Will your people use it? Designing AI output for the people who act on it

The question "will our people use it?" is usually asked last, after the technology questions, and it decides more than any of them. An integration that is right but ignored does nothing. One that is slightly less clever but lands in front of the right person at the right moment changes a shift.

Adoption is not a training problem to be solved after the build. It is a design problem, settled by where the output appears, how long it is, how often it is right, and whether the people who have to act on it were in the room when it was designed.

The floor decides, not the meeting room

Head office can approve an integration; only the floor can adopt it. The supervisor on the afternoon shift decides in about a week whether a new alert is worth looking at, and that decision is rarely reversed. Once an alert is known as "the thing that keeps going off", nobody opens it, however accurate it later becomes.

The reason is simple. A floor runs on habits that survive peaks, exceptions and shift changes. Anything new has to fit those habits or it is dropped at the first busy hour. A good design starts from the habits and works backwards.

Exceptions and shift changes are the test

In a quiet hour, people will read anything. The test is the exception: a truck arrives without paperwork, two carriers turn up for one dock, a cold room starts to drift during a peak. At that moment the supervisor has thirty seconds and one screen, and an AI suggestion either helps in those thirty seconds or gets in the way.

Shift change is the other test. The person who understood the pilot goes home and the person who arrives was not in the meeting. If the output needs explaining, it stops being used at the first handover. Anything the floor is expected to act on has to make sense to someone seeing it for the first time, mid-shift, without a manual.

Where the output lands, how long it is, how often it is right

Three design choices decide most of adoption, and all three are made before any AI is involved.

• Land where people already look. A supervisor who lives in the warehouse system and a radio does not open a new screen for fun. Output has to appear on the screen that person already watches, in the order they already work in. A dashboard nobody opens is a common way to waste a good integration.
• Keep it short. One line that says what was noticed, why it matters and what the suggested action is. "Bay 4 has been empty for forty minutes while two trucks wait in the yard" is enough. The evidence can sit one click away for the person who wants it.
• Be right often enough. Not perfect, but right often enough that acting on it is faster than ignoring it. The pilot looks at this directly with the people who receive the output: how often did they act, how often did they dismiss, and what did the dismissed ones have in common? Each kind of dismissal becomes a change to the design or a new check.

The fourth choice is quiet. An integration that speaks less, but only when it has something worth saying, is trusted more than one that comments on everything. Silence is part of the output.

Involve supervisors from the first conversation

The people who will act on the output should be in the room from the first conversation, not shown the result at the end. They know which alerts would be welcome and which would be noise, which screen they actually look at, and what a suggestion has to say before they would act on it without checking.

They also know the exceptions. A supervisor can list, from memory, the ten strange things that happened last month. Those ten become the first checks the integration has to pass, and the supervisor who contributed them has a stake in whether it does.

Pilot together, then widen

A pilot with a small group, one shift or one dock, is where adoption is designed in practice. The group uses the integration on real work and tells us what to change: the wording, the timing, the threshold at which an alert fires, the place it appears. We change it, and they tell us again.

When the group stops asking for changes and starts noticing when the integration is quiet, it is ready to widen. The next group hears about it from the first, which is the way anything is adopted on a floor.

An AI integration is adopted when a supervisor would miss it if it stopped, and that is designed in, not trained in.

This is how redfly designs the output of every integration with warehouse and distribution companies: with your supervisors in the room from the first conversation, piloted together with the people who will act on it, and landing in the redfly Portal as the short, plain alerts and suggestions your people already look for.