An AI assistant of your own: from pilot to daily use in a public body
Short answer
A pilot rarely fails on technology. It stalls on unclear rules, no owner and no measurable goal. Pick two or three tasks that come up daily, give each department only the sources it needs, agree in advance what you will measure, and make sure staff know what they may and may not enter. Open the assistant up widely only once those tasks are demonstrably faster.
Why pilots stall
- No clear task. "Give it a try" produces a few loose experiments, not an improvement in a process.
- No owner. Without a department that has to deliver something with it, nothing happens after the demo.
- Unclear rules. Staff who do not know what they may enter either avoid it, or use it for everything.
- Empty sources. An assistant without access to by-laws, work instructions and earlier decisions gives generic answers.
- Nothing measured. Without a baseline there is no story for the executive, so no follow-up.
Where to start
Pick tasks with a lot of text, high volume and recoverable mistakes. For public bodies, these usually work:
- Questions from residents: classify and prepare a draft reply from approved information.
- Summarising documents: applications, objections, advice and annexes, with source references.
- Searching policy: by-laws, policy rules and earlier decisions, walled off per department.
- Making letters easier to understand, with the specialist as final editor.
Conditions that make the difference
| Condition | What it means in practice |
|---|---|
| Roles per department | Social services, permits and communications each see their own sources. See Access control for AI |
| Clear policy | One page on what is and is not allowed, instead of a list of prohibitions nobody reads |
| Training | Half an hour per team on what the model can do and where it goes wrong. See AI literacy |
| A reporting point | One place where staff can report a wrong answer, with follow-up |
| A baseline | Lead time and hours per task, measured before the start |
From pilot to daily use
Scaling up is not about more users but about more confidence. Look at how often staff change the proposal, how often they fall back on the old way of working, and where answers go wrong. Expand only when those figures move the right way, and add every new application to your AI register.
With werqly the assistant runs on a node of your own organisation. Residents’ data stays in your own infrastructure, your IT department sets roles and sources in werqly AI Control, and the audit log provides the figures for your evaluation. See AI for local government and Generative AI in local government.
Frequently asked questions
Questions on this topic
Why do AI pilots in the public sector stall?
Usually not because of the technology, but because a task, an owner or a measurement is missing. A pilot without a process for the outcome to land in, and without a baseline to show the effect, ends after the demo.
Which application is the best place to start?
One that comes up often, involves a lot of text and where a mistake is noticed before it does harm: answering questions from residents, summarising documents or searching policy. Decisions about people are a poor starting point.
How do we make sure staff use it?
Give them access to their own department’s sources, one clear page on what is and is not allowed, a short training and a place to report wrong answers. Use grows when the work is noticeably faster.
What do we measure to decide on scaling up?
Lead time and hours per task, how often staff change the proposal, and how often they fall back on the old way of working. Those three figures say more than the number of users.
Sources
- Government of the Netherlands — Government broadens position on generative AI (April 2025, Dutch)
- Algorithm Register of the Dutch government
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.