AI in bookkeeping: what automation does well and what it does not
Invoice reading, posting and payment matching now run automatically. Here is where automation ends and human judgement has to begin.
Automation in bookkeeping has stopped being a promise. Reading an invoice from a file, assigning a cost category and matching a transfer to a document are routine today. It is worth knowing where that routine ends, because responsibility for the settlement still rests with the taxpayer, not with the software.
What automation genuinely does well
- Reading data from a document: number, dates, amounts, rates and issuer details. With a structured invoice the reading is flawless, because the data arrives in fields rather than in a picture.
- Repetitive posting: the same supplier, the same cost category, the same booking pattern every month.
- Matching payments to documents by amount, reference and counterparty, including batch transfers covering several invoices.
- Spotting anomalies: a duplicated document, an amount clearly out of line with history, a missing invoice for a cost that recurs every month.
- Checking the internal consistency of the records before the audit file is sent — totals, rates, completeness of numbering.
All these tasks share one feature: there is a single correct answer and it can be verified against historical data. That is the best possible ground for automation.
Where automation gets it wrong
Trouble begins where the outcome is decided by context outside the document. An invoice for a laptop looks identical whether the machine ends up on a desk at the office or in a child's room. What determines the treatment is how it is used, and that is simply not visible in the file.
- Classifying an expense as tax deductible where the purchase has a mixed character.
- Splitting a purchase between business and private use, including cars, phones and subscriptions.
- Judging whether a service bought from a foreign supplier has to be settled as an import of services.
- Unusual rates and exemptions that depend on the intended use of the goods rather than on their name.
- Everything that follows from the contract rather than the invoice: conditional rebates, acceptance stages, retention amounts.
What these situations have in common is that the document tells the truth and still is not enough to decide. That is why a well-built system does not try to guess: it flags such a document and sets it aside, instead of picking the most likely answer and quietly posting it.
The human role shifts, but does not disappear
In an accounting office that has adopted automation, the workload does not fall in proportion to the number of documents. It changes character: less data entry, more verification of exceptions and more conversations with the client about what the documents do not say. It is better work, but it requires a different shape of day and different skills than retyping invoices.
The conversation about price changes along with it. An office that charges per document earns less for the same service once automation arrives, so pricing the scope of care rather than the number of items becomes the natural direction. That conversation is better had before the rollout than after the first month.
Automation makes an excellent assistant and a terrible decision-maker. Responsibility for the settlement stays with the taxpayer.
How to adopt it sensibly
- Start with one area, preferably cost invoices — that is where repetition is highest and the risk of a single error lowest.
- For the first month, review every automatic suggestion and correct it; the system learns precisely from those corrections.
- Set a confidence threshold below which a document goes to a manual queue instead of posting itself.
- Leave a trail: who approved the posting and when, so a decision can be reconstructed during an audit.
- Do not automate an area you cannot check yourself — automation then only speeds up the repetition of an error.
Mind the dataAccounting documents contain personal data and commercial information. Before connecting an external tool, check where the data is processed, whether a processing agreement is in place, and whether the content of your documents is used to train the provider's models.
How this works in Księgowskaz
Cost invoices are pulled from the e-invoicing system and from a designated mailbox, and the software proposes a category based on the history of that particular supplier. The proposal is shown together with its justification — which earlier decision it rests on — so approving takes a second, and rejecting does not require hunting for where the suggestion came from. Documents with low confidence go to a separate queue instead of quietly entering the records.
The effect that matters to a small company is simple: closing the month stops being a full day of work and becomes a review of a dozen or so exceptions. The rest was posted along the way, on the days the documents actually arrived.
