AI document extraction accuracy is less about how many fields a tool fills and more about whether you can trust the ones it fills without checking. An extractor that says "please check" on the fields it is unsure about is worth more than one that is always confident, because you know exactly where to look. A confident wrong value is the expensive kind of error: it slips through.
What does "accuracy" mean for AI document extraction?
When someone says a tool is accurate, ask: accurate at what? There are two very different numbers hiding in that word.
- Of all the fields on the document, how many did it get right?
- Of the fields it gave you without a warning, how many were right?
The first number rewards a tool for filling every box. The second rewards it for being honest about doubt. If you run accounts payable or process shipping documents, the second one is what decides whether you can post data without reading every line yourself.
What is the difference between precision and coverage?
Two plain terms make this easier to talk about.
Precision is how often an accepted value is correct. High precision means when the tool says "this is the invoice total", you can believe it.
Coverage is how much of the document the tool fills in without asking for help. High coverage means fewer fields land on a person's desk.
You want both, but they pull against each other. Any tool can reach full coverage by guessing on every hard field. Precision is what tells you whether those guesses are safe.
| Extractor that always answers | Extractor that flags doubt | |
|---|---|---|
| Fields filled automatically | Every field, every time | Most fields; uncertain ones are flagged |
| When a value is wrong | Looks exactly like a right value | Usually flagged with a reason |
| How you find errors | Check everything, or find out later | Check the highlighted fields |
| Review effort | Unclear, so you end up checking all of it | Focused on a short list |
| Cost of a miss | Wrong payment, wrong customs value, rework | A few seconds to confirm a field |
The left column looks better in a demo. The right column is the one you can actually run a process on.
Why is "please check" better than a confident guess?
Think about where errors end up. A flagged field stops in front of a person who has the document open. They confirm it or fix it, and it is done.
A confident wrong field does not stop. It flows into the ledger, the payment file, or the customs entry. You find it when a supplier chases an underpayment, when a bank rejects a transfer, or at month-end when something doesn't reconcile. By then, fixing it means emails, credit notes and reversals.
SensiDocs is built on that trade. Every field is either accepted or flagged, and every flag has a plain reason: not found, failed a check, contradicted, bad format, not grounded in the document, or low confidence. It would rather ask than guess.
How does a second independent read improve accuracy?
Some fields are clearly readable. Others are borderline: a smudged digit on a scan, a total printed close to a subtotal, a date that could be read two ways.
For those doubtful fields, SensiDocs does a second read. The key word is independent. The second read never sees the first answer, so it cannot just agree out of habit. A value is accepted only when both reads land on the same thing. If they disagree, the field is flagged for a person.
This is the same reason auditors don't hand the reviewer the preparer's answer first. Two separate looks that agree are much stronger evidence than one look checked against itself.
Why should hard checks outrank model confidence?
AI models, including the Claude models from Anthropic that SensiDocs uses, are very good at reading documents. But reading correctly and the document being correct are different things. A model can read every digit on an invoice perfectly and the invoice can still be wrong, because the supplier made a mistake.
That is why SensiDocs runs deterministic checks next to the model:
- Totals: line items must add up to the subtotal, and subtotal plus tax must equal the total, allowing only for rounding of a cent or two.
- US bank routing numbers: validated against the ABA checksum.
- Container numbers: validated against the ISO 6346 check digit.
- EINs: checked for the correct format.
When a check fails, the field is flagged, no matter how confident the model was. Arithmetic does not have off days. A routing number that fails its checksum is wrong, full stop, and no level of model confidence should let it through.
What should you ask a document extraction vendor about accuracy?
Before you trust any accuracy figure, yours or anyone else's, work through this checklist.
- How was accuracy measured? On which documents, and how messy were they?
- Does the number count all fields or only accepted fields? These answer different questions.
- What happens when the tool is unsure? Does it flag the field, or fill it with its best guess?
- Does every flag come with a reason? "Low confidence" alone is less useful than "line items sum to 2,305.00, subtotal reads 2,233.00".
- Do arithmetic checks overrule the model? Or can a confident reading override a failed sum?
- Are identifiers checked by their real rules? Routing numbers, container numbers, tax IDs.
- Is there an independent second read for doubtful fields? Does it see the first answer?
- Can you trace each value to the page? Values that cannot be grounded in the document should be flagged.
- How does review work? Can a person fix a field in place, or do they export and edit elsewhere?
A vendor who can answer these clearly has thought about the cost of being wrong. A vendor who only quotes one big percentage has not told you much.
How should you test extraction accuracy yourself?
Use your own documents, and pick the awkward ones. A crooked scan. A phone photo. A long invoice with many lines. A bill of lading with a faint container number. At least one document where you already know there is a mistake.
Then score it the practical way:
- Of the accepted fields, were any wrong? This is the number that matters most.
- Of the flagged fields, was the reason clear?
- How long did it take to clear the flags?
If accepted fields are right and flags are quick to clear, you have a tool you can build a process around.