new, ambiguous and mixed cases, and above all on VAT and the expense-or-fixed-asset distinction. It is an aid, not a substitute: responsibility for the accuracy of the entries remains the accountant's.

Automatic categorisation of bank flows has become the backbone of connected accounting: the bank feeds the operations in, the tool proposes an account, and data entry seems to vanish. The promise is real, but its reliability depends on the type of operation. Poorly framed, it produces plausible but wrong entries that only surface at review, at the VAT return, or at audit. This article details what it does well, where it slips, and how to organise a targeted control that secures your accounts without sacrificing the time saving.

The stakes are not theoretical. The entries produced by this categorisation feed the accounting entries file (FEC), which the tax authority can request in case of an audit under article L47 A of the French Tax Procedures Code. Categorisation followed blindly is therefore not a mere management convenience: it commits the reliability of accounts that can be held against you.

Automatic categorisation assigns an accounting account to each bank operation, combining two mechanisms.

The first is the rules engine: a recurring label, an amount, a flow direction or a counterparty account triggers allocation to a given account. You write the rule once (for example, any transfer mentioning a landlord goes to the rent expense account), and the tool applies it from then on without intervention. The second mechanism is machine learning: the tool observes your past classifications and corrections, then proposes the most likely category for an operation that resembles what it has already seen. The richer and more consistent the history, the more accurate the proposal.

The two mechanisms have distinct strengths and blind spots. The rules engine is deterministic: it does exactly what you told it to, no more and no less, so it stays auditable, but it does not adapt to cases no rule covers. Learning is probabilistic: it generalises from examples, which makes it flexible on label variations, but opaque in its reasoning and sensitive to the quality of the history. A categorisation inherited from a poorly kept file passes on its biases: the model learns from erroneous classifications and reproduces them with confidence. In practice the two complement each other: the rule secures the known flow, learning proposes a hypothesis on the rest, and the accountant's role is to arbitrate where the proposal is neither certain nor verifiable from the statement alone.

On repetitive, recognisable operations (rents, software subscriptions, net salaries, social charge debits, bank fees), reliability is high and the time saving real. It is on this recurring flow that automation brings the most value, freeing up entry time to reinvest in control and analysis. This logic is part of the broader move to migrate to a cloud accounting tool, where the quality of the initial rules determines the lasting reliability of the file, and it feeds a financial management that relies on clean accounting data to steer cash flow.

Reliability drops as soon as the operation leaves the learned frame.

Four families of cases cause problems. First, new operations the tool has never seen: a first purchase from a new supplier, an exceptional operation, a new type of income. Learning has no precedent and offers, at best, an approximation. Next, ambiguous labels: the same bank wording can cover a deductible expense, a refund, or a movement between accounts. Then mixed operations: a transfer bundling several natures, a supplier payment split between purchases and shipping, an inflow net of a platform commission. Finally, and this is the most sensitive point, cases that require an accounting or tax analysis.

Two of these deserve constant vigilance. VAT first: the bank transaction carries no information on rate or deductibility. The tool sees a gross amount, not the split between the net base and VAT, nor whether that VAT is recoverable. The expense-or-fixed-asset distinction next: a durable equipment purchase must be recorded as an asset and depreciated, not expensed in the period. The bank statement says nothing about the useful life or the threshold applied. These are accounting qualification decisions, framed by the French general accounting plan, that examining the flow alone cannot settle.

Blind spots that depend on the activity#

Beyond these principles, the nature of the errors varies with the business model, which makes the idea of a one-size-fits-all setup dangerous. A few scenarios keep recurring in our files.

A company that buys or sells outside France stacks up VAT pitfalls. An intra-Community acquisition is settled net of tax and requires a reverse charge of VAT, which no bank flow signals; an import from outside the European Union shows VAT and customs duty amounts that are not expenses and that call for specific treatment. The tool, which only sees a transfer to a foreign supplier, cannot reconstruct this mechanism on its own. For an e-commerce or marketplace activity, the inflow arrives net of commissions and sometimes of VAT collected by a third party: recording only the credited amount distorts both turnover and the VAT due.

The expense-or-fixed-asset distinction also plays out differently by sector. An agency equipping a workstation, a practice buying medical equipment, a restaurant replacing its oven or a construction company acquiring durable tooling all face the same question, but with different thresholds, depreciation periods and uses. The tolerance threshold often applied to expense small equipment directly never removes the need to assess the actual useful life: a split purchase, a grouped order or an inseparable set can tip into a fixed asset even though each line, taken alone, looked modest. Here again the bank statement carries no such information; only the supporting document and a clear internal rule allow the call to be made.

A last frequent blind spot: operations with no management counterpart, internal transfers between the company's accounts, contributions or repayments of a partner's current account, expense reimbursements. Wrongly recorded as expenses or income, they artificially inflate activity and blur the reading of the result. These are discreet errors, because the amount is often round and the label neutral, but their cumulative effect can be significant over a year.

The table below summarises the expected reliability and the control reflex, operation by operation. It serves as a sorting grid: you do not control everything, you control where automation is weak.

