Key takeaways
- AI pilots rarely fail because of the technology, but because of a missing business case, ownership or employee involvement.
- Good starter use cases are frequent, clearly delimited, measurable and low in risk, such as text and document work.
- Since 2 February 2025, the EU AI Act’s prohibitions and the obligation to take AI literacy measures have applied.
- Every pilot needs a clear objective, a measured baseline and a fixed end before any decision on scaling is made.
Contents
Many companies run AI pilots that never make it into everyday work. The reason is rarely the technology. What is usually missing is a clear business case, a person who owns it, or the involvement of the people who are meant to work with the result. Starting from the process rather than the tool leads more quickly to results that last.
01Start from the process, not the tool
A common mistake: a tool is introduced first, and then people look for something to use it for. The reverse route works better. It begins with questions about your own business:
- Where does a lot of manual, repetitive work arise today?
- Where are customers or colleagues waiting for answers, documents or approvals?
- What does this process cost today – in time, errors and rework?
- How would we notice that things have improved?
02Suitable use cases to start with
Quotes, minutes, summaries, draft contracts or reports – wherever texts are produced to a similar pattern.
Sorting enquiries, drafting replies, answering frequent questions – with a person who checks and decides.
Making policies, manuals and project knowledge easier to find, instead of asking colleagues the same questions again and again.
Capturing documents, matching data, flagging discrepancies – tasks with clear rules and high volumes.
Good starter use cases are frequent, clearly delimited, measurable and carry a manageable risk. Where decisions about people are involved – in recruitment, for example – particular care is needed.
03Data and IT: the prerequisites
- Know where the relevant data is held and in what quality.
- Clarify access rights: who may put which data into which system?
- Plan the connection to existing systems such as ERP, CRM or document management instead of creating isolated solutions.
- Check providers for security, data processing and contract terms.
- Budget for ongoing operating costs, not just the introduction.
04Bringing people along
AI changes work – and that raises questions: will my job become redundant? Do I have to learn something new? Can I trust the results? These questions do not go away if they are ignored. What helps:
- Involve the people who run the process today – they know the exceptions.
- Train in a practical way, with examples from people’s own working day.
- Say openly what will change and what will not.
- Make early successes visible and share experiences within the team.
Practical tip
Involve the people who run the process today from the very start – they know the exceptions that are not in any process description.
05Governance, data protection and the EU AI Act
One person in senior management owns the topic and decides on priorities, budget and approvals.
Short, understandable rules: which tools are approved, which data may go into them and when a person checks the result.
As soon as personal data is processed, the data protection officer and, where there is one, the works council should be involved early on.
The EU’s AI regulation applies in stages. Since 2 February 2025, certain AI practices have been prohibited, such as social scoring or – with narrow exceptions – emotion recognition in the workplace. Since the same date, providers and deployers of AI systems, including companies that merely use AI, have had to take measures on the AI literacy of their staff (Art. 4). An amendment to the regulation in summer 2026 (the “Digital Omnibus”) clarified that no specific level of literacy is prescribed – the obligation to take appropriate measures remains.
Legal assessment & implementation
For the legal assessment of your AI applications under the EU AI Act and data protection law – such as risk classification and AI literacy under Art. 4 – talk to the law firm navyra.legal. navyra supports the implementation in your processes and systems.
06Step by step: from use case to scaling
- Prioritise use casesCollect use cases and prioritise them by benefit, effort and risk.
- Measure the baselineMeasure the baseline: how long does the process take today, and how many errors occur?
- Start a pilotStart a pilot with a clear objective, a fixed team and a fixed end date.
- Evaluate & decideEvaluate the results and decide deliberately: adopt, adjust or stop.
- Move into operationMove the use case into regular operation – with owners, training and support.
- Scale upOnly then roll it out to other areas, using the experience gained.
07Measuring the benefit
Without measurement, every pilot remains a matter of opinion. Which key figures fit depends on the process. Typical ones are:
- Time per transaction and lead time.
- Error rate and rework.
- Satisfaction of customers and employees.
- Actual use in everyday work.
This benefit has to be set against licence, operating and training costs. Only an honest comparison shows which use case is worth it – and which is not.
Jens Schneiders
Interim CEO for mid-sized companies, scale-ups and corporate ventures. Transformation. Restructuring. Stabilization. Handover.