AI document intake: where reading ends and recording begins

Extraction is a draft too, and a misread digit that flows unreviewed into the books is worse than a slow human ever was.

Every business runs on documents that arrive as files: the supplier's invoice, the till receipt, the bank statement, the government form, the delivery note somebody photographed in a hurry. Each one has to become structured data before the accounts make sense, and for years that meant a person typing. The tools that read these files automatically are now within a small company's reach, and they read faster than any person alive. The decision worth making is not whether the machine can read. It is what stands between the reading and the record.

The honest framing: extraction is a draft. A model produces its best reading of whatever it was given, and its confidence is a property of the typography as often as of the truth. A clean, sharp scan of a wrong number reads beautifully. The machine does not know what the invoice was supposed to say; it knows what pixels usually mean. Treating its output as fact skips the exact step that made the old manual process safe, which was a human who felt responsible for what got typed.

The failure versions repeat across every company that skipped the boundary. The amount misread from a smudged scan and posted straight into the ledger, discovered months later in a reconciliation that should never have been necessary. The form field filled from a similar-looking field above it, so the machine answered the question next to the one asked. And the read nobody questioned because it looked tidy, its confidence borrowed from good print rather than a correct read.

The practice keeps the boundary visible. Extraction writes into a review state, never directly into the record of truth: the reading lands somewhere a person confirms it, and only then does it become a posting. Spot-checks get weighted toward the fields that hurt when wrong: totals, dates, account identifiers, the numbers a supplier or a bank will argue about. And a known error that gets caught is treated as the system working, because the alternative is the same error traveling silently into the books.

We run our own assistant on exactly this rule. It reads what leads and clients send, it structures what it can, and anything ambiguous goes to a person rather than into a record, with no pretense that a machine's first read is the final word. The assistant says plainly that it is AI, and the extraction says plainly that it is a draft. Neither pretends to be the ledger.

Guidance, not guarantees: reading accuracy varies with scan quality, language, and layout in ways no general essay can promise around, and what counts as a proper record of account for your business belongs with your accountant or a licensed practitioner. Keep the review state between the reading and the record, weight the checks where errors hurt, and judge the system by the errors it surfaces, not by the speed of its typing.