Industrial AI is moving past dashboards that only notify. Agentic systems—models wrapped in tools, memory, and workflows—are being piloted to draft production schedules, open maintenance tickets, rebalance WIP, and propose parameter changes when constraints shift.
The upside is speed: an agent can read MES events, quality signals, and inventory state faster than a human can tab through screens. The downside is familiar to every plant that automated the wrong thing: an unconstrained agent can thrash setpoints, flood CMMS with noise, or optimize a local KPI while starving the next station.
What works in early deployments
- Narrow jobs — One workflow (e.g., reschedule after a tool down) with clear inputs and outputs.
- Tool boundaries — Read-many, write-few APIs into MES/CMMS with role-based limits.
- Human gates — High-impact actions stay approve-to-execute until trust is earned on soft suggestions.
- Audit trails — Every proposal and acceptance logged for quality and liability review.
What still breaks programs
Hallucinated constraints, stale digital twins, and agents that cannot see real-time OT state. Cloud-only loops fail when the plant network partitions. Without master data discipline—routings, calendars, skill matrices—the agent optimizes fiction.
Agentic AI is also not a substitute for process engineering. It amplifies whatever planning logic and data quality already exist.
What to watch next
- Agents measured on schedule stability and expedite rate—not demo fluency.
- Standard patterns for tool permissioning across MES, QMS, and CMMS vendors.
- Whether multi-agent setups (planner + maintenance + quality) reduce handoff delay or multiply conflicts.
Shop-floor agents become industrial when they are boring: bounded tools, logged decisions, and KPIs that match how the plant actually makes money.