“You get an agent! You get an agent! Everybody gets a personal AI agent!”
Everyone is about to be a manager without training.
We’re giving people something that can research, draft, use systems, and take action on their behalf. Then teaching them how to write a better prompt. