The colour that means you are supervising

There is a moment in an AI-assisted job when a person stops doing their work and starts checking a tool’s. It happens quietly. One screen you are the machinist; the next you are reviewing something a model produced — and nothing on the screen has told you that anything changed.

In consumer software that costs you a bad paragraph. Standing at a stopped machine it costs a shift. So when I built the assisted-repair flow in Strato, the first thing I worried about was not whether the agent would be right. It was whether the operator would know which of those two people they had just become.

01

The operator needs to know which job they are doing

Strato reserves one violet — a single hue, spent on nothing else — for anything the agent touches. Not as branding for an AI feature. As a boundary the eye can find without looking for it.

The palette had already committed its loud colours: orange is marking, red is alarm. The violet was not what happened to be left. It was picked because nothing on a machine is that colour — no lamp, no guard, no warning plate. When it turns up, you are no longer reading the plant. You are reading something reasoning about the plant.

It lands differently depending on who is standing there. An operator reads it as a change in posture: from here you are supervising rather than acting, and what follows should be weighed rather than obeyed. Someone deploying the system — or selling it — reads the same colour as a map of exactly where the AI starts and stops, which turns a vague capability into something they can point at and explain.

The signal only survives if it is never borrowed. The first time that violet is spent on a nice highlight or a secondary button, it stops meaning anything, and the operator loses the one cue that told them their role had shifted.

02

The manager asks. The operator is never told.

Strato’s console opens with a free-text box — ask anything about your fleet. Which looks like a contradiction, because a chat window is exactly the register I have been arguing against.

It belongs to the manager and to service. It never reaches the operator, and that is not a permission someone forgot to grant.

A manager spends the day with questions and very few immediate answers. A machine stops and all it offers for itself is an alarm code. The agent can suggest what might be behind it and read across the rest of the fleet — and it says, every time, that it is retrieving data for display, not processing it, and that it cannot act on the plant. It is a closer look at problems that previously had only an alarm to speak for them.

The operator meets the same intelligence from the other side. Nothing instructs them. They can reach through the agent for how-to-solve material — procedures, recurring patterns, what was done last time — but they are the one going to get it, when they decide they need it. The agent holds the knowledge. The operator does the work.

That distinction decides what a plant is left with after a few years. An operator who was walked through repairs by a chat window remembers what the screen said, more or less, and takes that memory with them when they leave. An operator who went and looked something up has used a procedure that is still written down, still correct, and still there for the next person.

03

The step people skip

The guided repair will not move past its first step until the operator confirms the spare parts are actually in their hands. On a six-step flow, in front of a machine that is losing money, that is a speed bump — and people are very good at getting over speed bumps.

They click through and read later. Under pressure that is often the right instinct; it genuinely is faster. The confirmation is there because something further down the line depends on a person saying what really went where. Put an inventory module or an MES behind it and that tick is the only moment a part is recorded as used, or as needing to be bought again. Skip it and the system’s picture of its own stock drifts away from the shelf.

None of which makes the friction good. If that step is genuinely in the way, the answer is not to quietly delete it:

Have the system check stock itself. If the part is not there, let it open the purchase ticket and move the maintenance to the day the component actually arrives. That is automation — the system carrying the whole consequence of a missing part, rather than the operator carrying it silently.

Taking the tick away and changing nothing else is not automation. It removes the only evidence the part was ever used.

04

Somebody has to be able to answer for it

Picture following a guided procedure, finding out afterwards that it was wrong, and having no way to retrace what the tool actually did. That is the situation the agent’s activity feed exists to prevent.

Strato’s agent writes down what it has done, timestamped, in ordinary language: what it diagnosed, what it proposed and to whom, what it ordered, what it closed without opening a ticket. Its own predictions included, set against what happened — predicted twelve days ago from charge-hold decay; the battery failed this morning at 07:42.

This usually gets filed under transparency, which makes it sound like a courtesy. An agent narrating its own work is not personality — it is close to the opposite. It is a standing admission that it is a tool, that it is not doing the job in your place, and that the person reading is still the one answerable for how this ends. It is a form of respect toward the operator, and toward whoever runs the line.

A line running partly on something nobody can inspect is not a quiet efficiency gain. It is a bill that has not arrived yet.

05

What the manager is actually buying

The console shows the share of requests the agent closed without a ticket ever being opened. An AI reporting its own success rate to the person deciding whether to trust it should make anyone suspicious.

It reads differently once you know who it is for. It is not the agent grading its own accuracy — it is there for the manager who signed for the thing, and the cost it is measured against is very concrete. Every ticket that gets opened waits for an answer, turns into a thread, and grows more serious every hour the machine stands still.

It is also where the boundary sits. Machine failures, stopped lines, anything holding up production — those still want a human to look and a human to confirm it is fixed. The agent’s job is the quick, recoverable repair and the small problem caught before it becomes a large one. It is not standing in for the manufacturer’s service team or a specialist maintainer, and that percentage is not claiming it is. It is proof of investment, and nothing more.

06

What it was before, and still is

The first version of Strato was a monitoring tool. Charts and numbers. If it were ever sold, that would still be the entry tier — and at its core it is a tool, not an automation workflow.

Worth saying plainly, because the pull with anything AI-assisted is to let the assistance become the product. It should not. The assistance is a layer a plant might want, marked in a colour that says where it begins and ends, sitting on top of something that works without it.

What makes it worth having is not that it decides anything. It is that anyone can make it fit their own production line.

Strato is a self-initiated concept. Its plant, fleet and data are simulated — held consistent in one model, but invented. Figures quoted here show how the interface presents information; none is a measured result from a deployment.