Agentic AI in Steel: Why Your Plant Already Has the Data, But Not the Decisions
It’s 11:40 PM. A planner at a hot strip mill gets a call, an export order has been pulled forward by two weeks. Somewhere in his head he’s now running a simulation no software is running for him: which heat gets bumped, which grade sequence breaks… whether the slab yard has the right slabs in the right corner of the bay… whether the crane schedule can absorb one more reshuffle without blocking a truck slot at 6 AM… He doesn’t open a dashboard for this. He doesn’t need one, the dashboard already told him the order changed. What he needs is something that tells him what to do about it, and then does the boring 80% of it for him.
This is the gap most steel plants live in today. Not a data gap. A decision gap.
We didn’t under-invest. We under-connected.
Walk into almost any modern steel plant and you’ll find a respectable stack: ERP, MES, APS, a quality system, a maintenance system, maybe an IoT layer humming away on the rolling mill. None of this was wasted money. Visibility today is genuinely better than it was fifteen years ago.
But ask a simple question: when the export order gets pulled forward, does the system reschedule the heat, re-sequence the rolling campaign, notify logistics, and flag the slab-grade mismatch, all on its own, within governance limits you’ve already approved? Or does it wait for the planner’s phone to ring at 11:40 PM?
For nearly every plant I’ve worked with or studied, the honest answer is the second one. The systems observe beautifully. They decide almost nothing.
What “agentic” actually means, in plant language
Most of the industry’s digital effort has answered three questions: What happened? Why did it happen? What should we do? That’s where dashboards, root-cause tools, and recommendation engines live.
Agentic AI adds a fourth question that changes the economics entirely: who’s going to act on it, and how fast?
Take energy. Most plants already have a dashboard that shows the arc furnace pulling peak load right as the time-of-day tariff spikes, and everyone can see it. An agentic system doesn’t just show it. It checks whether the current heat can tolerate a brief power dip without compromising metallurgy, shifts auxiliary loads, and logs the saved cost against the heat number… before the shift supervisor has even finished his chai. The information was never the bottleneck. The 40 minutes between someone noticing and someone acting… that was the bottleneck. Agentic AI is mostly in the business of closing that 40 minutes.
Where this actually bites in a steel plant
Production planning instability. One late raw material truck… one unplanned BOF delay… one customer expediting an order… and the whole sequence downstream starts absorbing shocks manually. Ask yourself: the last time your schedule broke mid-shift, how many people had to get on a call to fix it, and how long did that call take?
Inventory that’s wrong in both directions. Finished coil sitting in the yard for a grade nobody’s ordering this month, while a different grade is on back-order and a customer is calling daily. It isn’t usually a forecasting failure… it’s that nobody is continuously rebalancing inventory against the latest signal. By the time a human re-runs the numbers, the signal has already changed again.
The yard, the crane, the truck. Logistics in most plants still runs on tribal knowledge and walkie-talkie coordination… somebody’s jugaad, frankly, holding the dispatch yard together. It works, until the person who’s good at it goes on leave.
Knowledge walking out the gate. There’s a Ramesh ji or an Anand ji on every shop floor… thirty-plus years in, who knows by feel (he understands grades by smelling…joke implied) which grades can share a rolling campaign and which can’t, without consulting a single SOP. That’s not in your MES. It’s not in anyone’s documentation. When he retires, does that judgment retire with him, or has someone actually captured the pattern behind his decisions before he walks out?
Quality risk that’s visible only in hindsight. Most quality systems are excellent historians and poor prophets… they tell you beautifully what went wrong after the coil is already rejected.
A practical roadmap, not a moonshot
The plants that get this right won’t attempt a big-bang transformation. They’ll move in stages, and each stage should answer one question before moving to the next:
- Foundation: Is your data from ERP, MES, and APS trustworthy enough that you’d act on it without double-checking?
- Decision intelligence: Are your planners and dispatchers getting recommendations they actually trust, or ones they routinely override?
- Semi-autonomous operations: Where would you let AI act first and tell you after, versus where do you still want sign-off first?
- Autonomous execution: Which decisions are repetitive and low-risk enough that a human checking them is pure overhead?
- Enterprise agentic manufacturing: Are your production, quality, logistics, and energy agents actually talking to each other, or just each reporting up to a human coordinator?
Most plants try to jump from stage 1 to stage 4 because a vendor pitch made it sound easy. It isn’t, and the plants that try usually end up with an expensive pilot that nobody on the shop floor trusts.
The questions worth asking before the next budget cycle
- If your best planner took two weeks of leave tomorrow, would schedule adherence survive it… or quietly fall apart?
- When something goes wrong at 2 AM, how many of the next ten actions are genuinely judgment calls, and how many are just someone executing a known playbook by hand?
- What did your last “alert” actually save you, if no one acted on it for forty minutes?
- Whose knowledge would you lose the most sleep over, if they resigned next month… and have you actually captured why they decide what they decide, not just what they decide?
- Are you measuring AI initiatives by dashboards built, or by decisions that no longer need a human in the loop?
The real question for leadership
The transformation isn’t a future event you’re deciding whether to greenlight. It’s already running quietly in the plants that are pulling ahead… not the biggest plants, not even always the lowest-cost ones, but the ones whose decision-making cycle is shrinking fastest.
The honest question isn’t whether AI belongs in steel manufacturing. It’s whether you’re using it to automate the tasks around the decision, or to actually shrink the time between something changing and something being done about it.
Because in the end, that gap… between the 11:40 PM phone call and the fix… is where margin quietly leaks out of every steel plant. Closing it isn’t a digitalization project. It’s the next competitive advantage.
Disclaimer: The views expressed in this article are personal in nature, for informational purposes only, and do not constitute professional, operational, or financial advice for specific manufacturing facilities.
#AgenticAI #SteelIndustry #ManufacturingAI #SmartManufacturing #Industry40 #DigitalTransformation #ProductionPlanning #OperationalExcellence #SupplyChainAI #ManufacturingIntelligence #SteelManufacturing #AIStrategy