Update — 15 July 2026. In the public beta the advisory plane is served from the cloud tier, hosted in-country. The site-local reasoning tier described below has moved to the roadmap alongside time-series ML — we were not happy with what small models could honestly deliver on site hardware, and we would rather ship it right than ship it first. The advisory boundary is unchanged: no tier, present or future, can command equipment.
There is a quiet question behind every conversation we have with a control room: what happens when the model is wrong? On a heavy-industry site the cost of a bad command is not a re-render — it is a pump dead-headed against a closed valve, a conveyor started under a buried load, or a feeder energised while someone is still on it. So we answered that question before we wrote the marketing, and we answered it in the architecture.
The rule is simple and we do not soften it: AI explains what the PLC is doing; it never takes control.
Where control actually lives
On a Consystence site, every control action travels the same deterministic path it always has. The PLC executes the logic. Hard interlocks live in IO modules and safety relays, sized and certified against the relevant functional-safety standards — IEC 61511 for the process layer, ISO 13849 for machinery. The field wiring still follows AS/NZS 3000. None of that moves into a language model, an inference server, or the cloud.
What the platform adds sits beside that path, not inside it. We read tags, alarms, and command history. We reason over them. We hand an operator a recommendation or an explanation. The operator — or the existing PLC logic — decides what to do with it. The AI tier has no write authority over equipment, and there is no configuration toggle that grants it any. That is not a default we ship locked; it is an absence we built.
Advisory is not one thing
“AI” on an industrial site is too blunt a word, so it is worth being precise about the tiers of the advisory plane, because each is advisory in a different way:
- Site-local reasoning (on the roadmap). A small model running on hardware at the site — no tools attached, the live plant state fetched and handed to it — can explain a sequence of events in plain language: discharge pressure is trending up, the downstream valve is throttling, this is expected behaviour, not a fault. It draws on the device-type context, and it keeps working when the link to the cloud is down.
- Cloud advisory models. Larger models answer harder questions through read-only tools over tags, alarms, and history — richer grounding and multi-step retrieval, with inference kept in-country. The tools can query the plant; none of them can touch it.
- Time-series ML (on the roadmap). Edge models will watch for the drift and signatures that precede failure — a bearing warming, a flow curve flattening — and raise a flag early. They will not trip the machine; the existing protection does that.
All three produce words and numbers for a human to act on. None of them produces a command for a contactor.
Provenance makes advice auditable
Advice is only trustworthy if you can see what it was built from. Every sample in our historian carries its origin — measured, inferred, imported, simulated, replayed, or workspace. When the model says “flow is low”, you can ask whether that flow was measured by an instrument, inferred from pump curves, or replayed from a recorded incident. “Is this real or derived?” becomes a query, not a debate in the control room. An advisory system that cannot tell you the difference is just a confident guess.
Why we treat this as non-negotiable
Mining and process safety culture is built on a hard separation between protection and information, and for good reason. The people who run these plants have watched “smart” systems erode that line and have learned to distrust anything that blurs it. We think that distrust is correct. So rather than ask operators to trust that our AI will behave, we removed its ability to misbehave: the control path is deterministic, the interlocks are in hardware, and the AI is on the advisory side of a wall it cannot reach across.
On-device anomaly detection is on our roadmap precisely because it does not need a round trip to the cloud to be useful — it will keep flagging drift when the link drops — and for the strictest sites the AI tier can be disabled entirely. The plant keeps running either way, because the plant was never depending on the AI to run it.
That is what “advisory by design” means. Not a promise. A property of the system.