Software teams got copilots years ago. Controls engineers are getting a narrower, higher-stakes version: on-prem assistants that draft ladder, structured text, and HMI faces from specs, comments, and prior projects. Vendors and integrators are packaging these tools for air-gapped or tightly governed OT networks—because sending plant logic to a public model is a non-starter.
The industrial point is not autocomplete for fun. It is throughput under a skilled-engineer shortage, with change control that still owns every download to a PLC.
What copilots actually help with
- Boilerplate and variants — Motor starters, interlocks, and faceplates that follow plant standards.
- Review assist — Diff-style checks against naming rules and known anti-patterns.
- Knowledge capture — Turning tribal comments into searchable project context—without replacing RAG plant manuals on the floor.
What must stay human
Safety-rated logic, SIL arguments, and validated sequences are not “accept all.” Hallucinated tags and wrong fail-safe defaults are worse than slow engineering. Copilots that cannot cite the plant’s own standard library create audit debt. Agentic shop-floor actions are a different layer; this is engineering-time AI, not runtime control.
What to watch next
- Which OEMs ship copilots that stay fully on-prem with project-local models.
- Whether change-control workflows treat model output as draft-only by default.
- Measured engineering-hour savings on brownfield migrations versus greenfield toys.
Edge chips run inference on the line. PLC copilots try to write the line—under supervision.