Why most industrial AI experiments fail—and how to fix them.
Nearly nine in ten organisations worldwide now deploy artificial intelligence in at least one business function. Yet precious few have managed to embed it in core industrial operations. The problem is not the technology. Companies are launching pilots that solve nothing in particular.
Walk into most large industrial firms and you will find a flurry of AI experiments running in parallel—safety monitoring here, predictive maintenance there, perhaps a dash of energy optimisation or robotic automation thrown in. On paper, it looks like progress. In practice, these initiatives rarely escape the lab. Giedrė Rajuncė, chief executive of GREÏ, an AI-powered operational intelligence platform, has watched this pattern repeat itself. “What we see again and again is not a lack of ideas, but a lack of focus,” she says. “Companies can easily list dozens of AI use cases, but many start pilots without agreeing on the pain point they are solving, what success actually looks like, or who owns the outcome.”
The result is what Rajuncė calls “pilots for the sake of pilots”—a scattergun approach that fragments attention and slows decision-making. When several experiments run simultaneously, each competing for resources and executive buy-in, none receives the care required to succeed. This helps explain why roughly 95 per cent of enterprise AI initiatives fail to generate measurable impact on profit and loss, according to industry estimates.
Fear, hype and a shortage of nous
Why do companies persist with this? Rajuncė identifies three culprits. First, fear of missing out—no executive wants to be caught flat-footed while rivals experiment with AI. Second, top-down pressure to innovate, often imposed without consulting the people who will actually use the technology. Third, a lack of internal competence to manage pilots and implement change. “When those elements are missing, pilots may look successful on paper, but they rarely survive contact with real operations,” she says.
The mistake lies in starting with the technology rather than the problem. Instead of asking what AI can do, firms should identify one operational headache that is both painful and expensive—safety incidents, production bottlenecks, inefficiencies that eat into margins. “The most limited resource is time,” Rajuncé notes. “A pilot needs active management and coordination. When one person is responsible for several pilots at once, execution suffers, and ROI disappears.”
GREÏ’s own approach is stricter than most. Before launching a pilot, the firm insists on clear baseline data, defined success thresholds, pre-agreed next steps, dedicated users on the ground, fast-tracked IT support and hard stop dates. This discipline does more than improve outcomes—it also builds trust with frontline teams, who are often sceptical of startups promising algorithmic salvation. “Trust is critical, especially when a startup is the implementation partner,” says Rajuncė. “I want to know the technical task upfront so I can say with confidence the pilot will work.”
The broader lesson is straightforward. AI pilots should not be about ticking boxes or impressing the board. They should address specific, measurable problems with clear ownership and realistic timelines. Without that, companies risk burning through budgets and goodwill while their competitors—those who focus on one painful problem at a time—quietly pull ahead. “In the end, it’s KPIs, trust, and strong participation that determine success,” Ms Rajuncė says. “There is no other way to build a pilot that actually scales.”
Photo: Dreamstime.
