---
id: KB-DI-001
url: https://app.codecontract.io/help/documents-and-ai/how-the-ai-reads-your-documents
idioma: en
categoria: documentos-ia
audiencia: usuario
actualizado: 2026-08-13
tambienEn: [es]
relacionados: [KB-TL-001, KB-PR-002, KB-DI-002]
citadoPor: [KB-DI-003, KB-DI-002, KB-CD-001, KB-MA-001]
enLaApp: https://app.codecontract.io/documents
---

# How the AI reads your documents, and when to trust it

_What it extracts, what the confidence level means and who has the final word._

**Responde a:** how does the AI read documents · what is the confidence level of an extracted field · the AI misread a field what do I do · automatic data extraction from invoices · OCR for scanned documents

When a document arrives in a case it does not just sit there as a file: it is read, and the fields you asked for are extracted. Invoice number, amount, due date, company name. That is what later lets you search for "every invoice over €5,000" instead of opening fifty PDFs.

**En corto**

- The AI proposes the data; you confirm. It never closes a field on its own.
- Each field carries a confidence level telling you where to look.
- If you correct a field, the record shows that you corrected it and when.
- A blurry or skewed document lowers confidence; that is a warning, not a failure.

**The journey of a field** — Confidence does not decide for you: it tells you how hard to look. Confirmation is always human.

## What the confidence level means

It is not a quality score for the document: it is how sure the system is that it read that specific field correctly. The same document can have the amount at high confidence and the date at low, because the date was handwritten in a margin.

| Confidence | What it usually means | What to do |
| --- | --- | --- |
| High | Printed field, sharp, where it was expected | A glance is enough |
| Medium | It was read, but the format or position was unusual | Check it before confirming |
| Low | Shaky photo, rotated document, handwritten or struck-through field | Review it fully; sometimes it is worth asking for a better copy |

> [!WARNING]
> Low confidence does not mean the value is wrong: it means the system is not committing. Confirming it without looking is exactly the use it should not get.

## What to do when it reads something wrong

1. **Correct the field by hand** — It is immediate, and the record shows the final value came from you.
2. **See whether the document can be improved** — A photo with shadow or backlight reads badly. Asking for a retake costs less than correcting twelve fields.
3. **If it repeats with the same document type, say so** — A one-off failure on a bad photo is not the same as a pattern with a specific form, and the second can be tuned.

## What it does NOT do

- It does not decide whether a document is valid: you or your process do.
- It does not reject documents based on their content.
- It does not modify the original file, which is kept exactly as it arrived.
- It does not read in SmartCheck: there, documents are sealed as-is, uninterpreted.

## Frequently asked questions

**How long does reading a document take?**

Seconds. If it takes much longer it is not thinking: something has failed and it is worth reloading.

**Does reading a document consume credits?**

Yes, automatic reading is one of the actions that consumes credits. Correcting a field by hand does not.

**Can I ask for fields that are not in the document?**

You can ask for them as data for the person to fill in. If they are not on the page, reading does not invent them: it leaves them empty.

**Does it read documents in other languages?**

Yes. What affects reading most is image quality, not language.

**What if the document has several pages or several invoices?**

All of them are read. When one document contains several occurrences of the same group of fields, each is extracted separately.

## Ejemplos

**An accountancy receives 300 invoices a month and needs the number, date and net amount from each.**

- Defines those three fields on the process request
- Reviews only the ones flagged medium or low confidence
- Confirms the rest in bulk

→ About 30 get reviewed by hand instead of 300, and the ones reviewed are exactly the ones that needed it.

**Everything read is approved without looking.**

- Reviews whatever the AI flags as uncertain

→ The work shifts from keying to checking.

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

- Corrects the figure and records it

→ The indicator guides and the last word stays human.

**One document type always reads badly.**

- Checks whether the format matches what is expected

→ The problem is tackled at source.

**A hundred invoices a month are keyed by hand.**

- Lets them be read and reviews the uncertain ones

→ Time goes on checking rather than copying.

**The reading is distrusted and everything is checked.**

- Filters by confidence and starts at the bottom

→ Review concentrates where the risk is.