Thinking of Putting an AI Agent on Your Shop Floor? Read This First

A guide for plant managers and manufacturing heads.

If you run a plant, someone has probably pitched you an AI agent by now. The promise goes something like this: the software watches your lines, figures out what’s going wrong, and has the next step ready before your team has even finished reading the alarm. Some of that is true. Quite a lot of it is old software with a new label.

The survey numbers tell both stories. McKinsey’s 2025 State of AI study polled 1,993 people in 105 countries. Sixty-two percent said their organisations were at least trying out AI agents. Dig a little deeper, though, and only 23% had moved past trials to scaling an agent anywhere in the business. The other 39% were still experimenting. And the ones who were scaling mostly did it in one or two functions. In no single function did more than 10% of respondents say they were scaling agents.

People inside manufacturing see the same drop-off. “On an average let’s say it’s about 80% of the manufacturing industries adopt gen AI and foundational models. It’s about less than 30% which actually convert what we call as a pilot or a demo into something that is deployable. Even lesser than that from deployable to the networking side, which means at a wider scale, is about 5%,” said Dr. Vasant Jain, Chief Engineer, Myelin Foundry,

Gartner is blunter. It expects more than 40% of agentic AI projects to be scrapped by the end of 2027. The reasons it gives are costs that keep climbing, business value nobody can pin down, and risk controls that weren’t good enough.

So whether agents work isn’t really the question. The question is how you end up among the plants that get something out of them instead of writing off another pilot. Based on what we’re seeing, a few things make the difference.

Pick the slow decision first

An agent earns its place by shrinking the gap between something going wrong on the floor and someone doing something about it. In most plants that gap is filled with phone calls, spreadsheets and people walking between departments to find out what happened.

So before you look at any product, sit down with your team and list the decisions that always come too late. Which calls depend on one senior person who’ll retire in a few years? Which delays cost you output, scrap or a missed dispatch every time? Start there. If a use case doesn’t speed up a decision that actually hurts when it’s slow, you don’t need an agent for it.

Experts advice to go for agentic AI only where the value or ROI is clear. Rajesh Mittal, President & Managing Director, ISUZU Motors India, put it more directly “Be it an agentic AI or multi-agent platform whatever it is, it should be adopted where it solves your problem.” He added: “Unless it adds value to your business, any technology should not be implemented for the sake of technology.”

Keep the job narrow

Your plant is already split into roles. Maintenance doesn’t make the same decisions as quality, and the energy manager looks at different numbers from the production planner. Agents tend to work best when they’re set up the same way, with one role, the data that role needs and a goal tied to how that role is measured.

It’s tempting to buy one big assistant that sits on top of everything. Don’t, at least not yet. A narrow agent is easier to test and easier for people to trust. It’s also easier to switch off if it starts doing odd things. Try it on one role and one line. Widen it only once it’s proven itself.

Sort out your data before the AI arrives

Most plants have more data than they know what to do with. What’s usually missing is context. A temperature spike tells you very little unless you also know what the machine was doing, which shift was on and what material was running. If that information lives in registers, Excel files and the heads of a few old-timers, the agent will flounder, however clever the model behind it.

Three questions are worth asking your team before you sign anything. Can we actually connect the equipment, production and quality records for the area we want to start with? Do our systems use the same names for the same assets and products? And once the agent is running, whose job is it to keep that data clean?

If nobody has a confident answer, fixing that is project number one. It’s dull work. It also decides whether everything after it succeeds.

Decide what the agent may and may not do

Settle this before go-live. For most plants right now, a sensible rule is that the agent prepares and a human approves. It can suggest a likely fault, draft the work order or propose a change to the schedule, but a planner, engineer or supervisor makes the call.

Put the thresholds, approvals and escalation steps in writing. Make sure the system can stop and hand control back to a person when it hits something it wasn’t built for. Repairs should still be signed off and carried out by your technicians. These rules aren’t just about safety. They’re also what gets your people to accept the thing in the first place.

Pay attention when people say no

Your operators will reject some of what the agent suggests, and that’s fine. A good system learns from what happens afterwards: whether the fault got worse, how it was actually fixed, which recommendations were taken, tweaked or ignored.

Keep an eye on how often people override it. If the number stays high, find out why. Maybe the model has it wrong. Maybe the data feeding it has gaps. Or maybe its suggestions just don’t match how your plant really runs. It also helps to ask vendors whether the agent shows its working, meaning the data and logic behind each recommendation. Engineers are far more likely to act on something they can check.

Don’t underestimate the wiring

Getting a demo to run on a laptop is easy. Getting an agent to read from your MES, CMMS and historians, and then write back into your planning system, is the real job. Experts say connecting agents to older systems can be technically hard and expensive.

So set aside time and money for integration from the start, and ask vendors to show you that part rather than just the dashboard. A pilot that can’t talk to your systems hasn’t really proved anything.

Make sure you’re buying an agent

Have you heard of the term  “agent washing.” That’s when vendors take existing chatbots, assistants or robotic process automation tools and rebrand them as agentic AI, even though they can’t do much on their own. Talk around and the feeling you may get is that only a small share of the thousands of companies selling “agentic AI” are offering the genuine solve.

There’s a quick way to check. Does the product just answer questions or raise alerts? Or can it reason across your data, recommend an action within limits you’ve set, and get the work ready for someone to approve? If it’s the first kind, it might still be worth having, but you’re buying a chatbot or an alarm system. Pay for it as one.

What’s coming either way

Industry reckons that by 2028, at least 15% of day-to-day work decisions will be made autonomously by agentic AI, up from none in 2024. It also expects a third of enterprise software applications to include agentic AI by then, compared with less than 1% in 2024. In other words, agents will turn up in your plant through software you already pay for, whether you go looking for them or not.

That’s a good reason to do the groundwork now. Work out which decisions matter, get the data behind them in order, and agree on what the software is allowed to do. The technology will keep getting better on its own. Whether your plant is ready to use it is up to you.