Book a call
AI Readiness · Culture · Operating System

Before AI Comes to the Factory Floor, Fix the Human Operating System

AI can surface the problem. People still have to own it, raise it and solve it.

By Nash Krishnan · September 29, 2026 · 7 min read

A veteran machine operator and a younger colleague with a tablet talking beside an automated production cell
AI-generated illustrative image. No actual plant or employee is depicted.

The biggest barrier to AI adoption in manufacturing may not be the technology. It may be the operating system the technology lands on: who owns a problem, how it gets raised, how quickly someone responds, and whether people trust what happens after they speak up.

Picture an operator who has run the same process for eighteen years. He knows the sound the machine makes before a jam. He knows which material lot runs a little wet and which setting to nudge when the humidity changes. None of that is written down. It lives in his hands and his head.

One morning a new screen appears beside his station. Someone from outside the plant says it will "predict problems." Nobody explains what it will do with the data it collects, or what it means for his job.

He does not ask whether the algorithm is accurate. He asks the question almost every employee asks, out loud or silently: Is this here to help me, to monitor me, or to replace me?

If that question goes unanswered, the project is already in trouble, however good the software is.

What the research says about how people respond to AI

A growing body of peer reviewed research shows that how people respond to AI depends heavily on what they believe about it, not only on what it produces.

In a 2023 study in JAMA Internal Medicine, a panel of health professionals compared physician answers with chatbot answers to real patient questions and preferred the chatbot answers in most comparisons, rating them higher for both quality and empathy (Ayers et al., 2023). On output alone, AI can perform very well.

Yet the label matters. Researchers writing in PNAS found that AI generated responses made people feel more heard than human responses, but that advantage shrank when people were told the response came from AI. Negative attitudes toward AI explained much of the drop (Yin, Jia & Wakslak, 2024).

A larger set of experiments published in Nature Human Behaviour, nine studies with more than 6,000 participants, found the same pattern. Identical supportive messages were rated as more caring when people believed a human wrote them, and many participants were willing to wait for a human reply rather than receive an instant AI one (Rubin et al., 2025).

People are not simply against algorithms, either. Work in Organizational Behavior and Human Decision Processes showed that people often prefer algorithmic advice to human advice for estimates and forecasts (Logg, Minson & Moore, 2019). The resistance is less about the math and more about meaning: what the technology says about my role, my judgement and my future.

Inside workplaces, that meaning shows up as behaviour. A three wave study of 402 employees in the Journal of Innovation & Knowledge linked AI related job insecurity with knowledge hiding, partly through lower psychological safety. Employees who felt more capable of learning the technology were buffered from that effect (Kim & Kim, 2024). Experiments reported in the Journal of Business Research found that when AI related job insecurity was high, the path from knowledge sharing to employee thriving broke down. The authors recommend transparent communication about AI adoption and real upskilling (Nguyen, Ferm & Ngo, 2026).

None of these studies was conducted in an Ontario factory, and none proves what will happen in yours. Together they point to one practical conclusion: the human response to AI is part of the system, not an afterthought.

What this looks like inside a plant

In a factory, knowledge hiding does not look dramatic. It looks like an experienced operator who stops mentioning the workaround. A lead hand who lets the new system make the call rather than challenge it. A supervisor who waits to see whether the project survives before investing in it.

AI does not remove those behaviours. It inherits them.

If accountability is weak, the AI will flag a problem that nobody owns. If employees do not feel safe raising problems, they will not report the cases where the recommendation is wrong. If supervisors do not have a consistent management rhythm, alerts pile up with no routine to act on them. If improvement depends on a few experienced people, the knowledge the model most needs is exactly the knowledge most at risk of walking out the door.

In each case, AI simply gets layered on top of a weak operating system. The dashboard gets better. The plant does not.

Sometimes the first tool does not need to be AI at all. In one turnaround, a shared spreadsheet between two plants removed most of the hearsay. It only worked because the roles, routines and quality checks around it changed too.

The human operating system comes first

By operating system I mean the daily way a plant runs: who owns what, how problems move, and how people learn. It is not software. It is a handful of habits that decide whether any technology pays off.

Clear ownership. Every metric, alert and recurring problem has a named owner at the right level, and people know who decides what.

Daily routines. Short, consistent tiered huddles where the team looks at yesterday, today and what is in the way. This is where an AI insight gets turned into an action instead of a notification.

Escalation that works. A simple, visible path for raising an issue, with a response time people can count on. If raising a problem leads to silence, people stop raising problems.

Structured problem solving. A shared method for getting from symptom to root cause, so a flagged issue is solved rather than patched.

Trust. An honest answer to the question every employee is asking about AI, and a consistent response when someone challenges an output or admits a mistake. Trust is built by what leaders do the tenth time, not what they say the first time.

When these are in place, AI has something to plug into. Operators help the model get better because they see how it helps them. Supervisors use alerts in the huddle. Knowledge moves from a few people into the system and the team.

Five questions before your next AI pilot

  1. Which problem are we trying to solve, and who owns it today?
  2. Where in our daily routine will the AI output be reviewed and acted on?
  3. How will someone raise a concern or reject a recommendation, and how fast will they hear back?
  4. What have we honestly told the affected people about their roles, and what support will they get to learn?
  5. Whose knowledge are we relying on, and how are we capturing it with them rather than from them?

If those questions are hard to answer, that is useful information. It tells you where to start, and it is usually not the software.

Build the human operating system first. Then give it better tools.

If you are weighing an AI or automation project and want a clear view of whether your floor is ready for it, OpsAlign starts with a walk of the floor, not a sales deck. Book a 30 minute call or see our services.

Common questions

What is the human operating system in a factory?

It is how work actually gets decided and done: who owns which decisions, the daily routines that surface problems, how issues are escalated and whether people trust that raising something will lead to action.

Why do AI projects fail in manufacturing plants?

Many fail because the tool is layered on top of weak routines. If operator knowledge is not captured, problems cannot be raised without blame and supervisors are not leading change, AI simply automates the confusion.

Where should a manufacturer start before buying AI?

Walk the floor first. Check whether people understand where the business is going, whether problems are surfaced early, whether experienced operators' knowledge is written down and whether employees help design changes to their own work.

Sources

← Back to Insights