AI projects often don’t come past pilot stage, finds Valliance study
As AI spending comes under scrutiny from investors, consultants are making the case that the fault may not be with the technology, but a poor understanding of how to measure impact on pilots. A new study from Valliance has shown that close to half of AI initiatives at large European businesses remain at the pilot stage – while those that scale often see limited returns on investment.
After years of breathless hype, 2026 has seen consulting firms drastically reframe the way they discuss artificial intelligence. Most pointedly, PwC – which declared in 2023 that Generative AI was “so powerful and easy to use” it had already “reached a tipping point”, and was poised to “ reinvent entire industries” – softened its tone three years down the line, admitting that questions around returns on investment result in a “silence that often follows” in board-room meetings, because “all that AI activity isn’t producing measurable returns.”
The agreed narrative forming around this is not that proponents of the technology have over-promised – but rather that clients have mishandled the integration process. A mounting body of research from professional services firms now notes that the vast majority of AI projects get ‘stuck’ at the pilot stage, rather than being scaled to deliver value across a whole organisation. For example, late in 2025, IBM polled an international cohort of CEOs, and found while a year earlier, two-thirds expected to move beyond the piloting phase of AI changes, when the researchers circled back, 60% were still stuck in the nascent period of experimenting.
Now, a new study from Valliance has further reiterated this point, after surveying 1,000 senior leaders at Europe’s largest enterprises. Ultimately, the researchers found that, although investment is still accelerating – with 27% year-over-year growth, and spending averaging £39.2 million annually among UK organisations – value remains hard to come by.
Excitement remains
Looking at where enthusiasm is the highest, unsurprisingly, IT and technology companies remain most excited by AI – investing £45.8 million annually on average – while that is followed closely by those in the finance, insurance, and legal sectors at around £44.2 million. These are not small commitments to the belief of AI’s potential, but that makes the fact so many of them end up delivering the outcomes of a fleeting experiment even worse.
In this case, Valliance found that 40% of AI initiatives remain pilots by design – rising to 48% in mature organisations with established AI programmes. According to the firm, this is proof that “experimentation itself isn’t the problem” – rather, what happens next, “or more accurately, what doesn’t happen” is the key. Projects stall between proof of concept and production. Lessons aren’t learned or applied elsewhere and therefore “capability doesn’t scale”.
The gruesome sounding “pilotitis” is in fact to blam – with Valliance arguing that “endless experiments stalled by poor metrics, low adoption, and misaligned consulting models” mean that firms often waste the potential of pilots, because they misunderstand the potential applications of the results. Businesses stuck in pilot stage report lower success rates – 43% against 50% overall – while taking 6.6 months to see value, as opposed to 5.9 months, and enjoying weaker ROI – with only 20%reporting strong ROI, compared to in 76% of cases where a pilot is scaled in mature organisations.
Valliance subsequently claimed that mature organisations can lead the way here – treating pilots as rapid learning cycles with clear gates to production, “rather than endless science projects conducted for the sake of it”. They “measure obsessively and kill what doesn’t work quickly, invest in their people, and partner with firms whose success depends on their success, not on how many hours they can bill”. In similarly harnessing these metrics, the firm concludes that others can emulate their success.
Stephen Treloar, Valliance board advisor, said, “When electricity was first invented we didn’t go straight to inventing smartphones and TVs but now, it’s that undercurrent powering everything businesses do. AI will be the same. We’re only seeing the beginnings of what it’s capable of, and soon enough it too will be indispensable for any company that wants to stay competitive.”
Returns on investment
Valliance suggested that this can also see workers achieve demonstrable gains when empowered correctly. Harvard Business School research shows 25% faster output and 40% higher-quality work when AI is integrated properly into workflows. Regardless of whether blame for AI’s underperformance lies with the technology itself, or the people and organisations implementing it, however, a number of experts are questioning whether heightened ROI can actually be maintained across the economy.
In 2025, a now-infamous paper from MIT found that fewer than one-in-ten firms had seen positive financial impacts from implementing AI. That study did show that that portion of firms were doing very well indeed – Aditya Challapally, the MIT researcher who led the study, telling Fortune that some large companies and younger startups are "excelling" with AI because "they pick one pain point, execute well, and partner smartly with companies who use their tools." But rather than concluding other companies should attempt to replicate these policies across every other organisation, that may suggest that the ‘AI revolution’ simply does not make sense for larger, established players in a market.
At the same time, a number of leading adopters of AI tools have been caught red-faced in recent months, having suddenly seen charges from their suppliers rise exponentially. As largely loss-making organisations like OpenAI and Anthropic look to appease investors, some customers are finding it hard to justify spending on AI – as the costs begin to surpass the fees of human talent, without delivering returns. Axios’ Madison Mills recently reported that one company had accidentally spent $500 million in the space of a month on Anthropic’s models after failing to set spend limits.
