A clear methodology for AI pilot projects often determines success more than the choice of model, because it defines when a pilot ends and how its success can be measured. According to the Bitkom report, analyzed by bi2run By August 2026, approximately 41 percent of companies with 20 or more employees will be actively using AI; a year earlier, the figure was significantly lower. Nevertheless, many pilot projects never make it to production. This article describes a 90-day methodology that takes you from the initial idea to the decision to roll out or discontinue the project.
A 90-day AI pilot project methodology divides the work into four phases: defining the use case and success metrics, establishing data access and building a prototype, testing with real users, and then Decide whether to proceed with the rollout or cancel it. According to a March 2026 analysis by collectivebrain, without this structure, only about 19 percent of AI initiatives ever reach production.
Why Most Pilots Fail to Deliver Results
According to a widely cited analysis linked to MIT, approximately 95 percent of companies using generative AI see no measurable impact on their bottom line. This figure comes from a single study, not from a broad, independently replicated series of studies, and should therefore be interpreted with caution. However, it is broadly consistent with the findings of collectivebrain from March 2026, according to which only about 19 percent of AI initiatives ever reach production.
According to the same analysis by collectivebrain, pilot programs with clearly defined success metrics have significantly better chances of achieving a positive ROI than those without established metrics. The analysis recommends a pilot duration of six to eight weeks for the test phase alone—short enough to yield results, yet long enough for a genuine user test. According to the same analysis, one-third of companies using AI consider the technology to be more expensive than originally expected. The gap between 41 percent adoption and 19 percent production use typically arises during the transition from prototype to actual testing—precisely the phase that a fixed methodology is best suited to address. The article lists which government funding programs in Germany and Austria can help offset these costs. AI Funding Programs in 2026 in Germany and Austria.
90-Day AI Pilot Project Methodology
The following methodology expands collectivebrain’s basic user test into a comprehensive framework that spans from the initial idea to the final decision. Four phases with clear deliverables replace the open-ended „Let’s just give it a try“ approach that most failed pilot projects have in common.
| Days | Phase | Result |
|---|---|---|
| 1 through 15 | Use Case and Success Metrics | One use case, one measurable metric, one person in charge |
| 16 to 45 | Data Access and Prototype | A working prototype based on real, albeit limited, data |
| 46 to 75 | Test with real users | Usage data and feedback from real-world work situations |
| 76 to 90 | Decision | Rollout, revision, or cancellation, with justification |
The first 15 days often require more discipline than the weeks that follow. If you don’t define a specific use case with a single metric at this stage, the project almost always loses focus later on, because each department brings its own expectations to the table. The phase from Day 16 to 45 follows the same logic as a data quality assessment, but is applied directly to the prototype rather than to an isolated sample.
Between Days 46 and 75, what counts is actual use—not internal demo sessions in front of coworkers. A prototype used only by the project team itself does not provide reliable insight into whether the application will work in everyday use. The final 15 days are dedicated to making a clear decision: rollout, targeted refinements with a new timeline, or cancellation.
If the company decides to proceed with the rollout, operations will not automatically begin across the entire organization on Day 90. It makes more sense to take a second, broader step using the same principles, a new user group, the same metric, and a new timeframe of six to twelve weeks. This way, the expansion remains just as easy to evaluate as the original pilot.
Some AI initiatives reach full-scale operation, while all the others remain stuck in the pilot phase.
collectivebrain, March 2026
A pilot project without predefined termination criteria will go on forever. If no one decides before the start what outcome would justify terminating the project, every mediocre interim result will be sold as progress, and 90 days can easily turn into 300.
Before the Pilot: The Data Situation
A prototype in Days 16 through 45 stands or falls on the data source that feeds it. Our contribution to Data Quality in AI Projects describes how to quickly assess a data source before starting a project, without investing weeks in a data project. Anyone who wants to apply this methodology to a specific use case can do so in a Initial Consultation on AI Consulting An initial assessment.
Define the use case and metric together
The first 15 days of the methodology often determine how the rest of the process unfolds. A discussion before getting started helps define the use case and success metrics right from the start.
Frequently Asked Questions
How much does a pilot cost?
That depends heavily on the use case and data access; it’s impossible to provide a reliable estimate without knowing the details of the project. As a rough guide, a 90-day pilot typically requires a small, dedicated team for the entire duration—not just occasional support on individual days. Estimating the personnel costs for this team over ten to twelve weeks is a more realistic approach than any flat-rate daily rate calculation. In many cases, there are also costs for model access or infrastructure, which can only be roughly estimated before the first prototype is ready and therefore require a buffer in the calculation.
Who should be involved?
At least three roles: a business owner from the department using the use case, a technical lead responsible for data access and the prototype, and a person who ultimately makes the decision to cancel or roll out the project and takes ownership of that decision. If any of these three roles is missing, it is usually the final decision that gets delayed, because by day 90, no one feels responsible for making it. In practice, a pilot project without a designated decision-maker often ends in a postponement rather than a clear “yes” or “no.”.
When should you quit?
It’s best to base this on the metric established in the first 15 days, not on a gut feeling in week 12. If the prototype clearly fails to reach the defined threshold by day 75, termination is the planned consequence according to our own methodology. The team did nothing wrong; the metric was known from the very beginning.
Why torck
torck builds prototypes and production systems with its own teams in Maxhütte-Haidhof, Vienna, and Rabat, and supports pilot projects through to the decision on whether to roll them out or terminate them. We remain involved in the project even after the initial demo, until a sound decision has truly been made. From our own projects, we know which of the four phases tends to run over the most. A 90-day timeframe only works if the same people are on the project the entire time, which is why our teams work in a shared time zone with German-speaking project management. Anyone who wants to set up a methodology for their own use case is best advised to discuss this in a Initial Consultation on AI Consulting.
German 
English 
French