IA on the job: can employees challenge a proposal without putting themselves at risk?

An AI-generated proposal can become difficult to challenge once a manager has endorsed it. The management question is specific: how can reporting an inconsistency become genuinely possible, beyond a general instruction to remain vigilant?

What the literature establishes

Amy Edmondson (1999, Administrative Science Quarterly) identifies an association between team psychological safety and learning behaviour. Psychological safety refers to a shared belief that the team is safe for interpersonal risk-taking, including asking questions or acknowledging mistakes. For the “IA on the job” programme, these findings suggest examining the social conditions for checking outputs, rather than technical skills alone.

Managerial openness beyond the message

James R. Detert and Ethan R. Burris (2007, Academy of Management Journal) show that managerial openness is associated with employees speaking up with suggestions intended to improve organisational functioning. Their findings support a mediating role for psychological safety in this relationship. Applied to AI as a working hypothesis, this mechanism suggests that an invitation to check outputs will remain fragile if objections are treated as signs of distrust.

The premature conclusion

One might conclude that a team whose members feel free to speak will necessarily be able to correct AI outputs. These studies do not establish that: being able to voice a concern means neither identifying a technical error nor having the evidence needed to demonstrate it. Freedom to challenge must therefore be distinguished from verification skills and approval authority. (our executive and employee training programmes)

What the evidence cannot promise

These studies do not address generative AI and do not experimentally demonstrate the effect of a managerial instruction on error detection. Applying them to current AI use must remain a hypothesis to be examined locally. Simply counting objections would not suffice either: their number also depends on the opportunities for checking and the difficulty of the cases.

A practical check in Neuchâtel

As part of “IA on the job”, a workshop in Neuchâtel could use AI-generated inspection instructions based on fictional cases from microtechnology, watchmaking and precision microelectronics, without confidential data or any impact on production. A subject-matter expert would insert documented inconsistencies, and the manager would explicitly state that participants are expected to challenge them before approval. The measure would be the proportion of these inconsistencies actually reported before approval, with a record of the response to each report. Comparing exercises of similar difficulty before and after this change in instructions would provide a local indicator, without independently establishing a causal effect. To go further: explore the AI on the Job training in Neuchâtel, or browse our executive and employee training programmes in Switzerland.

In pictures: AI on the Job in Neuchâtel

• ai on the job
• Neuchâtel
• research
• training neuchâtel