Manufacturers are deploying AI across operations, but the harder question is whether those tools are producing measurable gains in cost, throughput and efficiency.
getty
Manufacturers are adopting AI aggressively, especially in operations. But the returns are harder to prove, and the way factories buy AI may be part of the problem.
In Grant Thornton’s 2026 AI Impact Survey, none of the 100 manufacturing leaders surveyed reported a significant revenue increase from AI. None reported significant cost savings either. Across the other industries in the same study, 12% of respondents reported each. A zero inside a sample that size isn’t a rounding error. It’s a warning.
On paper, manufacturing should be easy ground for AI. Factories already run on sensor data, repetitive processes and decades of automation. Yet the sector shows one of the widest gaps between AI activity and AI results anywhere in the survey. The technology usually isn’t the thing holding it back. The buying process is.
Adoption Was Never The Issue
Manufacturers aren’t ignoring AI. They’re already using it, especially in operations. In Grant Thornton’s survey, 64% of manufacturers reported efficiency gains, and 62% said operations is where they most want more AI, the highest share among the sectors surveyed.
What happens next is the problem. 48% are still stuck in pilots, against 34% across industries, and only 14% report faster innovation versus 31% elsewhere. Efficiency is real, but it isn’t the same as cost savings. A model that shaves a few minutes off a changeover looks great in a demo. Unless it turns into less scrap, fewer unplanned stoppages, lower inventory or fewer warranty claims, it never reaches the P&L in a form a CFO will sign.
They Bought The Anxiety, Not The Problem
Ask why they’re spending and the picture sharpens. 45% say competitive pressure is the main force behind their AI adoption. Not a bottleneck they’ve costed out, not a defect rate they’re chasing. The worry that a rival is a step ahead.
Anxiety is a bad procurement criterion. Grant Thornton’s analysts put the pattern plainly: manufacturers buy AI tools and then wait for the vendor to figure out how to deploy them. That sends money toward whatever peers are doing instead of the handful of decisions that actually move margin.
You can see it in where the spend lands. Operations sounds like the disciplined choice, but it’s the hardest place to make AI pay. A lot of the data is patchy or locked inside PLCs and older line equipment that was never built for it. Integration means touching machines nobody wants to stop. And on a running line, a wrong answer gets expensive fast. In the procurement reviews I’ve sat in on, the step that tends to get skipped is the same one: teams jump from “AI could help here” to “let’s run a pilot” without naming the number the project is supposed to move. The model gets scoped before the business case does.
Pilot Purgatory Has A Cause
The 48% stuck in pilots aren’t stuck on the technology. They’re stuck because no one owns a number. A pilot with no P&L target can’t really succeed or fail. It can only continue.
The broader research points the same way. MIT Media Lab’s Project NANDA, in its GenAI Divide study, found that after $30 billion to $40 billion in enterprise generative-AI spending, only about 5% of integrated pilots were pulling out real value. The rest showed no measurable P&L impact. The ones that worked tended to target a specific process with a clear owner, and most were bought rather than built. Externally sourced tools succeeded at roughly twice the rate of internal ones. For an industry that likes to engineer its own answers, that stings. Building in-house is often the right call on the floor. With AI it can backfire once integration and upkeep cost more than the model ever did.
Manufacturing also has less patience for soft wins. In a back-office function, an assistant that saves people an hour a day can justify more tinkering. On the floor the test is harder. The system has to move downtime, yield, rework, inventory or throughput, and do it without adding risk anyone will lose sleep over. That’s what makes the sector a useful proving ground. You can’t dress activity up as results when the results are supposed to come out as parts.
There’s a governance version of the same gap. Only 7% of manufacturers say they have a tested plan for when AI gets something wrong, the lowest of any sector. So most of the plants running AI in scheduling, quality or supply chain have never rehearsed the failure. These are companies that drill for fires and load-test their backup generators. The software now influencing decisions on the line hasn’t had the same treatment.
The Fix Is Procurement Discipline, Not A Better Model
The manufacturers who close this gap won’t be the ones holding out for a smarter model. They’ll be the ones who change how they buy. Start from a problem you’ve already costed, not from something a competitor announced. Name the metric up front, whether that’s scrap rate, unplanned downtime or days of inventory, before the first vendor walks in. Put one executive’s name against that number. Then run it like any other capital request, with a hurdle to clear and a date it dies if it doesn’t. That last part tends to get waived for AI, because the technology is impressive, and impressive is easy to confuse with valuable.
Before funding the next operations project, an executive should be able to answer four questions in a sentence each. What line does this move? By how much? Who owns the result? When do we stop if it isn’t working?
Manufacturing doesn’t have an AI interest problem. It has an AI proof problem. The companies that fix it won’t be the ones chasing the flashiest demo. They’ll be the ones that require AI clears the same bar as every other investment on the floor: show me the number, show me the owner, and show me when we stop if it doesn’t work.
Leave a comment