Document Processing Agent · Use case
How can a government agency turn paper documents into data with an AI agent on its own infrastructure?
A document processing agent reads scanned forms and letters, extracts the fields you define and returns them as structured data, with the page each value came from. A value it is unsure of goes to a person to check. The agent runs on infrastructure the agency controls, so citizens' documents are not sent to a public AI service.
Paper forms and scans.
Retyped into the case system.
- ✗ 01Staff retype paper forms into the case system, and scanned files cannot be searched.
- ✗ 02The documents hold citizens' personal data and cannot leave government infrastructure.
Where it stops for a person
When a value is unclear or a page cannot be read.When a value is unclear, a page cannot be read or the document is not a type it knows. A person checks those.
It reads and structures.
A person checks the doubts.
Step · 01
Scanned documents arrive through File Upload.
Step · 02
The agent reads each page.
Step · 03
It extracts the fields you defined for that document type.
Step · 04
It marks each value it is unsure of.
Step · 05
A person checks the marked values before the data is used.
Reads from · connectors
File Upload
Your own APIs
Documents arrive through File Upload and are read with Markitdown.
Acts through · MCP servers
Markitdown
It returns structured data. Writing that data into a case system goes through your own APIs, where you choose to connect them.
What you set up
- List the document types and the fields for each.
- Decide who checks the marked values.
- Write a test set of scans with the correct values.
Where it runs
The agent, its knowledge store and the model run on infrastructure you own or control. You choose the model and serve it yourself, so questions, documents and answers are not sent to a public AI service.
How you get it today
This page is labelled Use case: the agent is built in Katonic Studio. You describe the job, connect what is named above, add test questions and release the agent once it passes them. When the Catalog holds a tested template for this job, this page will carry the Template label and show how to take it.
Does it read handwriting?
Not as described here. Reading handwriting needs a handwriting service, and the one the Catalog lists is hosted outside your environment. A page the agent cannot read goes to a person.
How do we know a value is right?
Each value points to the page it was read from, and you score the agent against a test set of scans.
Where are the documents processed?
On infrastructure the agency controls, with a model it serves.