---
id: KB-GL-010
url: https://app.codecontract.io/help/glossary/what-is-the-confidence-score
idioma: en
categoria: glosario
subcategoria: documentos
audiencia: usuario
nivel: intermedio
actualizado: 2026-08-13
tambienEn: [es]
relacionados: [KB-DI-004, KB-DI-005]
citadoPor: [KB-DI-009, KB-GL-016, KB-GL-009]
---

# What is the confidence indicator

_How sure the system is about each value it read, and what to do with that._

**Responde a:** what does confidence on an extracted value mean · low confidence on a document · when to review extracted data · ocr confidence percentage

**Confidence indicator** — A measure of how sure the system is that it read a particular value correctly, based on the clarity of the original and on what is plausible for that field.

It is what makes automatic reading usable. Without it you would review everything just in case; with it you review what asks to be reviewed.

## How to read it

| Level | What it means | What to do |
| --- | --- | --- |
| High | The value was read clearly | Use it |
| Medium | It was read, but something does not quite fit | A glance |
| Low | The original did not allow a clean read | Check it against the document |

> [!IMPORTANT]
> A low-confidence value must not be used for anything with consequences — a payment, an official registration, an access decision — without being checked. Automatic reading saves time; responsibility stays with whoever decides.

> [!NOTE]
> If one document type always comes back low, the problem is how it arrives: ask for a better-quality original rather than reviewing each one by hand.

**Does confidence say whether the value is correct?**

It says whether it was read correctly, not whether the paper tells the truth.

**Can I filter by low confidence?**

Yes, and it is the efficient way to review a large batch.

**Does it rise if I correct it?**

Correcting improves reading of subsequent documents of the same type.

## Ejemplos

**A batch of two hundred certificates arrives for review.**

- Filters by low confidence
- Reviews the eighteen that come up

→ Review goes from a full day to half an hour, without giving up on checking what is doubtful.

**Everything is reviewed equally even though almost all of it read well.**

- Sorts by confidence and starts at the bottom

→ Review time concentrates where the risk is.

**A high-confidence figure turns out to be wrong.**

- Corrects the figure and records the correction

→ The indicator guides, but the last word stays human.

**Nobody knows what threshold to review at.**

- Tries a real batch and adjusts

→ The threshold is set from your own data rather than a default.

**One document type always comes out with low confidence.**

- Checks whether the extraction model fits that format

→ The problem is tackled at source rather than at every review.

**Two hundred figures are approved in bulk without looking.**

- Sets the uncertain ones aside before approving the rest

→ Bulk approval stops dragging along what was failing.