Extracting data from a stack of documents

Thirty invoices, fifty job applications, a hundred free-text responses to a questionnaire. Turning that into a table is the most spectacular time-saver an AI can offer.

Define the columns BEFORE

The mistake is saying "extract the important information". Each document will then produce different columns and nothing will be comparable.

"For each invoice, extract exactly these columns: supplier, number, date, amount excluding tax, VAT, amount including tax. If any information is missing, write ABSENT and invent nothing."

The "ABSENT" is essential: without an explicit instruction, a missing field will be filled with a plausible value, and you will never know which one is real.

Process in batches

Five to ten documents at a time. Beyond that, quality drops and you can no longer reasonably verify. You will assemble the batches afterwards.

Spot-checking

Don't check everything, that would be as time-consuming as doing it by hand. Pick three rows at random and verify them in full against the original document. If all three are correct, the batch is probably fine. If one is wrong, redo the whole batch: errors rarely come alone, they stem from a misunderstood format.

Watch out for personal data

Fifty job applications means fifty people. On a consumer-grade service, that is exactly the case you should not process without a contracted professional version.

Do this now

Take five documents of the same type. Define your columns, request the extraction with the ABSENT instruction, then check two rows at random.