AI for business emails and documents: a controlled workflow, not an unattended inbox
The useful unit is an accepted business result, not a document processed by a model. An email summary that misses a deadline can be more expensive than reading the original. Start by deciding which document class, fields and downstream action are in scope. Keep extraction, verification and execution separate: a plausible answer is not authorisation to update a record or send a report.
Define the evidence chain
For every output, preserve a reference to the source version and the evidence supporting important fields. Distinguish text extracted from a document, a computed value and an inference. When the source is missing or contradictory, the correct output can be an exception rather than a completed record. Replacing an attachment after approval should invalidate that approval; otherwise a safe review screen can authorise different content from the material eventually sent.
Compare three routes on the same representative sample. A template or deterministic rule may be sufficient for stable forms. OCR is relevant when the text itself must be recovered from an image. A language model can help with variation or unstructured content, but its result still needs field checks and an evaluation design. Do not describe an arithmetic demonstration on pre-entered values as proof of PDF extraction.
Acceptance is a workflow contract
| Gate | Required evidence | Blocking condition |
|---|---|---|
| Decision | Accept a verified result, not merely generated text | No silent export |
| Intake | Allowed source and source version | Unknown or unauthorised file |
| Extraction | Required fields and source references | Missing, conflicting or unsupported fact |
| Validation | Business rules and arithmetic | Unresolved mismatch |
| Release | Current human approval and permitted destination | Changed source or duplicate action |
Build an evaluation set containing ordinary documents and important exceptions: amended versions, unreadable scans, unusual tax lines, long email threads and malicious instructions embedded in content. Separate examples used to configure prompts from held-out tests. Report error severity and denominator, not an unexplained accuracy percentage. Review minutes and correction effort belong in operating cost.
Make the boundaries observable
- Record document identifier, source version, workflow version and processing state without exposing unnecessary personal content in logs.
- Use read-only access first; introduce outbound messages and accounting or CRM writes as separate authorised capabilities.
- Treat instructions inside a document as untrusted content rather than permission to change recipients or reveal other files.
- Provide a human exception queue with a named owner, not a fallback mailbox nobody monitors.
- Reconcile retries with the destination so a timeout does not create a second payment record, report or email.
Measure cost per accepted result
Count eligible inputs, correctly accepted outputs, rejected cases and rework separately. Total attributable cost includes configuration, data preparation, inference, review, corrections, hosting and support. Divide by accepted results only when that denominator is positive, and compare the same required quality with manual and deterministic routes. A fast system that moves more work into the exception queue may not be cheaper.
For ongoing support, agree who owns changes in templates, validation rules, model versions and source systems. A sampled quality review and replay suite should run after significant changes; automatic upgrades without acceptance evidence can undo the original business case. Support coverage, response commitments and third-party usage are quoted separately, not hidden inside the entry diagnostic.
A practical discovery brief contains one document class, a redacted authorised sample, required fields, current handling effort and the action that must remain controlled. LEMAIT's diagnostic is EUR 350–600 for one agreed process; a pilot is a separate bounded scope. Our Namirial and Epilot reviews analyse public supplier stories, not projects those businesses purchased from LEMAIT. This operational guide does not certify data protection or replace professional approval of legal or accounting documents.
Prepared with AI assistance and checked against the linked public sources by LEMAIT. Scenarios are illustrations, not customer results. No ranking, ROI or legal-compliance guarantee.