Initial context
A municipality of 35,000 residents, 410 staff (tenured, contract, temporary), HR department of six. Typical document volume: thirty years of internal memos, two hundred job descriptions, a collective agreement, internal rules, hundreds of DGCL and DGAFP circulars (national HR oversight bodies) archived without an index.
Pre-deployment assessment: 40% of the HR team’s time spent searching. Looking for the latest version of a procedure, a precedent handled the same way, an applicable circular. The rest — payroll, recruitment, training — was done under pressure.
Chosen scope: internal HR, not the general public
Key decision: the first assistant does not serve staff or residents. It serves only the HR team itself. Why?
- Simpler GDPR framing (an assistant that sees HR data for HR staff, no transfer outside the team).
- Simpler to acceptance-test (the six users flag errors in real time).
- Simpler to show value (gains are measurable across six people, not diluted across 400).
The second assistant — for staff — will be deployed in a second phase, once the first is in routine use.
Document ingestion: three sources, in order of usefulness
- The internal rules and their annexes (12 years of successive revisions).
- Internal memos and circulars since 2018.
- Job descriptions and current pay scales.
Total initial volume ingested: roughly 2,800 documents. Three quarters were loaded in two afternoons. The remaining quarter — old scans, corrupted Word files — was handled case by case the following week.
Daily use, six months in
Three use cases account for 80% of questions asked of the assistant:
- “What’s the procedure for long-term sick leave for an employee on a fixed-term contract of more than 3 years?"
- "What internship stipend did we apply last time for a master’s in HR?"
- "We need to redo the remote work memo, where is the latest approved version?”
The assistant answers in five to ten seconds, with exact citation of the source document. When it doesn’t find an answer, it says so.
Numbers after six months
The recovered hours are largely reinvested in recruitment — which was structurally behind schedule.
Three lessons learned
Start small, finish big. The HR team explicitly refused to open the assistant to staff before six months. That’s what made the pilot succeed.
Watertight scope is not optional. No document leaves the database, no document is used to train a model. See our security & GDPR page.
Adoption is won on UX, not on the pitch. If the assistant isn’t faster than opening the file share, no one uses it. Rule of thumb: three clicks maximum between question and answer.
Step-by-step in Practical guide: deploying AI in your local government. See also our local government page.
Twenty minutes on a sample of your HR documentation. We look at actual quality, identify the pilot scope, and frame compliance.
Book a demo→