Uptime-critical operations, proprietary designs, and plant data that shouldn't traverse the public internet. The economics favor the edge, as long as it's governed.
Manufacturing has the strongest case for local AI and the least tolerance for getting it wrong. The valuable data (process parameters, maintenance history, proprietary designs) is exactly the data that should never traverse the public internet. And the plant floor cannot take a dependency on someone else's uptime.
The quiet crisis is knowledge retention: decades of process expertise walking out the door in retirements, documented, if at all, in formats nobody searches. A private model trained to retrieve from your own maintenance logs and runbooks turns tribal knowledge into an asset that stays when people leave.
Edge economics are favorable and getting more so: inference on plant-local hardware is a fixed cost immune to token pricing, latency is deterministic, and an internet outage does not idle the line. The requirement is governance: hard OT/IT boundaries and stop conditions on anything automated that touches production.
Local inference at the edge and on-prem for design, maintenance, and process knowledge. IP-sensitive data stays on your network.
Hard network boundaries between OT and IT, least-privilege tool access, and stop conditions on any automated loop touching production.
An edge/colocation TCO model, then a private stack that turns tribal plant knowledge into a governed, queryable asset.
Models on plant-local hardware: deterministic latency, no internet dependency, IP on your network.
AI systems live on the IT side with explicit, audited crossings; nothing reaches into control systems uninvited.
Retrieval over maintenance logs, runbooks, and design history: the plant's memory, queryable.
Any automated loop touching production has hard limits and a human abort by design.
Edge versus colo versus cloud modeled on your duty cycle, not a vendor's slide.
No, and for IP-sensitive process data it shouldn't. Modern small models run on modest plant-local hardware, and the economics usually beat metered cloud pricing at industrial volumes. The TCO calculator gives a first pass with your numbers.
Retrieval-first: index the maintenance logs, shift notes, and runbooks you already have, then interview-to-document the gaps. The model's job is finding and synthesizing your knowledge, not inventing answers, and provenance is part of the design.
Only through governed, audited boundaries with stop conditions and human abort, and for most plants the answer for automated write-access is simply no. Advisory AI at the edge captures most of the value at a fraction of the risk.