Who is accountable when AI gets it wrong?
Practical boundaries for using AI at work: name the human decision-maker, check the evidence, and take responsibility for what happens to your people.
An AI-generated summary can look complete while leaving out the context that changes a decision. A draft can sound considerate while getting someone’s name, contribution, or circumstances wrong.
When that output affects a person, someone needs to own the decision. “The AI recommended it” does not explain why a leader accepted it or how the leader checked its work.
The framework is straightforward: Machines do. Managers decide. Leaders develop. Putting it into practice takes more than adding a human approval button.
Name the person responsible
Before using AI for a recurring task, decide who will review its output and act on problems. That person needs enough context, time, and authority to challenge it.
Be specific about the task. Preparing questions for a coaching conversation is different from evaluating an employee’s performance. Summarizing approved scheduling information is different from deciding whose hours should change.
A useful boundary describes what AI may prepare, what a person must verify, and which decisions remain with an authorized human.
Check the evidence behind the recommendation
A confident explanation is a starting point for review. Ask what information supports it and what information might be missing.
For example, an overtime flag can identify hours that need attention. It cannot establish why those hours occurred or whether the person needs support. That requires checking the schedule and talking with the people involved.
Keep observations separate from interpretations. “Three shifts ended later than scheduled” is an observation you can verify. “This employee manages time poorly” is a judgment that requires more context.
Make it possible to pause
If a process repeatedly gets names, facts, or instructions wrong, slow it down. Return to manual review while you identify the problem. People should know how to stop an automated action and whom to tell when something goes wrong.
Approval should be deliberate when an output affects an employee or customer. A busy reviewer missing a notification should not be mistaken for a considered decision.
Use tools approved for the work and follow the organization’s privacy rules. Remove unnecessary personal information. Preparing a conversation should not require uploading an employee’s private circumstances into an unapproved service.
Own the conversation and the follow-through
AI can help you rehearse difficult feedback, consider another explanation, or organize a development plan. You still need to listen to the person, correct your assumptions, and explain your decision.
When you make a mistake, address its effect on the person. Correcting the prompt may prevent a repeat, but it does not replace an apology or a repaired commitment.
Account for the time you recover
There is another decision to own: what happens to any capacity AI creates?
If a task becomes shorter, choose a specific use for that time. Prepare a coaching conversation. Recognize a contribution. Follow up on an employee’s goal. Then check whether that action actually happened.
The Field Guide’s human reinvestment prompt can help you review those choices. The point is to connect useful tools with a leadership commitment you can keep.