According to a new Manufacturing Leadership Council survey, more than 70% of industrial organizations are now using or planning to use generative artificial intelligence, the first step to agentic systems on shop floors. That’s a striking jump from two years ago, when just 46% of manufacturers used AI.
With labor shortages and downtime costs climbing, the appeal of agentic AI is obvious. But leading a maintenance and reliability software company has shown me that manufacturers who capture real value from agents will be the ones who do a specific kind of preparation first—and most leaders are skipping it.
Gartner forecasts that over 40% of agentic projects will be scrapped by 2028 due to spiraling costs, lackluster return on investment, or unforeseen risks. The cancellation rate isn’t a reason to avoid agentic AI: The upside is too significant to ignore. It’s a signal that the conventional rollout playbook of buying the platform, identifying use cases, and launching isn’t as straightforward in industrial environments.
WHY THE STANDARD PLAYBOOK FAILS
Part of the problem is that the agentic AI market is overheated. In 2025, venture capital firms poured $24.2 billion into agentic startups—up from $6.9 billion in 2024. That’s led countless companies to position themselves as agentic innovators whether or not the label fits. Today, thousands of vendors claim to provide agentic technologies—but only about 130 are selling true agentic solutions.
Overhyped technology paired with impatient timelines is a dangerous combination. About one-third of leaders expect positive ROI from their AI projects in a year or less, despite 95% of organizations struggling to generate value from past AI investments. The risk is that bullish leaders will overcommit to underbaked AI technologies, while underinvesting in the data and infrastructure needed to unlock real value.
The pattern isn’t new. Consider the generative AI boom. Around 80% of companies now use generative AI—but the vast majority aren’t generating value from their investments. McKinsey attributes this to an imbalance between horizonal and vertical deployments: It’s easy to roll out tools that summarize emails, but much harder to map AI tightly onto specific business functions. Fewer than 10% of such specialized use cases evergetpast the pilot stage.
The same pattern is now playing out with agentic AI: Organizations are rushing to deploy agents without doing the hard work of identifying use cases, collecting data, and building paths to value.
That’s a big problem. If a generative AI project fails, you’ve wasted time and money. But if an agentic project goes south, you’ve put unproven AI in a position to do real harm—ordering incorrect parts, issuing unnecessary work orders, canceling manufacturing runs, or triggering safety and compliance incidents. With agents, the cost of failure isn’t lack of ROI—it’s real financial and reputational damage.
THE SEQUENCE THAT WORKS
Manufacturers who get real value from agentic AI share a common pattern: They treat deployment as a sequence, not a switch. Three steps, in this order, are the pattern in practice:
1. Capture what’s in technicians’ heads. Some 43% of leaders see data readiness as the biggest barrier to their AI goals, but few manufacturers currently collect data from the work itself. The knowledge that matters most lives in people’s heads, not in any system an agent can read. For example, a senior technician knowing to replace a bearing at 8,000 hours instead of waiting for the alarm or recognizing which symptoms predict a failure two work shifts later. Until that context graph and the subsequent decision traces are structured and accessible, agents will make decisions without the context their human counterparts take for granted.
2. Start agents on reversible decisions. An agent that drafts a work order is doing something a human can review in seconds. An agent that auto-approves a $50,000 parts order or cancels a production run is making decisions that are hard to walk back. Start with the first category. Sandboxes, spending caps, and approval thresholds let agents earn broader authority by proving themselves where the cost of being wrong is measured in minutes, not millions.
3. Know what to leave alone. The instinct in any AI rollout is to keep expanding the footprint, but the most disciplined operators treat scope as a deliberate choice. According to the Deloitte Research Center for Energy & Industrials, 81% of manufacturers’ task hours will stay human driven. Of that, 39% of task hours will never be well-suited to agentic automation, 36% may benefit from it, and 6% may benefit from robotics. augmentation. Resisting the pressure to deploy autonomous agents in places they don’t belong is what separates a thoughtful program from a costly one.
Each step builds on the one before it. Without structured workflow data, agents can’t make context-aware decisions. Without reversible starting points, you can’t build the operational trust that justifies expanding agentic authority. And without clear scope discipline, you end up with agents in places that frustrate workers and create risk without delivering returns. Skip a step and the whole sequence collapses.
THE PAYOFF OF GETTING THE ORDER RIGHT
Manufacturing isn’t a sector that can iterate in public. A bad agentic decision shows up in scrapped parts, missed shipments, and occasionally, injured workers. The sequence above isn’t about being cautious for its own sake. It’s about giving agents the conditions they need to deliver results.
Manufacturers who structure workflow data first will have agents that can reason about real operating conditions. Those who start agents on approved/reversible decisions will build the trust required to expand their authority later. Holding line on scope will avoid the cancellations and rework that will define most of this market over the next two years.
Agentic AI is a powerful technology, but it’s not a plug-and-play. The manufacturers who treat it that way will end up in Gartner’s 40%. The ones who follow the sequence—data, approval/reversibility, scope—will be the ones running plants that, three years from now, have agentic AI doing meaningful work on the floor.
Chris Turlica is CEO of MaintainX.