Products
Document AI
Read documents structurally instead of retyping them
document-ai.txtas is → to be
Today a document is scanned, opened, read, retyped and checked, and corrected three times. Afterwards it is uploaded, read field by field, and a person confirms only the fields the reading was unsure about.Illustrative hours, not a measurement.
Invoices, delivery notes and forms arrive as paper or PDF, and somebody types them into the system. Document AI reads each one field by field and says how sure it is about every field. The unsure ones go in front of a person; the rest do not. Nothing is saved until somebody confirms it, and nothing is typed twice.
The same figures stop being typed twice, and the fields a person still checks are the ones the extraction was unsure about rather than all of them.
How it runs
- the document is read field by field, and each field says how sure it is
- the fields it is unsure about are shown to a person instead of guessed
- nothing is saved until somebody confirms it
- the confirmed result becomes the record, with nothing retyped
Where it does not fit
- poor scan quality
- documents with legal significance requiring full human review
- one-off formats not worth modelling
Size: a medium piece of work — an estimate from the pattern, not from your business.
What we would want to know before committing
- document types and monthly volume
- which fields anybody actually uses afterwards
- where the extracted data has to land
- current error and rework rate, if known
- run a sample of real documents before committing to a scope
- agree what happens to a low-confidence field
- confirm the destination system accepts writes