In this article
If your team can generate documents faster but struggles to make decisions, producing more documents will not solve the problem. Start with the management question: what should improve for the people who depend on this work?
Managing AI begins with a clear assignment
A good assignment states the outcome, the relevant information, the constraints and the person who will judge the result. Those are familiar leadership disciplines.
Ethan Mollick's essay on management and AI explores this connection. The practical implication is to make your expectations inspectable instead of assuming that an impressive response reflects the right priorities.
In a hypothetical regional operations team, “Use AI to improve reporting” is too broad. “Prepare a draft exception summary from the approved weekly figures, preserving uncertainty and requiring analyst verification” describes an actual job.
The second assignment also makes a failed result easier to recognise. A report that invents a cause for a variance has crossed the boundary, even if the narrative is fluent.
Separate delegation from accountability
Delegating preparation does not delegate responsibility for the decision. Identify the owner of the input, the reviewer of the output and the person authorised to act.
These roles may belong to one person on a small task. On a larger workflow, separate them explicitly so a missing review does not disappear between teams.
| Role | Practical responsibility |
|---|---|
| Task owner | Defines the problem and acceptable outcome |
| Information owner | Confirms approved sources and access |
| Reviewer | Checks claims, omissions and consequences |
| Decision owner | Approves the action and records the reason |
Do not write “the AI decided” in place of a decision owner. A tool's recommendation is an input to a process, not a person accepting consequences.
Review the difficult part yourself
Focus review on the judgement the task requires. In a status update, that may be distinguishing a delay from an unresolved risk. In a staffing discussion, it may involve fairness, context and information the tool should never receive.
Ask three questions before accepting the work:
- Which statements are supported directly by the source material?
- Which statements are interpretations or proposals?
- What important exception would change the decision?
In the hypothetical report, sales may have fallen while a location was closed for refurbishment. The numerical decline alone is insufficient to judge the team's performance. A manager who knows the operating context must supply that judgement.
NIST's AI Risk Management Framework provides a reference for managing AI risk. A team-level review should translate broad principles into named responsibilities and practical checks.
Give feedback that develops the workflow
When an output fails, identify whether the cause was a poor assignment, unsuitable input, model behaviour or weak review. Each needs a different response.
If the brief omitted the closure, add the relevant context. If the tool ignored an explicit restriction, tighten the process or choose a different approach. If the reviewer missed an unsupported claim, improve the review method rather than merely rewriting the prompt.
Keep a short failure log with the issue, consequence and corrective action. Avoid collecting mistakes to blame employees for experimenting in good faith; use the record to make the approved workflow safer and more useful.
Protect the team's ability to disagree
Invite staff to challenge both the output and the assumption that AI is suitable. A polished answer should not end a discussion that needs local knowledge.
In a meeting, ask a colleague to identify what the draft missed before asking whether everyone agrees. Give people a clear route to stop a workflow when confidential information, biased treatment or unreliable evidence appears.
Managers should also explain how the tool changes responsibilities. If staff believe that checking AI output counts as resistance, they may learn to approve it quietly. Make thoughtful review part of the work, with time allocated for it.
Run an assignment review with the team
Before introducing a recurring AI-supported task, ask three colleagues to review the assignment from different positions: the person providing the information, the person preparing the draft and the person using it. Each should describe what they need to trust the handoff.
In the hypothetical reporting team, the information owner might require the reporting period and source version to remain visible. The preparer might need a clear rule for missing figures. The recipient might need the requested decision separated from background explanation.
Put these requirements in the assignment, then test it on an awkward example. If the team cannot agree on an acceptable result, the management work is not finished. A more elaborate prompt is unlikely to resolve that disagreement.
After the test, ask whether the review effort fits the role and available time. Do not quietly make a junior colleague responsible for validating an analysis they lack the expertise or authority to assess. Give the review to someone qualified, narrow the task or choose another method.
Measure the outcome people actually need
Judge the workflow by verified usefulness: fewer missing actions, clearer decisions, appropriate turnaround and acceptable review effort. Prompt counts and generated pages are weak substitutes.
Choose one recurring assignment and document its owner, acceptance criteria and review point. Then observe where the process helps and where it adds work.
For further context, visit PiOfAI, the AI procedure guide and the guide to trust and credibility as a leader. Better AI use begins with management responsibilities that remain visible.
AI-assisted researched guidance from PeakBelt. Hypothetical examples are labelled in the text. See our editorial approach.