AI within public-sector boundaries, for plain language and faster handling
A province communicates with residents, businesses and partners about topics that genuinely matter: water, housing, nature, mobility, the economy. Provincie Zeeland wanted to apply AI to make that communication more accessible and to speed up internal processes, without giving up on diligence, GDPR and public accountability. No consulting circus, but a way of working you can defend inside the organisation.
AI in an environment where transparency, accessibility and GDPR are not afterthoughts
Public-sector language is precise, which is a good thing, but for a resident a permit letter is often a wall of C2-level sentences. At the same time provincial documents must meet WCAG, the Dutch Open Government Act and the retention rules that come with a public body. Every AI application therefore touches three layers at once: the resident, the employee and the legal officer. That demands an approach that not only works, but is also defensible to the corporate controller and the data protection officer.
On top of that sit the practical preconditions: data must not leave the EU, vendors must sign a processor agreement, models must not train on internal material, and every choice has to fit a board decision or procurement framework. The question was not whether AI is possible, but how to introduce AI in a way that survives the procurement, privacy and governance cycle without months of delay.
Three steps that move through governance, procurement and the daily work
Analysis and framework setting
We mapped the work processes where AI can add value directly and tested them against GDPR, WCAG and existing procurement frameworks. For every use case we wrote a lightweight DPIA and a risk assessment, so the privacy officer and corporate controller can review on a single page.
Strategy and vendor setup
We selected models and hosting parties that run inside the EU and sign a processor agreement, with a no-training policy. Per use case we documented which data may and may not enter a prompt, how long processing is retained and how an employee can flag an error.
Implementation and adoption
We rolled out the first use cases within a defined department, with a feedback loop to the editorial team and the privacy officer. Employees received working sessions instead of a manual, and operations sit with a permanent team inside the province so knowledge stays after our engagement.
Concrete applications, with the safeguards a public body needs
Each application started small, with a measurable goal and an exit path. Components are loosely coupled so a use case can be paused or replaced without stopping the rest. The list below is illustrative and is adjusted per engagement based on priority and legal review.
- 01Writing assistant that rewrites official texts to plain language (B1), with an audit trail per change
- 02Accessibility scan on PDFs and web pages against WCAG 2.2, with concrete fix suggestions
- 03Routing of citizen questions to the right team, with reviewed draft answers for the case handler
- 04Searchable archive of decisions and policy documents, with source references in every answer
- 05Meeting minutes and action lists, drafted from recordings with speaker recognition
- 06Working method with DPIA, processor agreement and EU-only setup, captured in an operational runbook
What the first use cases deliver in numbers
shorter lead time on standard letters to residents
documents checked for WCAG accessibility
use cases live with DPIA and processor agreement
* Numbers follow after the first measurement period and are verified with the client
“We have introduced AI in a way that passed our privacy officer and procurement. It saves time on work nobody misses, and we can explain to residents and the board how it works.”
next case