What is an extraction confidence score?
The confidence score is a value from 0 to 1 attached to every extracted field: unit price, total, name, number. It estimates how reliable the value read is and separates the data you can process automatically from the data worth a review. It is a decision aid, read field by field.
How is the confidence score calculated?
Koncile computes the score from several signals: how sharp the source document is, how complex its layout is and how the field was extracted. A blurry scan, heavy compression or an unusual structure lowers the score. The goal is to spot values that may be wrong and to prioritize reviews.
At what score should you review a value?
There is no single threshold. The right level depends on the business impact of the field: an amount does not carry the same stakes as a description. These are the guidelines we recommend.
Above 0.95
The value can be used automatically. A spot check remains useful on sensitive financial fields.
Between 0.85 and 0.95
Extra attention is advised, especially for amounts, sales tax rates and account identifiers.
Below 0.85
A review is strongly recommended, especially for data with a high business impact.
How do you use the score in your automations?
The score lets you write validation rules: send trusted values straight through, queue the rest for review. A few uses by team.
Finance and accounting
Focus reviews on amounts, sales tax and due dates, line by line on invoices.
Purchasing and logistics
Secure prices, quantities and SKUs on order and delivery lines.
HR and compliance
Spot risky administrative or regulatory data before it enters your systems.
Does a low score mean the value is wrong?
Not necessarily. A low score signals a hard extraction, often tied to document quality or a complex structure. The chance of error is higher, so a review is recommended. Binary fields (yes or no, present or absent) need special care, since an error there is always complete.
The score is available on every field, in the app and in the API response, including for handwriting.
An AI OCR built for security
Your documents are never used to train a model. Fully encrypted software, secure storage of extracted data, a SOC 2 certified and GDPR compliant platform.
Confidence score FAQ
Is the confidence score reliable?
It combines several signals from the document and the extraction to give a consistent estimate. Use it as a decision aid to secure your automations.
How should you read the confidence score?
Field by field, based on business impact: a unit price at 0.95 matters more than a descriptive field at the same score.
At what threshold should a value be reviewed?
As a rule, a value at 0.95 or below deserves attention, and a review is strongly recommended below 0.85.
Does the score depend on document quality?
Yes. A blurry, poorly scanned or heavily compressed document makes extraction harder and lowers the score.
Is the score available for every field?
Yes. Every extracted field carries its own confidence score.
Does the score work for unstructured documents?
Yes. Reading it then depends more on context and on document quality.
