Manufacturers should give their line operators explicit authority to pause a flawed AI instruction, with a named flesh-and-blood person responsible for resolving a problem and protection against punishment for any resulting delay.

Otherwise, asking employees to exercise judgment while rewarding uninterrupted throughput creates a conflict that management has left them to solve alone.

See also: Smart Industry's annual State of Initiative Survey is open!

Smart Industry's annual State of Initiative Survey asks readers about digital transformation as the industry looks toward 2027. One useful question for every plant preparing its next investment is whether an operator can challenge the software in practice.

The question belongs alongside Smart Industry's coverage of AI-enabled manufacturing software. In its December 2025 announcement of its elastic manufacturing execution system offerings, Rockwell Automation described embedded AI, analytics, and connected worker technology across operations.

As manufacturers evaluate such capabilities, I recommend treating the procedure for interrupting a questionable recommendation as part of the investment itself.

Consider a hypothetical assembly cell. An AI assistant produces setup instructions using an obsolete process sheet. The operator notices that its cited revision differs from the controlled document at the station. The shift is behind schedule, and the supervisor has been asking everyone to recover lost output. Following the recommendation would be easy. Raising the discrepancy means explaining another delay.

The operator should not have to win an argument about the entire AI program to resolve that mismatch. A specific, documented discrepancy should be enough to pause the affected instruction and obtain an accountable decision.

Give the pause a clear owner

Manufacturing already has a useful precedent. Toyota's description of jidoka explains that equipment can stop when it detects an abnormality, and operators can stop the line themselves. The purpose includes identifying problems and preventing defects from moving onward. Apply that principle to AI recommendations by defining which conflicts require intervention and who resolves them.

I recommend a short written rule: When an AI instruction conflicts with a controlled procedure, a quality hold, or an observed operating condition, the operator may pause that instruction and notify the designated supervisor or quality lead.

See also: AI investments rise, but so does a healthy dose of skepticism

The rule should identify the approved fallback and the person authorized to release the work. Any physical shutdown must follow the plant's established stopping procedures.

The National Institute of Standards and Technology's guidance on AI governance recommends defined paths for escalating risk concerns. Operators need clear authority over the affected step, along with escalation when the consequences extend beyond it.

They should also know what to do if the designated decision-maker does not answer. Name a backup and a response time appropriate to the process before deployment.

The NIST AI Risk Management Framework calls for clear responsibilities and communication, defined human oversight, and post-deployment mechanisms for appeal and override. My proposed plant rule makes those principles concrete at the point where someone must decide whether to act on an instruction.

NIST's guidance on managing AI incidents calls for documenting how organizations track and respond to errors. At the station, capture the recommendation, its source or revision where available, the discrepancy, the affected work, and the resolution.

Use the existing quality or production record when practical. The supervisor should own moving that record to closure, including telling the operator what changed.

The U.S. Occupational Safety and Health Administration's guidance on worker participation warns that incentives can discourage reporting of safety concerns. Apply the same attention to incentives when designing an AI escalation policy.

See also: The ‘AI-native’ mirage: Why pragmatic wrapping beats marketing washing

Do not reduce an operator's rating or incentive merely because a good-faith pause, made under the rule, slowed output. Evaluate whether the person followed the process and reported the available evidence. A concern can deserve investigation even when the instruction ultimately proves correct.

Rehearse the disagreement

The behavioral evidence gives manufacturers a reason to take this seriously. In her study of 51 teams in a manufacturing company, author Amy Edmondson found that psychological safety was associated with learning behavior.

Psychological safety concerns whether people believe they can take interpersonal risks within their team. Challenging a recommendation backed by a supervisor is precisely the kind of interaction leaders should rehearse.

Run a short exercise away from live production before introducing the AI workflow. Give an operator a realistic instruction containing a discrepancy, such as the obsolete revision in the hypothetical assembly example.

Ask the operator to use the actual reporting route. Have the supervisor respond, consult the approved source, and document a decision. Then test the backup route with the first contact unavailable.

NIST's guidance on testing human oversight recommends including relevant participants and using scenarios similar to deployment conditions. For a plant, that means rehearsing more than recognition of the wrong answer. Test whether the person receiving the concern knows what authority they hold and can reach the information needed to resolve it.

See also: Manufacturing’s hardest hire is the person running the automation

Include a case in which the AI instruction is correct and the operator's concern is mistaken. The supervisor should still acknowledge the report and explain the evidence. That response demonstrates how the rule works when speaking up does not uncover a defect. Ask participants afterward where they hesitated and what would make the same conversation harder during a busy shift.

Measure the response to a challenge

OSHA's program evaluation guidance includes response time and timely corrective action among useful safety measures. Adapt that approach to AI: Track time from an operator's report to a verified correction or an explained decision to proceed. Separate waiting time from investigation time. Review unresolved cases at shift handover, with an owner for each.

Track avoidable rework linked to the affected workflow, using the plant's established quality definitions and a consistent measure such as rework hours per batch. Compare similar products and operating conditions. Count time spent checking and correcting recommendations alongside any time saved. Do not assign a dollar value to a prevented defect unless the evidence supports it.

See also: Relay research finds critical gaps in factory-floor information-sharing, leading to failures

Finally, ask operators anonymously whether they know how to pause a questionable instruction, expect a timely response, and believe a good-faith report will affect their standing.

Compare those answers across shifts and follow up on specific obstacles. Treat a change in reporting volume as something to investigate. More reports might reflect better reporting or a less reliable system; the count alone cannot distinguish them.

Before expanding an AI workflow, have an operator demonstrate the pause and a supervisor demonstrate the resolution. Then inspect what happened to the last real concern. If the record is still open, the explanation never reached the worker, or the pause became a performance penalty, fix that management practice before extending the system to another line.