Every CEO I know has an AI initiative. Almost none of them can point to a dollar it has added to EBITDA.

That's not a knock on the technology itself. Over more than 35 years advising manufacturing and industrial leadership teams on performance improvement, I've watched a lot of tools clear the “does it work” bar and stall at the “does it pay” bar.

See also: Physical AI: Where opportunity really sits—now, today—for manufacturers

AI is doing the same thing right now, at scale, across nearly every industrial company I work with. The models are good. The dashboards are polished. The pilots get funded, extended, and occasionally turned into a press release.

What they don't do, in most cases, is show up in a quarterly number anyone on the leadership team is willing to defend in front of the board.

I don't think that's a technology problem. That’s a trust problem.

The gap between adoption and value

Gartner's 2026 CIO and Technology Executive Survey found that only 17% of organizations have actually deployed AI agents so far—even as more than 60% expect to within the next two years.

That's one of the steepest adoption curves Gartner tracks for any emerging technology. Read that gap the way an operator reads it, and it says something uncomfortable: Industry is scaling intent much faster than it is scaling proof.

I'd ask a simpler, more provocative version of that question to any group of peers: Has AI delivered any measurable financial benefit yet, or has it just changed how busy everyone looks?

See also: U.S. agencies report cybercriminals used AI-generated code to crack Siemens PLCs

On the plant floor, “busy” is easy to manufacture. Add a layer of predictive maintenance alerts, a forecasting model, a chatbot that summarizes shift reports, and suddenly there's more activity, more dashboards, more meetings about the dashboards.

None of that is the same as lower scrap rates, fewer unplanned downtime hours, or working capital that actually comes down. Activity is not value. Leaders who've spent real time on a shop floor know the difference instinctively. It's the ones evaluating AI from a slide deck who sometimes don't.

Why the constraint isn't the model

Here's what I've come to believe after walking plant floors across manufacturing, transportation, energy, and a dozen other sectors: The technology is rarely the bottleneck anymore. The constraint is decision quality—whether the humans receiving the model's output trust it enough to act on it, override a habit, or change a standard operating procedure because of it.

That trust doesn't show up automatically just because the model is accurate. It shows up when three things are true.

This is a pattern I've seen play out with almost every performance lever, not just AI. Executives gravitate toward the move that's tangible—more capital, more capacity, more technology—because it shows up cleanly on a plan and gives the board something concrete to point to.

The harder, less visible work of fixing how decisions get made rarely gets the same attention, even though it's usually where the real constraint sits. AI is just the latest example of that instinct.

First, governance has to be real, not decorative. A model recommending a maintenance schedule or a pricing change needs an owner, a review cadence, and a documented failure mode—the same rigor a plant would apply to a new supplier or a capital project.

See also: DX is diminished if your workforce isn’t upskilled

Too many AI initiatives skip this because governance feels like it slows down the “innovation.” In practice, the absence of governance is exactly what stalls adoption at the plant level, because operators have learned, correctly, not to trust a black box with no accountability behind it.

Second, the people closest to the work must be part of building the tool, not just receiving it. I've seen forecasting models that were technically sound get quietly ignored on the floor because the planners who'd use them every day weren't consulted on what “good” looked like.

A model nobody trusts doesn't get used, and a model that doesn't get used generates zero EBITDA no matter how sophisticated it is.

This is where a lot of enterprise rollouts quietly fail: The build happens in a conference room, and the first time the frontline team sees the tool is the day they're told to adopt it.

That sequencing almost guarantees skepticism—trust in a decision-support tool is built the way trust in a new supervisor is built, through repeated exposure to it being right, not through a rollout memo.

See also: ChatGPT won't run your plant, but here's what industrial AI needs instead

Third, leadership has to be willing to measure the thing they're actually trying to change—cost, cycle time, quality, cash—rather than a proxy metric like “number of use cases deployed” or “percentage of employees trained on AI.”

Those are activity metrics. They make for good slides. They don't answer the only question that matters to a manufacturing P&L: Did this change what we spend, what we make, or what we collect?

What decision quality looks like

The manufacturers I've seen get real value out of AI didn't start with the most ambitious use case. They started with a narrow, high-friction decision—reorder points on a volatile input, or root-cause triage on a recurring defect—where the cost of a wrong call was well understood and easy to measure.

They built the governance and the trust on something small enough to verify, then expanded once the model had earned credibility with the people who had to live with its recommendations. That sequencing matters more than the sophistication of the underlying algorithm.

It's a less exciting story than “we deployed enterprise-wide generative AI.” It's also the version that shows up in the numbers.

There's an organizational discipline underneath this that's easy to overlook: Someone must own the model the way they'd own a piece of equipment. That means a named person accountable for its accuracy, a schedule for checking whether its recommendations are still holding up as conditions change, and a clear process for what happens when it's wrong.

See also: Podcast: 'Muddy waters' of implementing AI and how manufacturers can avoid them

Manufacturing already knows how to do this—plants do it for machines, for suppliers, for quality systems. Most haven't consistently extended the same discipline to the algorithms making recommendations alongside them.

Until they do, “AI governance” will keep meaning a policy document nobody on the floor has read, instead of an operating habit that earns trust one correct call at a time.

The question I'd put to my peers

So, here's what I'd genuinely like to ask other manufacturing and industrial leaders, not rhetorically but as a real gut check: If you stripped away the dashboards, the training completion rates, and the press releases, what has AI done to your EBITDA this year?

Not what it might do. Not what the roadmap says. What has it done?

I suspect most honest answers land somewhere between “not much yet” and “we're not sure how to measure it.” That's not damning—new capability takes time to convert into results, and manufacturing has never been an industry that mistakes hype for cash flow.

But it’s a signal that the next phase of AI in industrial operations isn't about better models. It's about better decision architecture: Governance leadership that actually enforces, trust built at a scale people can verify, and metrics tied to the P&L instead of the press cycle.

The technology cleared the bar a while ago. It's our job, as leaders, to build the trust that lets it clear the one that actually counts.