An engineering consultancy we work with produces inspection and evaluation reports from site surveys. Consultants capture findings on a tablet, and M-Files generates a Word report from that data. The generated report is a solid first draft, but it is a template filled in by rules. Before anyone can review the engineering content, someone has to delete the instructions meant for the author, resolve the conditional wording and tidy the formatting.
That clean-up took staff time on every report, and it was the same work each time. TEAM IM set out to give it to Claude, Anthropic's AI model, without taking any decision away from the people who sign the report.
The review lives in the M-Files workflow. We added a workflow state called Claude review. When a report reaches it, a vault application sends the document to a review API that TEAM IM built and that runs in Microsoft Azure. The vault authenticates to the API through Microsoft Entra ID.
The API has no screens. It takes a document and a named task, runs Claude with a skill that holds the firm's own clean-up rules, and returns the result. A typical report takes about 30 seconds.
The vault application then checks the cleaned document in as a new version and moves the report to the next state. The version comment records what happened, so the history shows who changed what, including Claude.
Claude never overwrites a report. It returns the document with Word tracked changes, and the reviewer accepts or rejects each one. The people who already approve reports still approve them. They start from a cleaner draft.
If a review fails, times out or finds the document was edited while it was out, the report moves on unchanged and the version comment says why. A problem with the AI step never strands a report or blocks the workflow. While Claude has a report, the review state sets permissions so nobody can edit it.
The workflow calls Claude at four points:
Each pass is its own workflow state and a named skill on the same API.
Keep the AI step out of document generation. The review is a separate module with its own task queue and its own configuration. Switching it off sends reports straight on to the next state, and a fault in the review cannot stop a report being generated.
Consider which improvements belong to report generation versus Claude. Theoretically every fix that Claude is making to the report could instead be coded into the deterministic report generation. The decider is flexibility and time investment. Updating a Claude skill is fast and flexible. Code for new rules must be strictly defined for every case. Claude is effective at filling gaps that are difficult to formalize into logic.
Let the workflow decide. Which reports go to Claude, and when, is set in the M-Files workflow and the vault application configuration. The firm's administrators change it in configuration, not in code.
Write the outcome where people already look. Reviewers check the version history, so that is where the result goes. We considered Word comments and rejected them, because they would add more clean-up work for the reviewer.
Measure the cost early. We ran capped test jobs before any real report went through. With prompt caching, a clean-up pass costs well under a dollar per report.
Most documents in an M-Files vault already move through a workflow with named states and permissions. That makes the workflow a good place to put AI. Each state says when Claude acts, which documents qualify and where the result goes, and the vault history records all of it.
If you run M-Files and spend staff time correcting generated or incoming documents before review, talk to TEAM IM about what Claude could take on in your workflow. Our implementation team designs and builds changes like this one inside your existing vault.