How a Belgian Municipality Automated Citizen Requests with AI
From 2-week response times to same-day answers — without hiring more staff.
The Problem
A mid-sized Belgian municipality (name withheld for confidentiality) was drowning in citizen requests. Their team of four administrative staff handled everything from waste collection questions to parking permit applications to complaints about noise and potholes.
The requests came in via email, web form, phone, and walk-ins. Each one needed to be logged, categorised, assigned to the right department, and tracked until resolution. Average response time: 10–14 days. Citizen satisfaction: dropping.
The mayor wanted to hire two more people. The budget officer said no.
The Fly AI Solution
We built a three-part AI system over four weeks:
Part 1: Intake Agent
An AI agent that reads incoming emails and web form submissions, extracts key info (citizen name, address, issue type, urgency), and auto-categorises the request. It runs 24/7 and processes requests within 30 seconds of arrival.
Tech: Claude 3.5 Sonnet via Anthropic's EU API, integrated with the municipality's email server and existing ticketing system.
Part 2: FAQ Responder
For the ~60% of requests that are routine ("When is bulk waste collection?", "How do I apply for a parking permit?", "Where do I report a pothole?"), the agent drafts a response in FR or NL, pulls the relevant link or PDF from the municipality's website, and parks it in the staff member's draft folder for one-click approval and send.
Tech: Custom knowledge base built from the municipality's website, municipal regulations, and past answered emails. Fully bilingual (FR/NL).
Part 3: Routing & Escalation
For complex or sensitive requests (complaints, legal questions, urgent infrastructure issues), the agent flags them as "requires human" and routes them directly to the right person with a summary: "Complaint about noise from construction site at Rue X — mentions ongoing sleep disruption, requests inspection."
Tech: Rule-based routing combined with AI classification. High-urgency keywords trigger immediate escalation.
Implementation Timeline
| Week | What Happened |
|---|---|
| Week 1 | Discovery call + scoping workshop with admin team. Reviewed 200 past requests to identify patterns and edge cases. |
| Week 2–3 | Built and tested the intake agent and FAQ responder on a sandbox email account. Trained on historical data. |
| Week 4 | Human-in-the-loop launch: agent processes requests, staff reviews and approves drafts before sending. Fine-tuned based on feedback. |
| Week 5+ | Gradual autonomy increase. After two weeks of supervised mode, routine FAQs were approved for auto-send. Complex cases still flagged for human review. |
Results After 3 Months
- Response time dropped from 10–14 days to same-day for 60% of requests
- Staff time saved: ~18 hours per week across the team (equivalent to nearly half an FTE)
- Citizen satisfaction: jumped from 62% to 89% (measured via post-resolution survey)
- Accuracy: 94% of AI-drafted responses were sent without edits; 6% required minor tweaks
- Multilingual support: seamless FR/NL handling without duplication of work
What the Staff Said
"At first we were sceptical — we thought it would just create more work checking the AI's answers. But within two weeks we trusted it. Now we can focus on the requests that actually need our judgment, instead of answering the same parking permit question for the 50th time."
GDPR & Compliance
Because this is a public body processing citizen data, GDPR compliance was non-negotiable. Here's how we handled it:
- EU hosting: All data processed via Anthropic's EU instance (hosted in Frankfurt)
- Data minimisation: Agent only receives the email body and sender info, not full headers or metadata
- Retention: AI logs deleted after 30 days; citizen requests follow existing municipal retention policy
- DPA signed: Standard Data Processing Agreement in place with Anthropic
- Privacy policy updated: Citizens informed that initial triage may be automated
- Right to human review: Any citizen can request their case be handled by a human instead
What It Cost
| Scoping workshop | €450 |
| Agent build (intake + FAQ + routing) | €3,200 |
| Training & handover | €550 |
| Total project cost | €4,200 |
| Monthly running costs (API + hosting) | €180/month |
ROI: The time saved is worth ~€1,800/month in staff cost (18 hours × €25/hour blended rate). The system paid for itself in 2.3 months.
Lessons Learned
✅ What worked
- Starting with human-in-the-loop supervised mode built trust fast
- Focusing on the high-volume, low-complexity requests first (the "FAQ zone") delivered quick wins
- Involving the admin team from day one meant they became advocates, not resisters
⚠️ What we'd do differently
- We underestimated how many "edge case" requests there would be in month one — the long tail is real
- Initial training took longer than planned because past emails weren't well-categorised
- We should have built a dashboard for the team to see stats (requests handled, accuracy, time saved) — added it in month two
Could This Work for Your Organisation?
This approach works well when you have:
- High volume of similar requests (at least 50+ per week)
- A knowledge base or set of standard answers to work from
- Staff willing to review AI outputs initially (supervised mode)
- A commitment to GDPR compliance and transparency
It's especially powerful for public sector organisations, customer support teams, and professional services firms (legal, accounting, HR) where much of the incoming work follows patterns.
Want to explore AI automation for your team?
Book a free 30-minute discovery call. We'll look at your incoming request patterns and tell you honestly whether AI could help — and how much time and money you'd save.
Book a discovery call
