Forgeron3
/ MethodNov 17, 257 min read

Preparing your business to deploy an AI assistant

An AI project fails before the first line of code. Here are the five decisions to make before signing a vendor — and the ones too often taken in the wrong order.

F3
The Forgeron3 teamMarseille & Paris

1. Take stock of your document estate

Before picking a vendor, list your information sources: DMS, NAS, network shares, Google Drive, Dropbox, executive mailboxes, PST archives, scans still on paper. For each: estimated volume, dominant format, time period covered, internal owner, sensitivity.

This inventory takes half a day. It will save you two months of later course-correction.

2. Pick a pilot that scares no one

The ideal pilot has three properties: it’s useful to the team, it’s not critical if it errs, and it has a reasonable volume (between 200 and 2,000 documents).

Three examples that work well:

  • Onboarding assistant: feeds new hires with procedures, welcome handbook, HR FAQ.
  • Doctrine assistant (accounting firms, legal): fed with the technical knowledge base, for internal questions.
  • Catalog assistant: products, technical sheets, general terms — useful for sales and support.

To avoid as a pilot: anything touching customers or finance directly. You’ll build those assistants in a second wave, once the team trusts the tool.

3. Frame compliance before tech

Five questions to ask yourself before ingestion:

  1. Are there personal data in the planned documents?
  2. Are there sensitive data (health, minors, criminal)?
  3. Is the vendor’s hosting in France, under French jurisdiction?
  4. Does the vendor commit in writing to not train its models on your data?
  5. Is the DPA (GDPR article 28) signed before ingestion?

If any answer is “I don’t know”, stop everything and clarify. More detail on our security & GDPR page.

4. Mobilize three people, not ten

An AI assistant project succeeds with three roles, not a steering committee of fifteen:

  • A sponsor (executive or operations lead) who arbitrates trade-offs and unblocks document sources.
  • A business lead who knows the content and will validate answers during the acceptance phase.
  • A technical lead (IT or provider) who handles access, SSO authentication, ACLs.

No need for more on the pilot. Add the rest at the rollout phase.

Bottom lineThe broader the project team, the slower the pilot. Three committed people beat fifteen consulted ones.

5. Go / no-go criteria for scaling up

After three to six weeks of pilot, you should be able to answer yes to these five questions before rolling out:

  1. Does the assistant give correct answers in more than 85% of tested cases?
  2. Does it systematically cite the source used?
  3. Does it politely decline when the information isn’t in its base?
  4. Does the pilot team use it spontaneously (at least 5 sessions per user per week)?
  5. Have you measured a measurable time gain on the pilot case?

If a single answer is “no”, go back to pilot. Don’t roll out. Details in Making an AI assistant deployment succeed.

Typical schedule for a well-prepared deployment

W0–W2Inventory, pilot selection, GDPR scoping
W3–W6Ingestion, acceptance, adjustments
W7–W12Controlled rollout, second assistant

Three months between decision and a stable deployment, with measured results at each step. Faster, and you’re cutting corners on quality; slower, and prep was probably skipped.

Scope your pilot case

Twenty-minute demo on one of your real documents. You walk away with a sized pilot case and a six-week roadmap.

Book a demo