Agentic document intelligence: OCR, routing, review, and repair loops

Document automation often fails because it treats extraction as the whole product.

In reality, users do not only need OCR. They need documents to be routed, reviewed, corrected, validated, and turned into reliable workflow data.

That is the idea behind Dossier: an agentic document intelligence system where OCR is one step inside a larger operating loop.

The workflow

A practical document system needs to handle messy inputs:

Upload document
    -> detect document type
    -> choose OCR/provider route
    -> extract fields
    -> validate against rules
    -> flag risks and missing data
    -> human review when needed
    -> repair failed fields
    -> export structured result

The agentic part is not “let the AI do everything.” It is the ability to plan the next step, select the right provider, retry when extraction fails, and ask for review when confidence is not enough.

Why routing matters

Different documents need different strategies.

A clean invoice, a scanned contract, a handwritten form, and a photo of a document should not necessarily use the same OCR path. Some providers are better for tables. Some are better for layout. Some are cheaper. Some are faster.

Provider routing lets the system choose based on document type, confidence, cost, and latency.

Review and repair loops

The most important feature is often not extraction. It is repair.

If a field is missing or suspicious, the system should not silently produce bad data. It should:

• show the evidence region;
• explain why confidence is low;
• ask for human review;
• store the correction;
• re-run validation;
• improve future routing decisions.

This turns document AI into a controlled workflow instead of a black box.

Measuring value

Useful metrics include:

• percentage of documents completed without manual review;
• average review time;
• extraction accuracy by field;
• repair success rate;
• provider cost per document;
• number of high-risk fields caught before export.

These metrics connect AI quality to business operations.

What I learned

Agentic document intelligence is less about a single powerful model and more about orchestration.

OCR gives text. The product needs decisions: which provider, which fields, which risks, which repair path, and when a human should step in. That is where the system becomes useful.

Related project: Dossier.