AI pilot to production: why it stalls and what to do

Your AI pilot isn’t reaching production because it was designed for the demo, not for operations. The model is rarely the problem. What’s missing is prepared data, a success criterion agreed before you start, integration with your real systems and someone accountable when the AI gets it wrong. That takes method, not another model.

How many AI pilots actually reach production?

Few, and the figures agree even though they come from different sources:

• 95% of organisations get no measurable return from their generative AI pilots (MIT NANDA, July 2025).
• Only 4 in 33 AI proofs of concept make it to production (IDC, March 2025).
• More than 40% of agentic AI projects will be cancelled before the end of 2027 (Gartner, June 2025).
• Gartner also expects 60% of AI projects without AI-ready data to be abandoned.

None of these figures says AI doesn’t work. They say the jump from demo to operations is where almost everything gets lost. And that jump can be planned from day one.

Why do AI pilots stall?

A demo and a production system answer different questions. The demo asks “can it do this?”. Operations ask “does it do it every day, with real data, and what happens when it fails?”. Pilots that stall usually trip over one of four things.

The demo data isn’t your operational data

The test set was cleaned by hand, once, for the occasion. In production, data arrives incomplete and late, in formats that change, from sensors that drift. If nobody has measured what it takes to get that data ready every day, the pilot works exactly once.

Nobody agreed what success looks like

“It works well” isn’t a criterion. Without a metric, a threshold and a baseline to compare against, the pilot can’t fail, but it can’t prove it worked either. And what can’t be proven doesn’t get a production budget.

The pilot lives outside your systems

A model on a laptop or in a spreadsheet changes nothing in your operations. The value shows up when the result reaches the system your people work in: the SCADA, the ticketing tool, the operations console. That integration is often more work than the model itself, and it’s almost never in the pilot plan.

Nobody is accountable when the AI is wrong

Every AI gets things wrong. The question is whether you find out. A model that sounds just as confident when it’s wrong as when it’s right ends up ignored, and an alert nobody can check ends up muted. Every answer needs to say how far to trust it, warn you when the data falls outside what it knows, and reach a person who decides what to do with it.

How can you tell if your pilot is heading for a stall?

This table sums up the most common warning signs and what to do about each one. If you recognise two or more, change course before the pilot grows.

What changes when you take it to production?

The order changes: first you decide what will be measured and how it will be run, then you build. We do it in three steps:

1. A written assessment. 45 minutes with the team and a technical report that is yours, whether we work together or not. You come away knowing which use case pays off, which data it needs and where the risks are.
2. A pilot with your data. Closed scope and a success criterion agreed before we start. With your real data, not a dataset prepared for the occasion.
3. Production without black boxes. Operated, documented and handed over to your team, with training if needed.

The team that does the assessment is the team that designs, builds and runs it: the people you talk to are the people who build it. Nothing gets lost between the proposal and the delivery. And there are things we never do: resell licences, subcontract, deliver without testing, leave things undocumented or walk away without handing over the knowledge.

Do you have to build everything from scratch?

No. Much of what a pilot needs to reach production has already been built, and that makes the project cheaper:

• CogniX, our predictive AI engine, detects anomalies, predicts and classifies with your data. Every answer comes with how far to trust it, a warning if the data falls outside what it knows and which variables weighed in. It runs on your network or at the edge, with no telemetry.
• RAGNIX, answers with the manual in hand, serves your manuals to the AI assistant you already use, with complete procedures cited to the document, revision and section. If it finds no support, it doesn’t make things up: it says so. And the manual never leaves your network.

Both rest on the same idea: AI that enters critical operations has to be able to prove what it says. If it can’t justify itself, it doesn’t answer, or it tells you it isn’t sure. We go into it in what accountable AI means in practice.

Which way of working suits you?

Not every project starts the same way. These are the ways we work and when each one fits:

• Assessment: you know you have a problem, not what the solution is.
• Proof of concept: you want to check it with your data before investing.
• Fixed-scope project: the scope is clear, with a fixed price and deadline.
• Hours bank or extended team: you need senior capacity inside your team.
• Operations: you’d rather we ran it, with KPIs and regular reports.
• Training: your team will maintain it and wants to understand it in depth.

Standard operations are automated, monitored and redundant. If you need a human response at any hour, we agree it separately.

Have we done this before?

Yes. The team has brought AI into operations that couldn’t afford a black box. In behavioural analytics for a telecoms operator, the result was 63% fewer false positives and 22% more incidents detected. For the operators of a maritime surveillance system in defence, the team built AI that runs locally and answers with citations to the operating manual.

Where do you start?

With a 45-minute conversation about your pilot and a report you keep, whether you carry on with us or not. If your pilot has been stuck in demo mode for months, see how we work and book your assessment.