BUSINESS DOCUMENTS
Read your scans, return real office files.
PDFs come in as structured text, and what goes out opens in Word, comments and tracked changes included.
A document is a pack of files
An onboarding file is not one file: it is an ID document, a proof of address, a signed contract. Standards treats it as what it is, a pack.
Each item in the pack goes through its own processing and carries its own state. The extracted text then reads as a single document, in the order of the items.
The folder trees are laid out in advance. A folder template is instantiated on every record, so the same items always end up in the same place, whoever uploads them.
Scans and native PDFs take the same path
Reading accepts PDF, PNG, JPEG and WebP. One route for the document a program produced and for the photograph of a 1998 contract.
What comes out is structured text, page by page. A heading stays a heading, a table stays a table, and the layout does not dissolve into a soup of lines.
Reading runs in the background. The upload does not wait, and one unreadable item does not stop the others from being processed.
What comes out is a real file
Not HTML dressed up as a document. A DOCX built with the reference library, a PPTX, a spreadsheet, and the PDF obtained by office conversion rather than by an approximate print.
The file is produced in the agent's environment, then stored in your workspace: it becomes an attachment on the conversation, and links to the records it concerns.
Pick a document up without rewriting it
An incoming contract is not redone from scratch. It is opened at the level of the format itself, changed, and closed again. What was not touched stays strictly identical.
Native Word comments and tracked changes are supported, replies to comments included. Your lawyer receives an annotated document, not a summary of what should have been annotated.
The produced file is validated against the official schema of the format before it comes back to you.
How it works
A hundred and forty signed contracts to re-read before an audit.
The items are uploaded in bulk, into the folder of the record they belong to. The folder template files them in their place.
Each item is read: scans and native PDFs by the same route, page by page, in the background, with nobody waiting.
The text becomes queryable. The agent reads the pack as one continuous document, and cites the item and the page a claim comes from.
It annotates what is off. Discrepancies are placed as native Word comments in the original contract, corrections as tracked changes.
The deliverable returns to the workspace: one file per contract, validated against the format schema, attached to the record, ready for a human read.
The other pieces of this chapter
Reflexes
The computation an agent wrote once, replayed identically.
Skills
Your method in plain words, opened at the right moment.
Environments
An isolated, tooled machine that sleeps without losing state.
Hybrid search
Meaning and wording weighed together, filtered in the engine.
Agent memory
Memories that are records, not a black box.