95% Accuracy vs. 80% STP: What IDP Vendors' Numbers Actually Mean

Accuracy and STP rate come up in almost every IDP sales discovery call, and they’re two of the numbers buyers find most confusing. At first glance, both can look like measures of accuracy. It’s easy to assume they mean the same thing and can be used interchangeably, especially if you’re buying a data extraction tool for the first time. Our prospects ask about this often, so here’s what each number actually means.

A buyer sees “95% accuracy” on one slide and “80% straight-through processing” on the next. Then comes the obvious question: which number is true?

Both can be. They measure different things, at different levels, using different denominators. Accuracy measures whether extracted data is correct. STP measures whether that data moves through the workflow without human intervention.

1. Accuracy vs. STP: what they are and how they differ

Accuracy measures extraction quality. STP measures how much processing happens without human intervention. Both can be measured at the field level and the document level.

The confusion starts when they’re treated as the same claim.

Accuracy

Field-level accuracy

Field-level accuracy measures how many extracted fields are correct out of all the fields extracted. Results are checked against verified answers, so the number tells you how often the model gets it right.

If a system extracts 100 fields and 95 are correct, its field-level accuracy is 95%.

Several things influence it:

• Continuous model training on large volumes of real-world documents
• Using the best-suited model for each extraction, based on document type and industry
• Performance on previous extractions
• Internal checks and verifications the system runs on its own

Document-level accuracy

Document-level accuracy measures how many documents have every required field extracted correctly, out of all documents evaluated.

If a system processes 100 documents and every required field is correct in 80 of them, its document-level accuracy is 80%.

A document can have high field-level accuracy and still fail the document-level accuracy check. If 19 out of 20 required fields are correct, field-level accuracy is 95%. But because one required field is wrong, that document does not count as fully accurate.

Confidence scores

Infrrd’s IDP system assigns a confidence score to every extracted field. This is the system’s estimate of how likely that value is to be correct.

Accuracy tells you how often extracted values are right when checked against verified answers. Confidence tells you how sure the system is about a particular value before anyone checks it.

A high confidence score does not guarantee that a value is correct. A correct value can also receive a low confidence score and be sent for review. 

STP

STP measures the share of fields or documents processed without human review, correction, or approval.

In Infrrd’s system, every extracted field carries a confidence score from 0 to 1, where 1 is the highest confidence. Users can set the confidence threshold anywhere from 0.1 to 1. Fields that meet or exceed the threshold and pass any required validation checks move automatically to the next stage; fields below it are routed for human review.

Field-level STP

Field-level STP measures how many extracted fields move through the defined workflow without human intervention, out of all extracted fields.

If 7 out of 10 fields pass without human review, the field-level STP rate is 70%.

The remaining three fields need review. This means some fields can go straight through even when the whole document cannot. 

Document-level STP

Document-level STP measures how many documents move from intake to your downstream system without human intervention, out of all documents received.

If 70 out of 100 documents go through without review, correction, or approval, the document-level STP rate is 70%. 

If even one field needs human review, the document does not count toward document-level STP, even if every other field passes automatically. 

In a four-field document, one wrong total brings field-level accuracy down to 75%. The document also fails the document-level accuracy check. If a person corrects that total, it does not qualify for document-level STP either.

2. Why a high-accuracy vendor can still have a low STP rate

Two things can pull document-level STP below field-level accuracy: errors across different fields, and correct answers that still get sent for review.

A field may be correct, but is the system confident enough to release it without a human reviewer?

For critical fields in complex documents, accuracy alone isn’t enough. The system also needs enough confidence and supporting validation to decide whether the value can proceed without review.

When a field is correct, but the system’s confidence is low, it still sends that field to a reviewer. More review means lower STP. That’s how a system can deliver high accuracy and still have a lower STP rate: some accurate fields still need to be checked.

Right isn't enough. The system has to know it's right.

This is the part most accuracy claims leave out. A document only goes straight through if it meets the conditions for automatic release. Even if every field is correct, one low-confidence field can still send the document to a person.

That document can count toward document-level accuracy while failing to qualify for document-level STP.

STP therefore depends on two things accuracy doesn’t measure: does the confidence score reflect how reliable the result is, and is the release threshold set appropriately? A model that is often right but can’t reliably tell when it’s right will keep sending good work for review.

A vendor quoting field accuracy alone is describing extraction performance. It isn’t describing how much work your team will need to review.

3. What actually drives STP up

Better extraction helps. So does giving the system reliable ways to check its own work and determine what needs human attention: 

1. Correct splitting and classification. A document that’s split incorrectly or assigned to the wrong type can fail before extraction even starts. Field accuracy alone does not give you the full picture of these failures.
2. Honest, per-field confidence scores built into the system. As covered above, accuracy alone isn’t enough. The system needs to assess its confidence in each field and automatically send anything below the threshold for review. Inflated confidence pushes errors downstream. Overly cautious confidence floods the review queue.
3. Validation rules that confirm values. Line items that add up to the total. A loan amount that matches across three documents. A vendor ID that exists in your master data. Each successful check provides evidence that can help the system release a field it would otherwise hold for review.
4. Thresholds set by business impact, not one global number. A wrong invoice total costs more than a wrong fax number. Stricter thresholds for critical fields and more flexible ones for lower-risk fields can reduce unnecessary review while maintaining the required quality.
5. Corrections that feed back. Every reviewer correction provides information the system can learn from. If those corrections don’t feed back into the model and rules, the same documents can fail in the same way next month.
6. Your document mix. STP depends as much on your documents as it does on the vendor’s software. A rate measured on clean, typed invoices tells you little about handwritten forms or 300-page bundles.

These factors affect both field-level and document-level STP. Reducing review on individual fields saves work, but a document only goes straight through when nothing in its workflow requires human intervention.

4. How to make vendors' numbers comparable

Don’t just ask vendors for their accuracy, STP, or confidence numbers. Ask how they define each one and whether they’re measuring fields or whole documents.

These questions will help you understand what you’re comparing: 

• How do you define field-level accuracy? How do you account for incorrect, missing, or partially extracted values?
• How do you define document-level accuracy? Does every required field have to be correct?
• Are you quoting field-level STP or document-level STP? A high percentage of fields skipping review does not mean the same percentage of documents went through untouched.
• For STP, how does your system decide which documents or fields need human review?
• What counts as a human touch? Does a reviewer glancing at a flagged field count? What about a rule exception?
• Whose documents were used? Your sample set, a public benchmark, or the vendor’s own clean data?
• How many fields per document, and how many pages? A 95% rate on 8-field receipts and 95% on 60-field loan packets are different claims.
• When was it measured? During a pilot or after six months in production, with real volume and drift?
• What happens to documents that fail? Are they included in the STP denominator or dropped as “out of scope”?

A simple starting point: send every vendor the same 200 documents from your own workload and ask for field-level accuracy, document-level accuracy, field-level STP, and document-level STP on that set.

Metrics glossary

Seven terms, one line each, to help you read every vendor’s numbers the same way.

Conclusion: The easiest way to tell if an extraction system delivers

Beyond the metrics, four questions help you understand whether a data extraction system actually delivers:

• How many documents or fields reach downstream systems without manual intervention?
• If the vendor claims a high STP rate, how do they check that the data moving downstream meets your accuracy requirements?
• Can the system maintain consistent STP over time?
• What will they do when STP rates stall?

Ask these questions in every discovery call with an IDP vendor.

What matters is how much manual work comes off your team’s plate, so they can move faster and spend less time staring at documents, comparing values, and checking data by hand.