Proof of concept for AI: a reliable decision in six weeks

Which question should the proof of concept answer?

A proof of concept (PoC) is not a small rollout. It answers a decision question, for example: “Does an AI platform relieve our administrative office and case workers enough to make the introduction worthwhile?” or “Which of two offers solves our tasks better?”. Write the question down before you start, and decide who makes the decision at the end.

Typical mistakes are a PoC without a goal that runs on for months, and a trial with tasks nobody has in everyday work.

How do you structure the PoC?

Phase	Duration	Result
Prepare	1 to 2 weeks	Question, tasks, success criteria, test group, data release, informing the staff or works council
Set up and train	1 week	Access, groups, sample documents, short training
Test	3 to 4 weeks	Tasks in everyday work, measurement, weekly feedback
Evaluate	1 week	Report with measurements and recommendation

Six to eight weeks are enough for most questions. Plan a second, targeted round rather than extending the first trial.

Which tasks are suitable?

Choose three to five tasks that occur frequently, take time and whose result a specialist can check quickly:

• Answering questions on internal rules, service instructions or guidelines, with source
• Drafting letters, administrative decisions or reports from bullet points
• Summarising long documents or minutes
• Sorting emails and suggesting replies
• Rewriting texts in plain language (Einfache Sprache)

Unsuitable for a first trial are tasks that only work once a specialist application is connected, and tasks in which the AI decides about people. Assessing staff or citizens does not belong in a PoC; strict obligations apply to this under the EU AI Act.

How do you measure success?

Define the criteria in advance and measure the same way before and during the trial:

• Time: How long does the task take without and with AI? Samples are enough.
• Quality: How many answers are usable after review by the specialist, with minor changes, or not at all?
• Traceability: Does every statement lead to a source?
• Permissions: Does a person without access to a document get an answer from it? That should never happen.
• Acceptance: Would the testers want to keep working with it?

How measured time becomes a business case is shown in the guide Calculating the ROI of AI projects. To compare several providers, carry the criteria over into the trial setup; suggestions are in the guide Criteria catalogue for evaluating providers.

Which data do you test with?

As long as there is no data processing agreement, test with invented, public or non-personal documents: statutes, guidelines, service instructions, template letters. That is enough for most questions. Use real personal data only once the data processing agreement, the legal basis and, where applicable, a data protection impact assessment are settled and the data protection officer has agreed. This applies in particular to health and social data.

Who must be involved?

• Staff or works council: An AI platform is usually a technical facility that could monitor behaviour or performance (§ 87 Abs. 1 Nr. 6 BetrVG, the German Works Constitution Act, and the law on staff representation in the public sector). Inform them early and agree for the trial what is logged and what the data will not be used for. More in the guide model service agreement for AI.
• Data protection and information security: from the start, not only at the evaluation.
• Testers: train them briefly; this also counts towards AI literacy under Art. 4 of the EU AI Act.

What applies under procurement law?

A free trial from a provider is permitted as market research (§ 28 VgV, § 24 BVergG 2018, Art. 14 Abs. 3 BöB). However, it must not become a prior decision: anyone involved in preparing the tender may only bid if the advantage is offset (§ 7 VgV). So test several providers with the same tasks where possible, and document what you showed each of them.

A paid PoC is itself a contract. If its value stays below the threshold for direct awards in your country, it can be awarded quickly; it does not create any claim to a follow-up contract. If a solution has to be developed that does not exist on the market, an innovation partnership may be an option (§ 19 VgV). For an AI platform that can be bought ready-made, that is rarely the case.

Not legal advice: Your specialist departments review co-determination, data protection and procurement law in each individual case. Research as of: 11 October 2026.

How does this work with basebox?

For a first impression, register with your work email address and work with 1 million tokens, without a time limit, in your own account in our cloud. Apps, knowledge bases and settings are kept when you book the licence. On request, we show basebox in an appointment with your own sample documents.