Start with how a plant is actually orchestrated today, because it is rarely written down anywhere.

The scheduling system says which orders run. The maintenance system says which assets are due. Neither of them knows who can run line three alone at 3am. That knowledge is in the head of the shift supervisor, who holds an undocumented map of which of his people can do what, who is quick, who needs watching on the changeover, who has never actually done the isolation without help. He allocates work against that map every shift, correctly, and the map leaves the building when he does.

That map is the orchestration layer of the human side of your plant. It is a person’s memory. It is not backed up.

What the next five years actually do

Robots and physical AI are going to absorb routine execution, unevenly and slower than the conference talks suggest, but genuinely. Follow what that leaves behind.

The human portfolio does not shrink proportionally. It concentrates into exception handling: the abnormal state, the thing the design did not anticipate, the weak signal at 4am. In human-factors terms, that is knowledge-based cognition, which is the slowest and highest-error-rate mode people have. So the work that remains is the work humans are worst at, performed on higher-consequence equipment, by fewer people, with fewer peers nearby to catch a mistake.

Then the second-order effect, which is the one nobody plans for. Veterans are made by years of routine work. If juniors never do the routine work because a machine does it, the ladder that produces future experts is gone. You do not notice for a decade, and then you cannot hire your way out of it because the people you need were never manufactured.

Aviation ran this experiment already. Autopilot removed routine flying, and pilot skill decay plus startle response became the central safety problem of the industry. Not because the automation was bad. Because the automation was good, and the human role changed underneath it while everyone was still measuring the old thing.

The conclusion is uncomfortable and, I think, correct: the per-capita human-factors problem gets worse under automation, not better. Fewer people, each in the most error-prone cognitive mode, at higher stakes.

What orchestrating the human layer means

Not monitoring people. Three specific capabilities, all of which are currently either absent or living in one person’s memory.

Knowing the state, not the record. Not who holds a certificate, but who is currently able to perform this procedure on this machine, and when that is likely to have faded. Readiness is a state and it decays. A certificate is a photograph of a state on one date.

Routing work against the state. If the system knows that only two people on shift can perform the changeover unsupervised and one of them is on leave next week, that is a production decision, available before the week starts rather than after the line stops. This is the version of the human layer that a plant manager can actually use, because it is denominated in his units.

Capturing the second procedure. Every plant runs two versions of every procedure: the one in the binder, and the one that actually runs. The gap is where the improvisation lives, and the improvisation is mostly what keeps the plant running. Today that gap is invisible and it dies with the people who hold it. The only channel that can capture it is a channel the worker is already talking to, in the moment, for their own reasons.

None of this replaces a worker. It makes the worker’s judgement legible to the organisation, which is the opposite of replacing it. Right now that judgement is the most valuable and least documented asset in the building, and the standard plan for preserving it is hoping the person stays.

The five-year picture, stated plainly

In the morning production meeting, next to OEE and downtime and yesterday’s output, there is a number for readiness: by line, by shift, by machine. It moves. When it drops, somebody knows before the audit or the incident tells them. That is not a transformation programme. It is one number that does not exist today, and the reason it does not exist is that nobody has been measuring the human layer as a live state instead of a filing requirement.

It is not a fact about the people. It is a fact about the tools.

The same operator who “doesn’t use technology” is in three WhatsApp groups, one of which is the real communication channel of your plant. He watches repair videos on his phone at lunch. He books flights, moves money, argues about football in a group chat, and troubleshoots his own car from a forum thread. There is no literacy problem. There is a very specific refusal, and it is aimed at specific softwares.

Some of this is a hangover from real failures. But a lot of it was never tested at all. Push past “they’re not ready for it” and the honest answer is usually a guess dressed as a fact: someone assumed the floor couldn’t handle an inbox, so nobody built anything to find out.

That assumption does damage on its own.

It blames the floor for a rejection that never happened.

Where a tool was tried and abandoned, the reasons that came back, across roughly a hundred conversations with operators, supervisors and plant leadership, on about thirty sites in six countries, were never about age or education. They were about what the tool asked for and what it gave back.

• It went up and nothing came back. The classic version is the hazard report: you file it, and nothing observable happens, ever. After the third time, filing stops. It wasn’t a tool for the worker. It was a data-collection instrument pointed at the worker.

• It cost shift time and paid nothing back in the same shift. Production is measured hourly and rewarded. Anything that takes eight minutes and returns value next quarter loses, rationally, every time.

• It created exposure. A timestamp is evidence. Where investigations end at “human error,” a system that logs what you did and when is a system that logs who to blame. Workers read incentives for a living.

• It assumed desk-worker physics. A login. A password that expires. An email address the worker doesn’t have. Six sub-menus with gloves on. Any one of those was enough to kill it.

• It was in the wrong register. Not just the wrong national language, though that mattered too. The wrong voice, procedure prose written for an auditor, handed to someone under time pressure who needed two steps and got a page.

Nobody, in any of those hundred conversations, said “I don’t understand technology.” They said, five different ways, “this was not built for me.”

So what actually moves

The honest answer is that adoption on a floor is not won by training people to use software. It is won by making the first thirty seconds obviously worth their time, and then never breaking that deal.

Concretely, from what has actually worked and what has actually failed:

• No install, no login. A QR code on the machine or a link in the group the plant already uses. Opens in a browser. If there is a password reset flow, you have already lost a third of the floor.

• Answer before you ask. The first interaction should give the worker something they wanted: the two steps of the procedure they were unsure about, in their language, in ten seconds. A system that opens by testing them, instead of listening to them, is a system that has announced whose side it is on.

• One question, forty seconds. Not a course, not a module, not a quiz. A check on the decision points that carry consequence, short enough to do while the machine cycles.

• Anonymous and aggregate by default. This is the design decision that decides the whole thing. If a wrong answer can be traced to a name and used against that name, you will get theatre inside one quarter, and your data will be worse than no data. Aggregate tells you something more useful anyway: seven of nine people skipped verifying the isolation. That is a procedure problem, not nine personnel problems.

• The supervisor first, or not at all. A tool the supervisor does not champion dies on the floor regardless of who signed the contract. To the worker, the supervisor is the company. Any rollout plan that treats the supervisor as an end user rather than a co-owner is a plan to buy shelfware.

• Show them the change they caused. This is the only real engine, and it is the one almost nobody builds. When an operator’s answer reveals that step four is ambiguous, and the procedure gets corrected, tell the floor that the procedure changed because of what they said. Do it once and you have a system people talk to. Do it never and you have another data-collection instrument.

What does not work

Points, badges and leaderboards, on a floor where the tribal borders are already tense. Mandatory completion drives, which produce completion and nothing else. An app. Anything that starts with a two-hour induction on how to use the tool, because a tool that needs an induction has already failed the test.

Our CEO put it in a recent interview in a way I would not improve on: the biggest challenge is not the technology, it is the culture, and the answer is not to ask people to trust a new platform. It is to prove value fast enough that trust arrives on its own.

There is a measurable version of that. Instrument daily usage from week one of any pilot. If the floor is not using it at 60 to 70 percent within the pilot window, do not push expansion, do not add features, and do not blame the culture. Fix the deal you offered them.