'80% Confidence': What Does It Tell You About an Extracted Invoice?

The number on the screen

Many generic extraction tools show a confidence score. It may be per field, per invoice or per batch. A result might read "invoice total: 80% confidence."

It looks informative. It is worth asking what it actually lets you do.

We should say clearly that our own work is not built around confidence scores. We write rules that read specific values and run specific checks. What follows is how we would read such a number if a tool showed us one.

Questions the number leaves open

Take "80%." Several things are unclear:

What is wrong with the other 20%? Is a digit uncertain, a field missing, or a whole region unclear?
Where is it? Which field, which line, which page?
How would I know? If the tool shows 80% for an invoice, do I look at all of it?
What score is good enough? Is 80% fine for one vendor and risky for another?
What does the score measure? The tool's own estimate of how sure it is, or a test against something you defined?

A score is a summary. A summary does not tell a reviewer where to look.

What a reviewer actually needs

Think about the person at the desk. They have a stack of invoices and a limited amount of time. What helps them?

Something like this:

Invoice 4411: GST check failed. Tax shown does not match tax calculated from the rate.
Invoice 4415: GSTIN format check failed. Value has 14 characters.
Invoice 4420: state code does not match the state on the invoice.

Each line says which invoice, which check and what was found. The reviewer opens that one invoice, looks at one field and fixes it. It takes a minute.

Compare that with "Invoice 4411: 80%." The reviewer reopens the invoice and reads the whole thing, hoping to spot the problem. They may find it, or they may not.

Confidence and checks answer different questions

It helps to separate two ideas:

Confidence is the tool's feeling about how well it read something.
A check is a test against a rule you defined, which either passes or fails.

A value can be read with high confidence and still be wrong for your purposes. A tax figure can be read exactly as printed and still not match the line amounts. A GSTIN can be read clearly and be invalid.

Likewise, a value can have a low score and be correct.

For accounting, checks that tie back to numbers you can verify, such as line math, tax math, totals and master data, usually matter more than a general score. They tell you whether the figures agree with each other, which is what you rely on when you post.

Does a high score mean you can skip review?

We would be careful about that. If a team decides "above 90% we do not look," the decision rests on a number whose meaning is not fully visible to them. Over time, a few wrong values pass through because they were confidently misread.

If you use a confidence threshold, test it. Take a batch where you know the right answers, see where the threshold lets errors through, and decide whether that is acceptable.

What we do instead

Our scripts extract values by fixed rules for the layouts you actually receive. Then they run the checks you have set. A value that fails a check, or cannot be found, is flagged by name in a list for review, with the invoice and the check shown.

That does not make every invoice right. It means the things you defined as important are tested every time, and when one fails you can see which. Anything not set as a check is not checked, and we say so up front.

Questions to ask any vendor

What does the confidence number measure, and how is it produced?
When a score is low, does the tool say which field or check is the problem?
Can I set my own checks, and are failures reported by name?
What score do you suggest as a threshold, and what is that based on?
Can I see how often a high-confidence value was still wrong in a test batch?

If your team currently scans a whole invoice to find out why something was marked uncertain, you can show us your workflow and we can discuss what named flags would look like for your documents. A sample can go with the document automation assessment.

FAQ

What does an 80% confidence score mean on an invoice?

It means the tool estimates it is fairly sure about a value. It does not say which part is uncertain, or whether the figure fails any check you care about.

Is a confidence score the same as a validation check?

No. Confidence is an estimate of how well something was read. A check is a pass-or-fail test against a rule you define, such as line math or GSTIN format.

Should I skip review above a certain confidence score?

Be careful. Test any threshold on a batch where you know the right answers before relying on it.

Want named flags instead of a bare score?

Show us how you review invoices today. We will sketch what a flagged list would look like for your checks.

Show us your workflow