Forty percent of AI initiatives at Europe's largest enterprises are kept as perpetual pilots — never killed, never scaled — "by design." That figure does not fall as organizations get better at AI. It rises. Inside the most mature AI programs, the share of permanent pilots climbs to 48%, even as average AI spend hits £39.2 million a year and grows 27% year over year (Valliance, via Consultancy.uk, 2026). More experience, more budget, more sophistication — and more pilots that go nowhere. That is the paradox, and it is the clearest signal yet that the thing breaking AI returns is not the experiments. It is the missing decision after them.
The reflexive read of that data is "AI pilots fail, so be careful about starting them." That is exactly the wrong lesson. The sharper finding buried in Valliance's survey of 1,000 senior leaders is that experimentation is working fine; what is absent is the kill-or-scale gate — the moment where someone looks at a pilot, against pre-committed metrics, and rules it either dead or into production. For a Head of Operations at a 200-FTE company finalizing next year's AI plan, the AI pilot-to-production problem is not a reason to slow down. It is a reason to install one decision your enterprise peers, with all their budget, still haven't.
The Paradox: Maturity Should Cure Pilotitis. It Doesn't.
The intuitive model of organizational learning says the more AI projects you run, the better you get at finishing them. You build muscle: you learn which use cases scale, you kill the duds faster, your hit rate climbs. Valliance's data says the opposite is happening at the top of the market. The organizations with the most established programs are the ones accumulating the most perpetual pilots — 48% versus the 40% baseline (Valliance, via Consultancy.uk, 2026).
That should stop you. It means "pilotitis" — the condition of endlessly experimenting without ever committing — is not a beginner's disease that maturity cures. It is a structural condition that maturity amplifies, because mature programs have the budget to keep more pilots alive indefinitely and the political complexity to avoid killing any of them. Each pilot has a sponsor. Each sponsor has a reason it is "still learning." Nobody owns the verdict. The result is a portfolio that grows at the edges and never resolves in the middle.
Valliance is precise about the mechanism: the failure is not the experiment, it is what happens — or doesn't — next. Poor success metrics, low adoption, and consulting engagements structured to extend rather than conclude all push pilots into a holding pattern. The £39.2 million average spend is not buying production systems. A large share of it is buying the right to keep running experiments that no one will ever formally end.
Why "Experimentation Failed" Is the Wrong Diagnosis
Here is why the distinction matters operationally. If the diagnosis is "AI doesn't work," the treatment is to deploy less. If the diagnosis is "we never decide," the treatment is governance — and the upside of getting it right is large and well-documented.
When AI is genuinely integrated into a workflow rather than left in pilot purgatory, the performance delta is not marginal. The Harvard Business School–BCG field experiment on knowledge workers found that consultants using AI properly completed tasks roughly 25% faster and produced work rated about 40% higher in quality than the control group (Harvard Business School & BCG, 2023). That is the prize sitting on the other side of the gate. The perpetual-pilot organization is paying full price for AI and collecting almost none of that return, because the 40% quality lift only materializes when the tool is in production, in the daily workflow, with adoption — not when it is being "evaluated" in a sandbox for the eighteenth month.
The cost of not deciding shows up in the macro data too. MIT's widely cited 2025 analysis found that roughly 95% of enterprise generative-AI pilots delivered no measurable impact on the profit-and-loss statement — only about one in twenty crossed into real financial return (MIT, via Fortune, 2025). Read alongside Valliance, the picture sharpens: the problem is not that AI can't pay off — the few that integrate it well do extremely well. The problem is that the overwhelming majority of pilots are never forced to prove they pay off, so they sit in the 95% by default.
That is the diagnosis. Not "experimentation failed." Experimentation succeeded and was never converted. The missing organ is the decision gate.
The Cost of a Missing Gate Is Bigger Per-Dollar for the Mid-Market
It would be comfortable to file this as an enterprise problem — £39.2 million budgets, 1,000-leader European multinationals, a scale of waste a mid-market company will never reach. That comfort is misplaced, and the reason is arithmetic.
An enterprise running £39.2 million of AI spend can carry a dozen zombie pilots as a rounding error. The waste is real but diluted across a vast budget; the perpetual pilot is a line item the CFO tolerates. A 200-FTE company cannot dilute anything. If you are running three AI pilots and two of them are perpetual-by-default, you are not wasting a rounding error — you are misallocating a meaningful fraction of a discretionary technology budget that was hard to win in the first place. The percentage of waste may be similar; the survivability of that waste is not. The mid-market feels every stalled pilot in a way the enterprise is structurally insulated from.
