Key takeaways
- Treat adoption as part of the project scope and budget, not a training task at the end.
- Put operators and technicians from every affected shift on the project team from use case selection onward.
- Document the purpose, data use and limits of each system, involve unions early, and keep the commitments.
- Build trust through advisory mode, open sharing of results and errors, and a respected path for overrides.
Adoption is the project
An AI system that works technically but is not used has produced nothing. On the plant floor, adoption depends on whether the people closest to the process believe the tool helps them, whether it fits the way work actually gets done across all shifts, and whether they trust it enough to act on it. These are not soft concerns to address after go-live. They determine whether the business case is realized.
Most resistance in manufacturing is rational. Experienced operators have seen systems arrive with promises and leave with nothing changed. Technicians know that a false alarm at 2 a.m. costs them sleep. Workers reasonably wonder whether a new system is meant to watch them or replace them. Address these concerns directly and early.
Involve the floor from the start
Bring operators, technicians and supervisors into the project during use case selection, not at training time. They know which problems matter most, which data is unreliable, which workarounds exist, and how a new step will fit into a shift. Practical ways to do this:
- Include at least one operator or technician from each affected shift on the project team, with time allocated for the work.
- Walk the process with them before designing the solution, and ask what would make their job easier.
- Let them help define success criteria, including what counts as a useful alert.
- Use their feedback during shadow mode to tune thresholds and wording.
People who helped shape a tool tend to defend it to their peers. People who had it handed to them tend to look for its faults.
Be clear about purpose and data use
State plainly what the system is for and what it is not for. If a vision system watches an assembly station to catch missed steps, say whether its data will be used to evaluate individuals. If a monitoring system tracks machine states, say whether it will be used for staffing decisions. Ambiguity breeds the worst assumptions.
Write down the purpose, the data collected, who can see it, how long it is kept and what it will not be used for, and share that document with everyone affected. Keep the commitments. A single breach, such as using safety camera footage for discipline after promising otherwise, can end trust in every future project.
Working with unions
In represented workplaces, new technology that affects job content, workload, monitoring or staffing may be a mandatory subject of bargaining, depending on the collective agreement and applicable labor law. Involve labor relations and union leadership early, before a pilot starts, rather than presenting a finished decision.
- Share the problem you are trying to solve and why it matters to the business and to workers, for example fewer emergency call-ins or less repetitive inspection.
- Explain what data will be collected and how it will and will not be used.
- Discuss effects on job content, training needs and any changes to classifications.
- Consider a joint committee that reviews the pilot and its results.
Get legal and labor relations advice on obligations specific to your agreement and jurisdiction.
Training that sticks
Training for AI tools should focus on the job, not the technology. Operators need to know what the tool tells them, what they should do about it, what to do when they disagree with it, and how to report problems. A short explanation of how the model works helps build appropriate trust, especially its limits: what it was trained on, what it cannot see, and when it is likely to be wrong.
- Train on the line, with real examples, during paid time, on every shift including nights and weekends.
- Pair training with updated work instructions so the new step is part of standard work.
- Identify champions on each shift who can answer questions and relay feedback.
- Refresh training when the model, thresholds or workflow change.
Building and keeping trust
Trust grows from a track record. Start in advisory mode, where people can compare the tool's suggestions with their own judgment. Share results openly, including mistakes. When an alert catches a real problem, tell the story at the shift meeting. When the system gets something wrong, explain what happened and what changed.
Give people a clear way to disagree. If an operator overrides a recommendation, record the reason and review overrides regularly. Some will reveal model errors; others will reveal training gaps. Both are useful. Never punish someone for a reasonable override, and never let the system become an excuse for removing judgment from people who are accountable for safety and quality.
Supervisors and middle management
Supervisors decide what actually happens on a shift. If they see the tool as extra work imposed from above, it will quietly fall out of use. Make sure they understand the purpose, have a say in the workflow, and see the results in the measures they are judged on. Add the tool's outputs to existing routines such as shift start meetings, tier boards and maintenance planning, rather than creating a separate screen to check. Track usage by shift and line, and follow up where it drops.
Measuring adoption
Decide in advance how you will tell whether the tool is being used, and review the numbers as regularly as you review output and scrap. Useful measures include:
- Share of alerts or recommendations that were reviewed within the agreed time
- Share acted on, and share overridden with a recorded reason
- Active users by shift, line and role
- Time from alert to action
- Feedback submitted, and how quickly it was addressed
Look for patterns rather than single numbers. If night shift reviews far fewer alerts than day shift, the cause may be staffing, training or a supervisor who was never brought in. If overrides cluster on one product, the model may be weak there. Short, regular conversations with users, such as a few minutes at a shift huddle, usually explain the numbers faster than any report. Treat a drop in adoption as a signal to investigate, not a reason to blame the people on the floor.
Related use cases
Workforce & training
Workforce training & skills (AI/XR)
AI, VR and AR tools train operators and technicians faster, track skills and deliver guidance at the point of work.
Operations & production
Manual assembly analytics & operator guidance
Cameras and sensors guide operators step by step and verify each manual assembly step, catching errors before the part moves on.
Safety, health & sustainability
Worker safety & EHS vision analytics
Computer vision on existing cameras spots unsafe conditions such as PPE gaps, pedestrian and forklift conflicts and zone intrusions.
Maintenance & reliability
GenAI knowledge assistants & copilots
Chat-based assistants answer maintenance and frontline questions from manuals, work orders and procedures, with sources cited.