Reading this table gives a simple rule: reliability is inversely proportional to the novelty and tax complexity of the operation. The recurring can be delegated, the qualifying must be controlled.

Automation speeds up entry, but responsibility for the entries remains human.

Automatic categorisation does not replace accounting judgment. Responsibility for the accuracy of the entries, which form the basis of VAT returns, the tax package and the annual accounts, remains the accountant's and, ultimately, the owner's who approves them. A control targeted on risk cases, rather than on the whole flow, is the right answer: checking dubious proposals, verifying VAT against invoices, reclassifying asset purchases, handling new operations. This is precisely the value of support on corporate taxation: turning a flow of plausible entries into genuinely reliable accounts, and arbitrating the qualifications the tool cannot settle. On more complex structures, such as a group with a holding company, automatic categorisation multiplies blind spots (intra-group flows, current accounts, recharges) that call for a dedicated review.

Well used, automation does not remove control: it shifts it to value-added cases. The time saved on recurring entry funds the verification time where the risk is concentrated. It is also one of the points on which the role of the chartered accountant is changing: less manual entry, more control, qualification and advice on the areas the tool cannot settle alone.

Data governance: the other side of the topic#

Automation also shifts part of the topic onto the data. When bank flows pass through learning engines, often hosted in the cloud, two questions are added to that of accounting accuracy: confidentiality and traceability.

On confidentiality, this data falls under professional secrecy and the GDPR. It is legitimate to know where it is hosted and who can access it, which ties into the debate on sovereign AI versus the US cloud for financial data. The firm favours a written usage framework, restricted access and, where possible, the pseudonymisation of data and the preservation of professional secrecy before any automated processing, all framed by a company AI charter. On traceability, a categorisation produced by a model remains an entry like any other: it must be supported by a document and reconstructable on request. Keeping a history of the rules applied and the manual corrections is not a luxury, it is what makes the file defensible in case of an audit.

Our view: delegate the recurring, keep control of the qualifying#

In the files we take over, the most common mistake is not using automation, it is trusting it indiscriminately. The tool excels on the recurring and gets it wrong on the exceptional, the ambiguous and the tax-related: these are two different reliability regimes, calling for two different treatments.

Our approach rests on three principles. First, invest in the quality of the categorisation rules at the outset and at each new case: a well-set rule is worth more than ten after-the-fact corrections. Second, focus human control on the three red zones (VAT, expense or fixed asset, new operations) rather than rechecking everything. Third, treat VAT as a permanent control point, not an automatic deduction: this is where the easiest reassessments to avoid sit. The tool learns from corrections and gains reliability over time, but that progress never dispenses with accounting judgment. You save time without losing reliability, provided you target.

A services SME entrusts us with its accounting after two years kept in-house on a connected tool, with systematic validation of the automatic proposals. On takeover, several anomalies emerge. VAT recovered on expenses that did not entitle it (partial catering, fuel), for lack of reconciliation with invoices. Several IT equipment and furniture purchases expensed directly in the period, when they were fixed assets to be depreciated, distorting both the result and the balance sheet. Internal transfers between the company's own accounts recorded as expenses, artificially inflating costs.

None of these errors was crude: each proposal from the tool was plausible. That is precisely what made them invisible to an owner validating on the fly. The correction required revisiting the documents, restating the entries and adjusting the VAT already declared. Since then, the file runs with a refined set of rules and a monthly control targeted on VAT and asset purchases: the time saving of automation is kept on the ordinary flow, and the risk is handled where it lies.

In practice: organising a targeted and efficient control#

Here are the operational reflexes to make automatic categorisation reliable without spending unreasonable time on it.

• Build and document your categorisation rules at the outset, then at each new recurring operation, rather than correcting afterwards.
• Systematically reconcile VAT against invoices: the bank carries neither the rate nor the deductibility.
• Set a qualification rule for expense or fixed asset and apply it to every durable equipment purchase.
• Isolate and reclassify internal transfers and movements between accounts, which are not expenses.
• Handle purchases outside France separately: reverse charge of intra-Community VAT, customs documents on import.
• Control recurring operations by sample, and new, mixed and exceptional operations exhaustively.
• Check the matching of customer and supplier receipts before review, to leave no unexplained gaps.
• Keep a trace of the rules and corrections, to keep a file that can be justified at audit.

A few pitfalls keep coming up when relying on automatic categorisation.

• VAT is never deductible just because the tool proposed it: deductibility depends on the nature of the expense and the invoice, not the bank flow.
• A durable equipment expense passed as a charge distorts both the period's result and the balance sheet: the expense-or-fixed-asset call is a decision, not an automatism.
• Internal transfers, refunds and capital contributions recorded as expenses or income artificially inflate activity.
• A categorisation inherited from a bad rule propagates: a setup error repeats across dozens of entries until the rule itself is fixed.
• Flows that pass through learning engines fall under professional secrecy and the GDPR: confidentiality, hosting location and processing traceability deserve to be framed.
• The entries feed the FEC, which is requestable at audit under article L47 A of the Tax Procedures Code: traceability and consistency of allocations matter.
• Trusting the tool does not transfer responsibility: it remains with the accountant and the owner who approves the accounts.