AI Is Taking Over Steel's Most Dangerous Jobs
Steelmakers are moving AI out of dashboards and into machines that inspect, grind, and haul on the shop floor, a shift POSCO frames as physical AI. The South Korean producer deploys Boston Dynamics' quadruped Spot robots to inspect blast furnace areas, detect gas leaks, and flag temperature anomalies in zones unsafe for human crews.
How It Works
The models process camera feeds, vibration signatures, and thermal data in real time, and instead of lighting up a control room screen, the output steers a robot's path or a machine's next adjustment. Computer vision grades surface defects at line speed with consistent severity calls. Generative AI systems ingest maintenance manuals and work-order history, then answer technicians' questions mid-repair.
Where It Is Already Live
JFE Steel's Chita Works runs a robotic system for grinding small seamless pipes that locates the part, detects defects, and optimises its own movements, lifting processing speed by 60%.
Nippon Steel is building autonomous heavy-duty transporters with TIER IV to move material through its Nagoya plant.
U.S. Steel developed MineMind with Google Cloud, its first generative AI application, guiding maintenance crews through repairs at its iron ore operations.
Tata Steel's Kalinganagar plant runs more than 260 AI algorithms covering charge planning, furnace control, vision-based quality checks, and predictive maintenance.
What It Costs, What It Replaces
The AI-in-steel market stood at $9.84 billion in 2025 and is projected to reach $35 billion by 2035. The barriers are capital cost and legacy plants whose ageing control systems cannot feed clean data. What the systems replace is inspection rounds in hazardous zones and inspector-to-inspector variability in defect grading, not metallurgists: ArcelorMittal's digital twin deployments across European plants cut energy use 12% and unplanned downtime 30%, with operators still making the calls.
Read The Case Study
Primetals' AI lead explains where plant-floor AI is heading next, including AI-accelerated plant simulation.