Which fields do you need, and where do they come from?
Start by narrowing the scope: which document types the workflow handles and which fields it really needs from them. Scoping like this is where AI document processing projects begin. It is tempting to extract everything, but every extra field needs its own checks and adds review work. For each field you keep, define the expected format, whether it is required and where in the document it is normally found.
Give the workflow an explicit state for information that is missing or ambiguous. If a document has no date or shows two different totals, the system should say so openly and not force a value that merely looks plausible. A blank marked “missing” is easy to deal with; a confident wrong value can travel all the way into your accounts.
Prepare anonymised examples from the document layouts you receive. Include the awkward ones too: scans of varying quality, multi-page documents and corrected versions, where these occur in your work. A system tested only on clean digital files will struggle the first time a photographed invoice arrives.
Why is extraction not the same as validation?
A value can be read correctly from the document and still be unsuitable for your records. The invoice number may be extracted accurately but belong to an invoice that has already been entered; the total may match the page but not the sum of the line items. Extraction answers “what does the document say?”, validation answers “can we use this?”.
Wherever possible, check identifiers, totals, dates and the required relationships between fields with explicit rules: does the supplier exist in the system, do the lines add up, is the date within a sensible range? Send uncertain cases to a reviewer and show them the relevant part of the original document next to the extracted value, so they do not have to search through the file.
Before anything updates a system, decide how duplicate and revised documents will be recognised. And where a clear rule exists, such as arithmetic or matching against a supplier list, keep that step deterministic. There is no reason to ask a model to estimate what ordinary code can calculate exactly, every time.
How do you keep updates traceable and recoverable?
For each document, record the version that was processed, the extracted candidate values, the result of validation, who approved it and what was finally written to the system. When a question comes up a month later, this record shows exactly where a value came from and at which step it changed.
Test the failure scenarios before launch. What happens if the AI provider does not respond in time? What if the receiving system gets the same request twice? Decide who corrects a wrong field once it has been entered, and how a handover that failed halfway is resumed without creating duplicates or losing documents.
When you review results, look at quality field by field and at the total processing effort, not only at how fast extraction runs. A quick extraction step that produces a long review queue can cost more than the manual process it replaced. The figure that matters is the time and effort needed to get a correct record into the system. The same measure applies to any AI workflow automation, from emails to support requests.
Checklist: document checks
- Define the fields, their formats and how missing or ambiguous data is marked.
- Show the source evidence next to each value and use deterministic rules wherever they fit.
- Test duplicates, revised documents and updates that fail partway through.
- Measure accuracy for each field and the total review effort, not just extraction speed.
Example: invoice data on its way to accounting
From an incoming invoice, the system extracts the invoice number, the supplier and the total into a draft. Rules check that the required fields are present and that the arithmetic adds up, and a reviewer sees the relevant part of the original document next to each value.
Only approved data is passed on to accounting. If the same invoice arrives again, it is linked to the existing record, so no second invoice appears in the system.
