Sanjay Brahmawar, CEO, QAD|Redzone. Works with manufacturers to help frontline teams move faster and make better decisions.getty​The headlines this year belong to the giants: Fortune 500 manufacturers unveiling "agentic factories" and launching billion-dollar platforms. This makes for a compelling story, but it hides the more important one: The real test of agentic AI is unfolding in the middle of the market.As I've talked to many manufacturing leaders, the same mistake surfaces again and again. They drop an AI agent onto an antiquated process and wait for the step change. What they get is speed—applied to the wrong thing.An agent layered over a broken workflow doesn't fix the workflow. It runs the broken workflow faster. For example, a manufacturer drowning in unplanned downtime deployed an AI agent to dispatch maintenance the moment a machine faltered. After the agent was implemented, technicians moved faster, tickets closed faster, response times dropped, but the downtime didn't. The real problem was a scheduling process upstream that ran the machines too hot to begin with. The agent never touched that. It serviced the symptom at speed, and the plant kept breaking in the same places.Independent industry analyst Robert Kramer has documented the underlying problem as plainly as anyone. Writing in Forbes, he describes how legacy ERP systems in manufacturing "quietly undermine capacity, flexibility and competitiveness in ways that become harder to unwind over time."But I'd take his diagnosis one step further. If that's the foundation, then automating it without rethinking it doesn't fix anything—it industrializes the very habits holding a plant back. It mirrors a conviction I've held for years: Infrastructure and data discipline matter more, not less, in the age of agents.Why Mid-Market Manufacturers Are Struggling With Agentic AI Adoption​None of this is an argument against agentic AI in the mid-market. The momentum is real, and the barrier is finally falling. In fact, Upwork's research on small and midsize businesses found that leaders at "74% of SMB report that AI has improved their productivity," with more piloting agents than sitting on the sidelines in every function surveyed. Turnkey platforms and implementation partners carry much of the technical load, so you no longer need a data-science department to begin.But access is not an advantage, and the friction lands harder on a mid-market P&L than on a Fortune 500 balance sheet. Three realities deserve honesty.​First, data readiness. As Forrester notes, data quality is often the single biggest blocker to AI adoption. An agent is only as good as what it can see. Many midsize plants still run on older systems with patchy shop-floor data.Second, cost predictability. Many companies are reporting sticker shock on their AI usage-based bills, and I suspect this will particularly impact small manufacturers. Many companies are now writing hard usage caps into their contracts, but custom-built agents can carry six-figure costs before they return a dollar, in my experience.Third, the ROI cliff. Gartner has predicted that more than 40% of agentic AI projects will be scrapped by the end of 2027 due to rising costs and unclear returns—a risk a midsize firm can afford far less than a conglomerate.​Three Rules For Mid-Market AI Success​So, here's my plain read. While the macro-vision of an autonomous factory carries high risk, task-specific applications have already become proven wins. Order-processing automation, materials-data lookup and document handling are exactly the high-friction, low-risk tasks midsize plants are drowning in.As I've learned from talking with manufacturing leaders, this is the real shift underway on the plant floor: moving from systems of record to systems of action, from software that documents what happened to software that does something about it. That shift is worth making. It's also where teams get hurt if they skip the fundamentals.​ The opportunity is genuine, but the winners won't be the ones chasing the agentic-factory vision the headlines sell. They'll be the ones who do three unglamorous things:• Fix the data foundation first. An agent can't act on data it can't trust or reach. The plants that get it right start with the boring part—making "line 3 stopped" mean the same thing in every system before a single agent goes live. I've watched teams spend the first month deploying nothing, just making the data real. It's the least exciting phase and the one that decides everything after it.• Start with one narrow, high-volume workflow. Instead of betting on a moonshot, prove value where the repetition is highest and the risk is lowest. The test I give teams: Find the task your best people do 50 times a day and shouldn't have to. On the floor, that's usually triaging downtime alerts or chasing quality-hold paperwork—high-frequency, forgiving if the agent gets one wrong. Nail that, earn the trust and the next one is easier to green-light.• Cap costs before the meter runs. An agent bills more like electricity than software—every task is a meter running—and I've watched a promising pilot become a bill nobody warned the CFO about. Set a hard monthly ceiling and a spend alert before the first agent runs, and start with a workflow whose usage you can predict.In other words, look at the whole business before you automate any part of it. Modernize the process, then let the agent amplify it. Done in that order, agentic AI becomes what it should be: friction removal that lets people operate with clarity, not another layer of speed on a system already straining.I'd still rather lead a midsize plant in 2026 than at any point in my career. The tools have finally caught up to the ambition. The only open question is whether leaders move with the discipline the moment demands—because the future of manufacturing won't be decided in the headlines. It will be decided on the floor, by the companies in the middle that choose to fix the foundation before they accelerate.​​Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?