There is a second asymmetry. The enterprise has governance functions — a transformation office, a portfolio review board, a CIO whose staff can at least see the zombie pilots. In a 200-FTE company, the pilot your operations lead started with a vendor over the spring has no review board waiting to judge it. If you do not personally build the gate, there is no gate. The Valliance finding that even mature enterprise programs fail to kill pilots is, read correctly, a warning: if organizations purpose-built to govern this still can't, the mid-market will not back into the discipline by accident. It has to be installed deliberately.
The Counter-Argument: "Bain Says the Use Cases Are Already Scaling"
The strongest objection to all of this comes from a credible source, and it deserves a straight answer. Bain & Company's 2026 executive survey advances something close to a counternarrative — that across most use-case categories, enterprises are moving AI from pilot into production, and the "everything is stuck in pilots" story is overstated. If Bain is right, the kill-or-scale gate is solving a problem that is already resolving itself.
Both findings can be true at once, and holding the tension is more useful than picking a side. Scaling is uneven. An organization can be genuinely moving its best two or three use cases into production while simultaneously accumulating a long tail of pilots that will never resolve — the 40% to 48% Valliance measured. Bain is counting the winners that crossed over; Valliance is counting the backlog that didn't. The mid-market lesson is not "ignore Bain." It is that the organizations succeeding at scaling are precisely the ones making explicit scale decisions — and the ones drowning in perpetual pilots are the ones that never built the mechanism to decide. The gate is what separates Bain's scaling story from Valliance's stalling story. You want to be on the right side of that line on purpose.
The supporting evidence points the same way. EY's Work Reimagined research found that while nearly nine in ten employees now use AI in some form, only around a quarter of organizations are actually positioned to convert that deployment into high-value outcomes (Deloitte, State of AI in the Enterprise, 2026). Usage is universal; conversion is rare. The differentiator is not access to AI. It is the discipline to decide what to do with each deployment.
Build the Kill-or-Scale Gate: Three Components for This Quarter
The fix is narrow, cheap, and entirely inside your control before the next budget cycle locks spend in. You do not need a transformation office. You need a gate, and a gate has three parts. Apply it to every active AI pilot you are running today, and make it mandatory for every new one.
First, a fixed evaluation window. Every pilot gets an end date at the moment it starts — 60 days, 90 days, one quarter, whatever fits the use case, but a date, named in advance. The single behavior that creates perpetual pilots is the open-ended timeline: a pilot with no end never has to face judgment. Close the timeline and you force the verdict. If a pilot reaches its window without a decision, the default is not "extend." The default is "kill."
Second, pre-committed success metrics. Before the pilot runs, write down what success looks like in numbers you will actually have at the end — hours saved per week, error rate reduced, throughput per head, adoption rate among the intended users. The reason mature enterprises stall, per Valliance, is poor metrics and low adoption: pilots that were never given a clear bar to clear can always claim they are "still learning." A pilot with a pre-committed metric either hits it or it doesn't. Define the bar before you have a stake in moving it.
Third, a binary verdict and a named owner. At the end of the window, one accountable person — not a committee — renders one of exactly two decisions: scale it into production with a real budget and adoption plan, or kill it and reclaim the spend. There is no third option. "Extend for another quarter" is the disease, not a decision. The HBS–BCG upside — 25% faster, 40% better — is only collectable on the "scale" side of that verdict, and only when scaling means real integration, not a bigger sandbox.
None of these three requires headcount or a platform. They require deciding, in advance, that a pilot earns its continuation by clearing a defined bar by a defined date, judged by a named person. That is the entire governance floor for AI pilot-to-production, and it is the difference between a portfolio that converts and one that quietly grows zombies.
The aggregate story of Valliance's 2026 data is that the organizations with the most AI maturity are the worst at finishing what they start — and that the mid-market, which cannot afford their tolerance for waste, has the most to gain from a discipline they never built. The decision in front of a Head of Operations this quarter is not whether to run more AI pilots. It is whether a single pilot on your books today has an end date, a success number, and a name attached to the verdict. Pick your most expensive active pilot and give it all three before the budget closes. The kill-or-scale gate is not the brake on your AI program. It is the only thing that turns the spend into a return.
