Deploying generative AI where the content is sensitive

Transmission reports sit in a sensitive context. Quality, completeness and formulation weigh heavily, and generative AI must not replace professional judgment. When the ethics board ruled in the autumn of 2025 that the experiment was too vaguely defined and insufficiently aligned with privacy guidelines, the European AI Act and the rules around transmission, everything came onto the table at once: privacy, human end-responsibility, training, risk and governance.

Each of those questions had its own department, its own expert and its own pace. Because each question was handled separately, advice came in that stood loose from other advice, not integrated and not supported. Contradictions remained, and the experiment fell still in the meantime.

The ethics board's ruling was not read as a line through the initiative, but as the moment to tackle it differently. In late 2025 the experiment was reshaped into a project. The core of the work was the bridge between the operational knowledge of the working group and the shop floor and the broader requirements around quality, privacy, ethics, risk, training and governance.

The application was deliberately positioned as a supporting quality guide: it helps with preparation and formulation, but takes over no assessment or decision. The employee remains owner of the content and end-responsible for the report. A coherent framework was built around that. The goal was sharply demarcated, with explicit usage and privacy conditions. Fixed quality and validation moments followed, and a legal and ethical review. The application was formally included in the AI governance, the AI inventory and the prompt database.

Users received targeted communication and short, tailored training moments. Whoever had completed that training received access to the application via a licence key. At launch it was also explicitly communicated where the boundary lay. Whoever is competent enough to write a transmission report themselves, does so without AI support. And within this sensitive context, only the approved application is allowed: self-made variants or other AI tools are henceforth forbidden here. Follow-up runs afterwards on two fronts. In terms of content, the quality department monitors use within the quality management cycle: is the application being deployed as intended, and do the reports meet the quality requirements? Technically it's about monitoring, testing, version and access management, and maintenance whenever the model, the legislation or the way of working changes.

Equally determining was how that happened. The lead figures had ties to both the central services and the provincial operations and could therefore switch quickly. Between all those involved, connection was missing, and that was deliberately organised: directors, managers and a sounding-board group stayed continuously involved, so that cross-fertilisation emerged and every open question got a jointly supported answer. Experts' advice was tested against daily practice each time, after which it came back adjusted.

The shift from an innovation question to an organisation question. No longer was the central issue whether generative AI could support, but how this experiment could become, within the requirements of the AI Act, a safe, high-quality and manageable way of working that the whole organisation can use. From that moment on, technology, professional practice, privacy, quality, training, risk and governance were designed as one whole instead of as loose advice.

An approved way of working with which generative AI can be responsibly deployed in a sensitive client and process context, with human end-responsibility as a fixed point. Alongside that, a reusable approach for subsequent GenAI initiatives, and a smaller distance between the central services, the provinces and the shop floor.

Employees experience that reviewing pays off, and experts see what their advice is worth once it has been through practice.

Every organisation knows two streams. The surface current is the visible: content, structure, rules, planning. The undercurrent is what lives beneath: trust, relationships, meaning. A plan in the surface current only succeeds when the undercurrent is on board, and a strong undercurrent only delivers lasting results when the surface current is clear and supported.

This trajectory worked because both were developed at the same time. Three principles made that concrete. Make sure the lead figures are anchored in multiple parts of the organisation, centrally and locally, because they keep the lines short. Make mutual dependence explicit: the starting point was that nobody could do this alone, and precisely because of that, equality emerged instead of defence. And organise the translation as a fixed process, in which central advice is first tested against practice and vice versa, before anything is introduced.

That AI experiments become durable through technology alone is not true. They become durable when they get translated into roles, quality conditions, risk management, training, decision-making and ownership. That is the core of human-centred digitalisation: not just building a solution, but designing a practice in which people know what the technology does, what it doesn't do, and what they themselves remain responsible for.