Steering committee meetings tend to follow a familiar pattern. Someone presents a pilot result. The room nods. Then someone asks the only question that matters: what happens when this rolls out to the other four plants? Usually, nobody has a confident answer. That’s not a technology gap. It’s a readiness gap that got skipped on the way to the demo.
The data backs this up, and it’s worth sitting with the scale of it.
Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from under 5% in 2025. That’s the adoption curve everyone’s watching. What gets less attention is Gartner’s other forecast from the same research: more than 40% of agentic AI projects will be cancelled by 2027, mostly over unclear business value. Same firm, same year, opposite trajectory. Adoption and failure are climbing together.
Manufacturing shows this pattern just as clearly. Grant Thornton’s 2026 AI Impact Survey found 48% of manufacturers are piloting AI, but only 10% have fully integrated it into operations. Zero percent of manufacturing respondents in that survey reported significant revenue uplift from AI, compared to 12% across all other industries surveyed. The pilots are happening. The value mostly isn’t landing.
This problem hits harder the smaller the organization gets, and the reasons are structural, not a matter of trying harder.
Reason one: the benchmark probably wasn't built for this OEM
Most of the AI ROI numbers circulating in industry content, the case studies, the vendor decks, the conference keynotes, come from large enterprise deployments. That’s not a coincidence. Enterprises are the ones with dedicated AI teams, budget for extensive documentation, and a marketing department eager to publish the win. A mid-sized OEM applying one of those numbers directly is solving for the wrong scale before the pilot even starts.
A 30% downtime reduction at a billion-dollar manufacturer with a dedicated data engineering team and five plants to average across is a different proposition than the same claim applied to a single-site OEM running $50M in revenue with one IT generalist. The benchmark isn’t wrong. It’s just not describing that starting point.
Reason two: mid-sized OEMs get one real attempt, not five
A large enterprise can run three or four parallel pilots, quietly retire the ones that don’t pan out, and keep the one that worked. Nobody outside the building ever hears about the failures. A mid-sized OEM rarely has that luxury. Leadership greenlights one pilot, watches it closely, and if it doesn’t show a defensible result within a quarter or two, the appetite for a second attempt tends to disappear along with the budget line.
That changes the calculus entirely. Picking the wrong first use case doesn’t just cost the resources spent on that pilot. It often costs the entire initiative, because there usually isn’t a second chance built into the plan.
Reason three: governance catches up eventually
Deloitte’s 2026 research found only one in five companies has a mature governance model for autonomous AI agents. That statistic tends to read as an enterprise problem, security teams, compliance frameworks, the kind of thing a Fortune 500 company worries about. In practice, it shows up just as often at smaller OEMs, just in a different form: nobody owns the decision the AI system makes, nobody’s accountable when it’s wrong, and the pilot quietly stalls because no one wants to be the person who signs off on scaling something they can’t fully explain to the plant manager.
KPMG’s Q1 2026 survey found 65% of leaders cite scaling difficulty as their top ROI barrier. That’s rarely a model performance problem. It’s almost always an ownership and process problem that the pilot phase never had to confront, because a pilot, by design, has someone watching it closely. Production doesn’t work that way.
Reason four: no two OEMs run the same shop floor
This is the one that tends to get underweighted the most. A high-volume, low-mix line and a low-volume, high-mix job shop aren’t starting from the same place, not in data density, not in process consistency, not in how clean the historical record even is. McKinsey’s 2025 State of AI survey found only 21% of AI adopters had actually redesigned a workflow around the technology. The other 79% layered a tool onto an unchanged process and then were surprised when the results looked thin.
That gap matters even more once variation across OEMs is factored in. A workflow redesign that works cleanly on one production line doesn’t transfer automatically to a plant with different equipment, different data maturity, and different team bandwidth. The generic playbook doesn’t generalize the way it looks like it should on a slide.
The payoff difference between the two groups isn’t small. McKinsey’s data shows only 13% of companies at the basic adoption stage report meaningful value from AI. That climbs to 48%, more than triple, for companies that fully redesigned how the work gets done before scaling. Same technology. Very different outcomes, and the difference traces back to preparation, not the tool.
The common thread
Every one of these four reasons traces back to the same root cause: the organization didn’t know its own readiness or its own ROI range before it committed to a pilot. Not the technology’s fault. Not a sign the team didn’t work hard enough. A sequencing problem, solved by asking the right question in the right order.
Put plainly: the pilots that fail usually weren’t wrong about the technology. They were wrong about the starting point.
What to check before the next pilot
Before picking a use case, get specific about three things. Where does the data actually live, and can it be trusted? Does the team have the bandwidth to act on what the pilot surfaces, not just run it? And what’s the realistic dollar range for this specific workflow, grounded in the organization’s own downtime, scrap, or cycle-time numbers, not an industry average pulled from someone else’s plant?
That third question is the one most pilots skip entirely, and it’s usually the one that determines whether the leadership meeting six months later is a celebration or a post-mortem.
This is the exact sequence Signal AI, by Gowitek, is built around: capture context on the workflows, run a structured readiness score, surface ROI bands grounded in the organization’s own numbers, then build a 90-day roadmap sequenced by what’s both ready and worth doing first. It starts at $0, because this is the question worth answering before capital is committed, not after.
If the last AI conversation ended with “let’s pilot it and see,” it’s worth asking one question first: see what, exactly, and against what number.