Can AI hallucinate bookkeeping data?
AI hallucinates when it is asked to make something up. untxt. is not asked to invent bookkeeping facts. It reads the document in front of it, extracts visible values into fixed fields, leaves missing fields empty, checks the result against the original text, and sends uncertain work to review before anything reaches QuickBooks or Xero.
The fear is reasonable
Bookkeepers are right to distrust confident AI guesses. A wrong answer in a chat window is annoying. A wrong vendor, date, total, or account posted into the books is your cleanup problem. The issue is not that AI exists. The issue is whether the tool is allowed to invent a finished-looking answer where the document does not support one.
Extraction is different from invention
untxt. is not asked to explain tax law, choose a business strategy, or guess what a client meant. It is asked to read what is on the page: vendor name, document date, total, tax, payment terms, line items, and the surrounding document context. If the value is not visible, the right behavior is to leave it out or flag it, not fill the field to make the screen look complete.
The fields are constrained
The document data goes into a fixed accounting schema. Required fields like vendor name, document date, and total are tracked. If one is missing, it is exposed as missing instead of quietly invented. Extra fields do not become trusted data just because the model produced them; they are separated from the canonical fields and can be reviewed.
The original text gets a second read
After extraction, untxt. also reads the document as raw text. That second pass is used to check extracted values against the words and numbers actually present on the document. This is not magic, and it is not a guarantee that every bad scan becomes perfect. It is a guardrail against the specific failure bookkeepers hate: a confident value that does not appear on the page.
Account mapping is bounded by your chart of accounts
When untxt. proposes an account, it does not get to invent a new one. It chooses from the chart of accounts you provide. The same applies to known vendor IDs and tax codes: if the ID is not in the supplied list, it does not become a valid match. A Staples receipt might be proposed as Office Supplies. A mixed invoice might need line-level review. The proposal is still a proposal.
The line is review
The answer is not "trust the AI." The answer is: do not let an uncertain read become an invisible entry. untxt. prepares the batch, shows the original beside the extracted data, flags uncertainty, and waits for approval. The bookkeeper still owns the judgment. The software removes the entry work around it.