A useful pilot should leave an organisation able to make a decision. Buying a tool and collecting enthusiastic comments won't necessarily do that. I would start with a task the team understands well enough to judge without the software vendor in the room.
For a small charity, drafting public event notices is one possible starting point. The source information can be approved in advance, and a staff member can check the result before publication. I use that example below to show how I would structure a trial.
Why the social sector needs its own criteria
Germany's welfare organisations are already discussing specific uses of AI. At its April 2026 event on care and algorithms, the BAGFW placed applications such as speech input alongside questions about professional care itself. BAGFW: Pflege zwischen Fürsorge und Algorithmen.
Staff need to judge the result against the work they are responsible for. The BAGFW's paper on AI in long-term care also calls for care expertise to inform development from the outset. BAGFW discussion paper.
For the event-notice example, the team can turn this into a concrete review: Is the location right? Are access requirements accurate? Has the draft invented free admission? A notice may read smoothly while giving somebody the wrong reason to attend.
Count the work that happens after generation
Before the trial, record how the team currently prepares a few comparable notices. Include the time needed to check details and correct the text. During the pilot, record the same stages for AI-assisted drafts.
Set a review date, perhaps four weeks away, and check that enough comparable notices will be written before then. If only one event is planned, the calendar needs changing.
If drafting becomes faster but checking becomes slower, report both. Don't count the generated text as finished work. Record errors found after release as well as those caught during review.
Choose someone who can reject a draft and someone who can pause the trial. The staff doing the work should help define unacceptable errors before the tool is introduced.
The proposed trial uses public information. A project involving service users' personal records would need its own assessment with the responsible data-protection and professional leads before those records were used. Success with an event notice would not justify extending the same setup to care documentation or decisions about support.
Try the arithmetic before writing the pilot report. Suppose a notice usually takes 25 minutes to draft, review and correct. With AI, prompting takes four minutes, checking takes twelve and rewriting takes eleven: 27 minutes in total. These are invented figures. Their purpose is to show why a faster first draft can still leave the team with more work.
For each actual notice, record the total time alongside the errors corrected and the release decision. A misspelt street name and an invented accessibility claim both take time to fix, but have different consequences for visitors.
An IHK Düsseldorf statement published on 24 August 2026 recommends beginning with a specific task and expanding only after practical experience. Dr Kerstin Vorberg makes that recommendation in a business-training context; a charity would still need to choose its task using its own professional criteria.
Give funders a decision they can inspect
At the end, retain examples of the task, the review criteria and the corrections required. Report total preparation time and the support staff needed to complete the process. A small pilot can be worth funding even when its conclusion is to stop or narrow the intended use.
The BAGFW's description of rückenwind³ places digitalisation and AI experimentation within workforce and organisational development. Eligibility still depends on the programme’s conditions. BAGFW: Wohlfahrt digital.
The strongest outcome might be quite limited: the tool helps with one kind of public notice, while another remains easier to write directly. That gives the organisation a basis for its next investment.
If the immediate need is staff practice with digital deception, Medienkundig's organisational pilot offers a place to discuss that training. Foundations interested in funding or testing digital learning with a community can explore Medienkundig's partnership options. Both pages are in German. A learning pilot would need its own evaluation; the time comparison above is designed for drafting work.
Sources checked on 10 September 2026. Cover: photograph by Headway / Unsplash. Illustrative photograph, not a Medienkundig project or training session.