What Companies Should Clarify Before an AI Pilot Project
Eight questions to answer before starting an AI pilot — from data access and success criteria to team setup and the decision path after the pilot.
Why start with a pilot instead of going big immediately?
The temptation is understandable: AI promises efficiency gains, competitive advantages and lower costs. So why not introduce it across the whole company right away?
The answer is practical. Most failed AI projects do not fail because the technology is impossible. They fail because the goal is unclear, the data foundation is weak or the people who should use the system do not trust it. A limited pilot reduces those risks to a manageable size.
The 8-point checklist
1. Which concrete problem should be solved?
AI is not an end in itself. The starting point should always be a specific, measurable problem:
- Employees spend too much time searching for information
- Standard customer responses take too long
- Documents are still classified and routed manually
Rule of thumb: if you cannot describe the problem in one sentence, the scope is probably too broad.
2. Which data is available?
No data, no AI. Clarify upfront:
- Where is the relevant data stored? SharePoint, Confluence, file system, email?
- Which formats are involved? PDF, Word, structured databases?
- How current is the data?
- Are there access restrictions?
Important: a pilot often only needs a compact, well-maintained data base. You do not need a full data warehouse to start.
3. What are the success criteria?
Define how success will be measured before the project starts:
- Time saved per request, for example from 15 minutes to 3 minutes
- Answer quality, reviewed by the specialist department
- User acceptance after four weeks
- Error rate, including wrong or incomplete answers
4. Who is the pilot user group?
Not every department is equally suitable for the first pilot. Strong pilot users:
- Do a lot of knowledge-intensive work
- Are open to new tools
- Can give constructive feedback
- Are representative of the later user base
5. Which budget and timeline are realistic?
A typical AI pilot at OzyCore:
- Duration: 6–10 weeks
- Budget: €10,000–25,000, depending on complexity
- Scope: one use case, one data source, one pilot group
6. Who decides the next step?
Clarify before the start:
- Who evaluates the results?
- Which criteria trigger the decision for a rollout?
- Who owns the budget for phase two?
7. Which data protection requirements apply?
Especially in Germany, you should clarify:
- Will personal data be processed?
- Is a data protection impact assessment required?
- Which GDPR requirements apply to the processing?
- Does the works council need to be involved?
Read more in our post on AI and data protection in Germany.
8. What happens after the pilot?
There are three possible outcomes — all of them are useful:
- Success: rollout to more departments
- Partial success: adapt the approach and run a second pilot
- No success: valuable learning at limited investment
Common mistake: too many use cases at once
The biggest mistake in AI pilot projects is usually organizational, not technical. Too many stakeholders want to test their use cases at the same time. The result is diluted learning and unclear responsibility.
Our recommendation: one pilot, one use case, one pilot group. Sequential beats parallel.
Conclusion
A well-prepared AI pilot takes a few weeks, costs a fraction of a full rollout and gives leadership reliable evidence for the next decision.
The eight questions in this checklist help set up the pilot properly and avoid typical pitfalls.
Ready for a pilot project? In a non-binding first conversation, we can discuss your use case, clarify the prerequisites and outline a realistic pilot plan.