Artificial intelligence is coming into control rooms. Not because anyone decided it should, but because it is in every tool the people there already use. The question is not whether it comes in. It is how.
What an AI sees when it looks at a plant
Show a language model the readings of a waterworks and it sees 31TT0231.PV, FT0203, AI:1064. It does not know which is a temperature, which machine it sits on, or that one of them has been stuck for half a year. So it guesses, fluently and confidently. In an email a guess is an annoyance. In a plant that cleans a town's water, a confident guess is a danger.
One door, and what stands behind it
So we built a door. Any AI assistant can come through it: Claude, Copilot, a model running on the plant's own PC. It speaks MCP, the standard such assistants already use to reach tools. Behind it are 10 tools, and every one of them can only read. What any AI gets on the way in:
- Meaning, not just data. Every reading as the plant actually uses it: what it measures, which machine it is on, and the evidence for that.
- What we do not know, first. The readings we could not name are listed before anything else, so nobody fills the gap with a guess.
- Numbers with a source. Every figure carries an id that leads back to the exact readings it came from. The assistant is told it may not add a number of its own.
- No hands on the controls. Nothing behind the door can change the plant. A work order is a draft a person approves.
- Where the data comes from. Every answer says whether the plant is real or simulated, so no assistant can pass one off as the other.
It understands people the way they talk
81% against 20%questions in Norwegian or English read right, against the keywords we had before
An operator does not type keywords. They ask hva er galt med skrubber 2? A small model on the plant's own machine reads what they mean and picks a machine from the plant's own list, nothing else. The answer is computed from the data, never written by the model.
When two readers on the machine disagree about what was meant, it asks back, one click. On fresh questions it asked 16 times in 60, and offered the right reading 14 times. It still missed 6 of 11 wrong readings without asking, and we say so.
A guarantee, not a score
at most 5% wrong, 99% keptfrom 200 checks people on site answer, held in every test on 20 plants
Every vendor shows a confidence score. A score is a feeling with a decimal point. We let the plant choose the risk it will accept, and then prove it: someone who knows the plant checks a random sample, and the console guarantees that of the readings it keeps, at most that share are wrong. The ones it is unsure of, it refuses.
And where it is not good enough yet
When a sensor dies, the console can estimate it from the readings it moves with. We tested that where it is hardest: a real office building, the estimate learned in winter and tested in summer. It beat simply showing the last value on only 8% of readings. So the console now offers an estimate only where it has clearly earned it, and says which season it was learned in. We would rather show you that than sell you a sensor that fails in July.
A ladder, not a leap
Nobody should hand a plant to an AI in one step. Trust is built the way it is built with a new colleague: first they watch, then they point things out, then they suggest, and only much later do they act, within limits, with someone looking on. Each step is unlocked only by measured results from the step before.
| step | the AI | a person | today |
|---|
| 0 | reads and explains the plant | checks a sample | running |
| 1 | says when something does not add up | decides what to do | running |
| 2 | drafts work orders, answers questions | approves or discards | running |
| 3 | sends approved, low-risk work into other systems | approves every time | next |
| 4 | adjusts within limits set in advance | can override any time | years away, through the plant's own control system |
Why it matters
The plants that keep water clean and lights on are run by small teams with old systems and no data department. They should get the best of what AI can do, without betting the town on it. The understanding of a plant should belong to the plant: in a file it keeps, through a door any AI can use, with every number traceable and every gap admitted. That is the door we are building.