How long does a private AI rollout actually take?
A transparent breakdown of timelines: 72-hour pilot proof-of-concept, 4-week security audit and data integration, 8-week production cutover and operational SLA handover.
Why many AI projects get stuck in the pilot phase
To many vendors, a fixed price AI project sounds like a risk. For many companies, it is the only way to get started at all.
The classic path looks like this: workshop, concept, a quote based on time and materials, extension. After a few months there is a pilot that never leaves the pilot phase, and a budget nobody knew about beforehand. The business unit has invested time, IT has set up access, and management is waiting for a basis for a decision. In the end, the answer to the question that matters is missing: Is it worth scaling up?
Figures from the market confirm this pattern. A 2025 MIT study concludes that 95% of generative AI pilots show no measurable return. The study's methodology is limited, yet its direction matches what we hear in conversations with companies. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. Analyses by Deloitte as well as Capgemini and Microsoft show that only 6% of all AI initiatives pay for themselves in less than a year.
These figures do not argue against AI. They argue against projects without a clear scope, without a measurable goal and without a plan for operations. Projects like these end up in the pilot graveyard.
Fixed price as a principle: three conditions
A fixed price is neither a discount nor a teaser offer. It is a method for resolving uncertainty before the start rather than during the project. For it to work, three conditions must be met.
The scope is narrow. One use case, one process, one clearly defined data set. If you want to connect several departments and systems at once in the first step, you do not need a fixed price offer, you need to set priorities.
The goal is measurable. Before the start, we agree on a baseline together: How long does a task take today, what is the error rate, how much capital is tied up? Without this starting value, there is no way to tell after the pilot whether it improved anything.
Both sides know in advance what is not included. This list is at least as important as the scope of services. It prevents a pilot from gradually turning into a project without end.
A fixed price does not mean that nothing may change during the pilot. If it turns out that a different scope looks more promising, we discuss this openly and agree on the next step anew, instead of quietly expanding the current scope.
This clarity is the real value of a fixed price, more than the price itself. As the client, you see the risk shift: you no longer carry the uncertainty about the effort, we do. That forces both sides to be precise before the start. How we bill the pilot, implementation and operations is described on our Pricing page.
What a pilot delivers in 72 hours and what it does not
| What a pilot delivers | What a pilot does not deliver |
|---|---|
| Demonstrate a concrete use case with your real data | Replace production operations |
| Show whether data quality and interfaces hold up | Cover all special cases and exceptions |
| Provide a first measurement against the baseline defined beforehand | Provide a reliable long term value over months |
| Make open questions for data protection, IT security and the works council concrete | Anticipate approval by data protection and the works council |
| Create a basis for the decision on scaling up | Provide a finished integration into all target systems |
A first internal pilot in 72 hours is deliberately small. Its purpose is to make a decision possible, not to produce a presentation. The following overview shows where its strengths and limits lie.
"Internal" means the pilot runs with a small group of users from the business unit, not with customers and not in daily business. This lowers the risk and is what makes fast implementation possible in the first place. For the pilot to be ready in 72 hours, a data extract, a contact person in the business unit and the technical environment must be settled in advance. If the pilot runs in an environment you control, your data stays within your company even in this early phase.
From pilot to production
A pilot answers the question of whether a use case holds up. The path into production answers the question of whether it works in everyday use. We plan 4 to 10 weeks for this, depending on data, interfaces and approvals.
This phase covers what a pilot deliberately leaves out:
At the end of this phase comes an acceptance check against the baseline, this time with real volumes and over a longer period than in the pilot. Only this measurement shows whether the effect seen in the pilot holds up in everyday use.
The range of 4 to 10 weeks depends less on the technology than on the organization. If the data is clean and approval processes are well established, things move faster. If there is no solid data foundation, it takes longer. According to the analyses cited, 80% of initiatives require upfront investment in data architecture. A pilot reveals this need early, before larger budgets are committed.
Once the system goes live, operations begin. With us, monitoring, retraining and updates are part of operations, not a new project. For availability, we set 99.9% as the operational target. Here too, the scope is agreed in advance so you know what to expect.
Who a fixed price start is right for
A fixed price for getting started is a particularly good fit if
A fixed price start is less suitable if the goal is still open or if several use cases are to be evaluated at the same time. In that case, choosing the use case comes first.
Expectations about the return should also stay realistic. Individual use cases typically need 3 to 12 months to pay back the effort, large platforms 2 to 4 years. A pilot does not shorten this period. It ensures that you know early whether the path is worth taking.
For data protection and IT security, one rule applies from the outset: whatever we promise must be verifiable. We describe the evidence we provide under Security.
Frequently asked questions
Sources
- MIT (2025): Study on generative AI in companies (95% of pilots without measurable return). The methodology is limited; the figure should be read as a trend.
- Gartner (2025): Forecast on agentic AI (more than 40% of projects canceled by the end of 2027).
- Deloitte (2025) as well as Capgemini and Microsoft (2025), cited in nexpatch.ai/de/blog/orpheon-vs-plattformen (6% of AI initiatives pay for themselves in less than a year, individual use cases in 3 to 12 months, large platforms in 2 to 4 years, 80% of initiatives require upfront investment in data architecture).