How many dormant POCs do you need to pile up before admitting the problem isn't the tool?
A mid-sized manufacturing company — let's call it what it is, one case among hundreds — launched three AI pilots between 2023 and 2025. An internal chatbot for HR support. A demand forecasting model. A writing assistant for the sales team. Total budget: roughly $200,000. Result in production: zero. All three projects stayed frozen in demo mode, rubber-stamped by the executive committee, never deployed. The CIO moved on to another role in the meantime. The licenses are still running.
This story isn't unusual. It's become the norm.
Pilot purgatory, by the numbers
MIT published a study in 2025 that sent shockwaves through the tech ecosystem: 95% of generative AI pilots in enterprises produce zero measurable impact on the bottom line. Not "partially fail." Produce nothing.
Gartner, for its part, had predicted by mid-2024 that 30% of generative AI projects would be outright abandoned after the proof of concept by the end of 2025. Reality turned out worse: according to the RAND Corporation, the overall failure rate hits 80%, with a third of projects abandoned and nearly 30% generating no value whatsoever.
Globally, we're talking about $684 billion invested in AI in 2025. Over $547 billion produced no demonstrable business value. Read that number again.
In the U.S., the picture is strikingly similar: surveys show that while a majority of small and mid-sized businesses report experimenting with generative AI — up sharply from prior years — only a fraction use it consistently. Most remain stuck at the stage of occasional experimentation — curiosity-driven usage, not transformation.
What kills pilots (and it's not what you think)
The temptation is to blame the technology. The models hallucinated, the data wasn't ready, compliance requirements complicated things. These obstacles are real, but they're not the root cause.
The MIT report (NANDA Initiative, July 2025) identifies a far less dramatic culprit: the lack of upfront strategy. Objectives are vague ("explore AI"), there's no business-side sponsor, and nobody has defined what measurable success would look like. Companies launch a POC to show they're doing AI, not to solve an identified problem.
In practice, three patterns come up again and again.
The showcase POC. A project launched to impress the board or reassure investors. The demo works on a hand-cleaned dataset. Nobody planned for integration with existing systems, access management, or maintenance. The pilot "succeeds," then dies of neglect.
The technology-first POC. The IT team picks the tool before defining the need. They test GPT-4, then Claude, then an open-source alternative, compare benchmarks, fine-tune prompts. Six months later, the business side still hasn't been involved, and the project solves no real pain point.
The ownerless POC. The project is championed by a lone enthusiast. When that person changes roles, goes on vacation, or loses steam, the pilot stops. There's no recurring budget, no dedicated team, no path to production.
McKinsey, in its State of AI 2025, confirms the diagnosis: 88% of companies use AI, but only 23% have managed to deploy agentic AI systems at organizational scale. The gap between "we use ChatGPT" and "AI is transforming our operations" is a chasm.
The ones who succeed do exactly the opposite
Allianz Partners took a radical approach in late 2025: rather than launching yet another pilot, the insurer deployed autonomous automation straight into production across its UK and DACH markets. Claims processing went from 29 days to 3.5 days. The company projects $330 million in annual savings by 2027.
Their approach isn't magic. It rests on a simple principle: start with the most painful business problem, not the most exciting technology. Claims processing was expensive, slow, and generating customer dissatisfaction. The use case was crystal clear, the ROI calculable before a single day of development.
That's the dividing line between the 5% who succeed and the 95% who get stuck. The former don't ask "what can we do with AI?" They ask "which process costs us the most, and can AI compress it?" The distinction is decisive.
What this means for small and mid-sized businesses
58% of SMB leaders consider AI a matter of survival within three to five years. Yet fewer than half have adopted a formalized AI strategy. We're looking at collective cognitive dissonance: everyone knows it's important, nobody knows where to start, so they launch a pilot to feel like they're doing something.
Here's what actually works, observed across deployments that stick.
First, identify a repetitive, costly, and measurable process. The phone system, inbound lead qualification, invoice processing, appointment scheduling. Not "AI in service of innovation" — a concrete problem with a known cost.
Second, set a success criterion before you start. "Reduce processing time for X by 40%" or "free up Y hours per week for team Z." If you don't know what success looks like, you won't know you've failed either.
Third, budget for production from day one. The POC represents only 20 to 30% of the total cost. Integration, training, maintenance, monitoring — that's where the real game is played. A company that allocates 100% of its budget to the pilot has, by design, zero budget for what comes next.
Fourth, appoint a business owner, not just a technical one. AI that works in a company is driven by someone who understands the process it's transforming, not just someone who understands the APIs.
SMB AI deployments that follow this logic reach positive ROI in six to seven months on average, according to cross-referenced data from multiple consulting firms. That's not an outlandish timeline. But it requires you to stop treating AI as a subject of perpetual exploration.
The real risk isn't a failed pilot
The real risk is that the pilot becomes an end in itself. That it serves as an alibi to say "we're doing AI" without ever transforming anything. That the $200,000 spent by the manufacturing company at the beginning of this article gets multiplied a thousandfold across the economy with nothing to show for it.
42% of companies abandoned the majority of their AI initiatives in 2025 — up from 17% in 2024. The acceleration of abandonment is as rapid as the acceleration of adoption. This isn't a sign of maturity. It's a sign that many entered through the wrong door.
The question to ask isn't "should we be doing AI?" It's "are you ready to see it through, or are you funding a demo that nobody will ever deploy?"