What can AI take over in accounting?

What works well

Extract documents. Amount, date, invoice number, supplier.

Suggest postings. Based on previous postings of the same supplier. Very reliable for recurring documents.

Report irregularities. Duplicate invoice, unusual amount, different bank details. The last point is an effective protection against fraud attempts.

Search. „What did we spend on packaging materials in 2024?"

What should not be left to the model

Calculating. A language model estimates numbers, it does not calculate. Sums, tax amounts and balances belong in program code.

Posting without approval. A wrong posting often only becomes apparent during year-end closing.

Tax assessment. Whether an expense is deductible is decided by no AI.

The sensible setup

1. Model reads the document and suggests posting
2. Rule-based check: Does the math add up? Does the tax fit?
3. If consistent with previous postings: automatic pre-posting
4. Approval by a person
5. Posting

What this achieves

Time savings occur with the many similar documents. Exceptions still take time – but accounting can focus on them.

Data protection and storage

Documents contain personal data. For cloud processing, a data processing agreement is required. The original remains unchanged storage obligation – the AI processing changes nothing about it.