Sort incoming documents by type before filing
Documents arriving by email or upload are sorted by a decision model into a fixed set of types, such as invoice, contract, or certificate, and filed in the right folder, with unclear ones left for the office manager.
Author's figure: Author's figure: 6x faster than gpt-5.6-luna for classification and splitting (DocJev). Reported by the builder, not measured by us.
The use case
Small offices receive a steady stream of PDFs and scans: supplier invoices, signed contracts, insurance certificates, tax forms, and statements, all landing in one inbox. Jev reads the extracted text of each document and picks one type from your list. Documents classified with a confidence of 0.85 or higher are renamed and moved to the right folder. Anything lower goes to an unsorted folder for the office manager. Jev does not read amounts into the books, approve invoices, or decide whether a contract is acceptable. The office manager still checks the unsorted folder, and the people who own each folder still act on the documents in it. Start with the types you receive most often, and add new ones only when the other pile grows large enough to matter.
Example workflow
- Documents arrive as attachments in Gmail or uploads to a Google Drive intake folder.
- Make extracts the text from each PDF with an OCR step and sends the first pages to Jev as the state.
- Jev answers a Choice question, what type of document is this, with the allowed answers supplier invoice, contract or agreement, insurance certificate, tax form, bank or card statement, receipt, and other.
- Documents above the confidence threshold are renamed with the type and date and moved to the matching Drive folder.
- Documents below the threshold or labelled other go to an unsorted folder, and the office manager gets a daily Slack summary.
- The office manager files the unsorted ones and notes any repeat mistakes so the type descriptions can be improved.
Tools
- Jev (TypeSafe AI)
- Gmail
- Google Drive
- Make
- Slack
- Google Sheets
Risks and guardrails
- Scanned documents with poor text extraction get misfiled — guardrail: documents with very little extracted text go straight to the unsorted folder.
- A misfiled invoice could be paid late — guardrail: the invoice folder owner also gets a weekly list of anything moved out of unsorted.
- Sensitive documents pass through more tools — guardrail: limit folder access and check each tool's data terms before connecting.
A similar pattern in a public build: https://madewithjev.com/builds/construction-plan-classifier. See also Decision models.
Related recipe
Receipt Capture and Categorization for Bookkeeper Review: An agent extracts vendor, date, amount, and tax from receipts in email, Drive, or photos, suggests a category from your own chart of accounts, and queues drafts in QuickBooks or Xero for a bookkeeper to approve.
Want it set up for you? Request a quote.
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Public build with a similar pattern: madewithjev.com/builds/construction-plan-classifier
- DocJev by Jerry Liu on Made with Jev: madewithjev.com/builds/docjev
- DocJev: original post or repo by Jerry Liu: github.com/jerryjliu/docjev
- Unstract blog: Jev separating invoices from pro formas and routing packet pages: unstract.com/blog/jev-and-where-a-system-one-model-fits-in-document-processing
- Gmail (official page): developers.google.com/workspace/gmail (checked 2026-11-19)
- Google Drive (official page): developers.google.com/workspace/drive (checked 2026-11-19)
- Make MCP server (official page): developers.make.com/mcp-server (checked 2026-11-09)
- Slack MCP server (official page): docs.slack.dev/ai/slack-mcp-server (checked 2026-11-09)