How to succeed in your AI assistant project

1. Pick the pilot (week 1)

The ideal pilot: a case where value is clear, scope is limited, and users are identified. Three criteria:

• A function where documentation is dense (HR, quality, legal, support).
• Document volume between 500 and 5,000 files.
• Five to twenty pilot users — no more, no less.

The classic mistake: starting with the most visible case (customer service, marketing) because it’s exciting. Visibility increases risk, and risk kills pilots.

2. Frame compliance (week 2)

Before the first ingestion, validate with your DPO:

• The DPA provided by the vendor (GDPR article 28).
• A DPIA if needed (see GDPR and AI assistants).
• Entry in the records of processing activities.

Two weeks, not three months. If your legal cycle is longer, you have a governance problem — not an AI problem.

3. Prepare documentation (weeks 3-4)

Three mechanical passes:

1. Triage: active / archive / obsolete.
2. Deduplication by hash.
3. OCR cleanup on scans.

Details in Good documentation makes a good assistant. Do not rewrite your documentation — that’s the trap that turns a three-month project into a two-year project.

4. Ingest and validate (week 5)

Before ingestion, write 20 to 30 test questions. Easy, realistic, hard. Check after ingestion: 85 percent correct answers minimum, 100 percent with source citation.

Below 85 percent, the problem is in the documentation, not the AI. Go back to step 3.

5. Onboard users (week 6)

Two one-hour sessions per user, run by a business champion (not the IT project lead). Three topics:

• How to frame a good question (see Prompting like a pro).
• How to read an answer with citations.
• How to flag a bad answer.

6. Measure in routine (months 2-3)

Six KPIs to track from go-live:

Details in Measuring AI assistant effectiveness.

7. Expand, or not (months 4-6)

At the end of the pilot, two honest questions:

1. Is the ROI there? If no, adjust the scope rather than restart a new pilot.
2. Do users come back on their own? If no, that’s a usage problem, not a technology problem.

If yes to both, expand to a second use case. Not three in parallel. One that succeeds and creates the precedent.

To go further, see Calculating and maximizing ROI and Integrating an AI assistant into your IT infrastructure.