Making it safe to report problems during AI adoption

Rehearse the moment a concern is raised

In a synthetic operations team, a staff member realizes that an AI-written customer update cited an obsolete service policy. They corrected the draft before it went out, but they hesitate to tell the manager because the rollout dashboard celebrates speed and the team has been praised for “using AI confidently.” A manager can make the next report easier or harder in a few seconds. “Why did you use the tool wrong?” places the person on trial. “Thank you for catching it; what source did the assistant use, what did we check, and who else may have the draft?” opens an evidence review.

The rehearsal card has two sides. The reporter states the artifact, potential consequence, evidence that showed the error, whether anything was sent, and immediate containment. The manager records acknowledgment, a non-punitive fact-finding question, owner, correction route, and follow-up date. If the incident involved real harm or a policy breach, the normal accountability and escalation process still applies. The card is not an amnesty rule; it prevents fear from obscuring facts needed to respond.

The American Psychological Association's 2024 Work in America survey describes psychological safety in terms of speaking up, sharing concerns, and discussing mistakes. It is survey evidence about workplace experience, not evidence that this particular rehearsal improves outcomes. An original workplace study on speaking up and joint problem solving examines the relationship between perceived safety and action in a healthcare setting. Its context differs from an office AI rollout; it supports the distinction between voice and a response that can solve a problem, not a universal effect size for knowledge workers.

Sources: [1] [2]

Give the report an observable route

The synthetic team chooses one route for AI-output concerns that everyone can find. The report asks for the task, affected document or system, what the AI contributed, source version, observed discrepancy, whether an output left the team, and what the reporter has already done. It does not ask the person to prove a root cause before anyone will listen. A separate owner handles triage, checks whether the same outdated policy reached other drafts, and records a disposition. The manager returns a short update so the reporter knows the issue was not swallowed by a form.

Do not measure safety by the number of reports alone. A rising count may mean more problems, greater willingness to speak, a new reporting channel, or all three. A falling count is not automatically success. Pair counts with examples of response quality: Was the concern acknowledged? Was the correction visible? Did the owner close the loop? Were people who challenged a senior AI-generated answer treated fairly? Review these questions with care for confidentiality and without turning the channel into a performance score for the reporter.

This is distinct from a broad employee-trust statement about data use and programme promises. The specific decision here is whether someone can surface a work error and see a fair, effective response. A team may trust an approved tool and still be afraid to tell a manager that a draft failed. The rehearsal tests the latter condition in a bounded, observable way.

Check the response, not just the invitation

After a synthetic drill, ask the reporter to explain what happened next. Can they name the owner and expected update? Did the manager distinguish an honest discovery from a reckless breach? Was the underlying document, prompt, or review step corrected? Repeat the exercise with a concern raised by a junior person about a senior person's draft. If the route works only when a manager is already sympathetic, the team has not yet built a dependable reporting practice.

This article does not diagnose anyone's mental health, prove an organizational culture change, or promise incident reduction. Two Prune's assessment product can observe task-bounded verification and communication behaviors in a simulation, with evidence and limitations for human interpretation. It does not judge an employee's psychological safety or make automated employment decisions.