Who is responsible in AI-assisted accounting?
Automation speeds up recording, lowers error rates and simplifies reconciliation. But the consequence of a wrong classification is borne by the professional; that is the one thing technology has not changed.
The debate about AI in accounting usually starts from the wrong question: will machines put accountants out of work? The right question is this: when a decision is made with automation and turns out to be wrong, who is responsible? The legislation is clear. Returns are filed under the professional's signature, and the accuracy of records and their conformity with source documents is the professional's obligation.
Where automation helps, and where it falls short
| Task | Contribution of automation | Human decision required |
|---|---|---|
| Document reading and data extraction | High | At verification level |
| Creating repetitive entries | High | Sample checking |
| Account and bank reconciliation | High | Investigating exceptions |
| Account classification | Medium | Yes — the substance of the transaction governs |
| Exemption, relief and withholding decisions | Low | Yes |
| Tax planning and structuring | Low | Yes |
The distinction becomes clear here: processing data and forming a judgement are not the same thing. Reading the amount on an invoice is data processing. Deciding which account it belongs to requires establishing the true nature of the transaction, and that is a judgement.
The legal frame of responsibility
- The professional is responsible for the conformity of the signed return with the books and documents. An error made by a tool does not remove that responsibility.
- The taxpayer is responsible for the accuracy and complete delivery of documents. Automation does not change this division.
- The software provider's liability is limited by contract and does not extend to tax liability.
- Tools processing personal data create data protection obligations; where client data is processed must be known.
Five rules for responsible use
- Record the basis of the decision. Where an automated classification is accepted, the criterion it rests on should be written down.
- Keep an exceptions queue. Transactions the tool is unsure about must be routed to a person for decision.
- Test for systematic deviation rather than sampling alone. Errors in automated entries are not random; a rule set up wrongly once is wrong every time.
- Use past decisions as evidence. How a similar transaction was treated for the same client is the most reliable measure of consistency.
- Keep the decision record even when the tool changes. Reasoning must not disappear with a software migration.
Legal basis
- Law No. 3568 (responsibility of professionals)
- Tax Procedure Law No. 213, repeated Article 227 (responsibility for signing returns)
- Personal Data Protection Law No. 6698
This article is general information and does not replace professional assessment of a specific matter. Amounts and rates relate to the stated year; please verify the current provisions before acting.
