What is an AI pilot?

An AI pilot is a small, time-boxed build that runs on your real data, with your real users, to answer a question you cannot answer on paper: does this work here, and is it worth doing at scale?

Pilots exist because of an economic shift, not a fashion. Building software has become dramatically less expensive and faster. That did not just reduce the cost of projects, it changed which strategy is rational. When a build cost €250,000 and took a year, you had to be certain before you started, and certainty was expensive to buy. Now that the same capability costs a fraction of that, the less expensive move is to stop trying to be certain and start finding out.

Why pilots became the right unit of work

For thirty years the sensible approach to software was to specify heavily up front, because changing your mind mid-build was ruinous. That logic held while building was the expensive part. It no longer is. Modern tooling, foundation models and managed infrastructure have collapsed the cost of a first working version, and the expensive part is now knowing what to build. When the answer costs more than the experiment, run the experiment. That is the whole argument for piloting, and it is why the practice arrived with AI rather than before it.

What a pilot is actually buying

Information, not software. A pilot that ends with a working tool but no clearer decision has failed; a pilot that ends with no tool but a firm, evidenced "no" has succeeded and saved the production budget. Three things are usually worth more than the prototype: whether your data supports the use case (it frequently does not, and this is the most common killer), whether the people whose work changes will actually use it, and what the real error rate is on your inputs rather than a benchmark.

How to run one: five decisions before any code

1. One process, named. Not "AI for customer service" but "the first-response draft for warranty emails". 2. A number someone already owns. If nobody currently reports the metric, nobody will notice it improve. 3. A pass mark written down first. Decide what result would justify production, before you see any results. 4. A real user, not a stakeholder. The person who does the work daily, not the person who approved the budget. 5. A stop date. A pilot without an end date becomes a permanent side system that nobody maintains.

The new failure mode is pilot purgatory

Low-cost experiments create a failure the expensive era did not have. The old risk was betting €250,000 on the wrong thing. The new risk is running eight pilots and shipping none of them. Each one looked promising, none had a written pass mark, so none could be declared finished or dead. The discipline that matters now is not the entry criteria, it is the exit criteria. A pilot needs a defined way to fail as much as it needs a way to succeed, and someone with the authority to call it either way on the stop date.

What the failure statistics actually say

You will have seen the number: 95% of enterprise GenAI pilots deliver no measurable P&L impact. It comes from The GenAI Divide: State of AI in Business 2025, published by MIT's NANDA initiative, which drew on 52 executive interviews, a survey of 153 leaders and a review of roughly 300 public deployments. Gartner published a second figure the same year: more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.

Both numbers are worth reading properly rather than repeating. The MIT report defined success narrowly and deliberately: deployment beyond the pilot phase, with measurable KPIs, and ROI assessed six months after the pilot ended. On that definition, efficiency gains, lower churn and faster pipeline all count as nothing. The report also describes its own interview findings as directionally accurate rather than company-reported. So 95% is not 95% of pilots collapsing; it is 95% failing a strict test that most pilots were never set up to sit.

Which is the useful part. That strict test is exactly the bar to design for. If a pilot cannot name the KPI it moves, the person who owns that KPI, and the date six months out when someone checks, it will land in the 95% by construction, whatever the model does. Both findings point at the same cause, and it is not model quality: it is integration into a real workflow, and deciding in advance what would count as proof.

What an AI pilot costs and how long it takes

At Crux Digits: a €2,500 audit first, to establish whether the process is worth piloting at all: roughly one to two weeks, and it regularly ends with "a €4,000 script does this, do not build AI". Then a €20,000 Production-ready MVP running on your data for 4–6 weeks, and production from €50,000 if the pass mark is met. Fixed prices, because a pilot whose cost is open-ended is not time-boxed in any meaningful sense.