For thirty years the sensible way to buy software was to decide carefully, once. Specification, sign-off, a long build, a launch. That process was not bureaucracy for its own sake — it was a rational response to the cost of being wrong. When a build ran to a quarter of a million euros and a year of calendar time, changing your mind halfway through was ruinous, so you bought certainty up front and paid for it in months.
That constraint has quietly stopped applying, and most companies have not updated the strategy that grew out of it.
What actually changed
The cost of producing a first working version of a piece of software has fallen sharply. Managed infrastructure removed the server work. Foundation models removed the need to train anything for a large class of problems. Modern tooling removed a great deal of the plumbing. What used to be a team and two quarters is now, for many well-scoped problems, a small team and a few weeks.
The important consequence is not that projects got cheaper. It is that the ratio changed. Building is no longer the expensive part of software — deciding what to build is. And when the answer costs more than the experiment, the rational move is to stop buying answers and start running experiments.
Why one pilot is the wrong number
If you accept that framing, a single carefully chosen pilot starts to look like the old strategy in new clothing. You are still trying to pick the winner up front; you have just made the bet smaller.
When experiments are genuinely cheap, the better approach is a small portfolio. Three or four narrow pilots across different processes will teach you more about where AI pays off in your specific business than one pilot on the process that seemed most promising in a meeting. The reason is unglamorous: intuitions about which process is the good candidate are frequently wrong, and they are wrong in ways nobody can predict from a slide. The data readiness is worse than expected here; the users are more enthusiastic than expected there.
You are not trying to be right about which one works. You are trying to find out cheaply.
The constraint is no longer budget
Here is the part that catches companies out. Once experiments are cheap, money stops being the limiting factor and something else takes over: the organisation’s capacity to make decisions about the results.
Every pilot produces a conclusion that somebody has to act on. Ship it, kill it, or extend it. Each of those requires a person with authority, a moment of attention, and the willingness to say no in public. That capacity is finite, it does not scale with your software budget, and it is the reason companies end up with eight promising pilots and nothing in production.
So the practical limit on how many experiments to run is not what you can afford to build. It is how many results you can actually decide on. For most mid-sized companies that number is small — three or four at a time, not ten.
What this looks like in practice
- Pick three processes, not one. Choose them to be different from each other — a document-heavy one, a customer-facing one, an operational one — because the point is coverage, not a single best guess.
- Give each a written pass mark and a stop date before anything is built. Cheap experiments without exit criteria are how you get the eight-pilot problem.
- Name one person per pilot who can kill it. Not a committee. If cancelling requires consensus, nothing will be cancelled.
- Schedule the decisions, not just the builds. Put the go/no-go conversations in the calendar at the start, and treat missing one as a failure of the process.
- Expect to kill most of them. A portfolio where everything succeeds was not ambitious enough to be informative.
The uncomfortable implication
If experimentation is cheap and decision capacity is the bottleneck, then the most valuable thing a company can improve is not its technology budget. It is its willingness to conclude things — to look at a working prototype that cost twenty thousand euros and say, clearly and on the record, that it is not worth taking further.
That is a cultural capability, not a technical one, and it is the thing that separates companies getting value from AI from companies with a lot of pilots.
Where to start
Our audit exists to answer the cheapest question first: is this process worth an experiment at all? It costs €2,500, takes one to two weeks, and it regularly concludes that a conventional script solves the problem for a fraction of the price of anything involving a model. That is a successful outcome. A proof of concept, if one is warranted, is €20,000 over four to six weeks, and production starts from €50,000.
The prices are fixed and published because an experiment with an open-ended cost is not really an experiment.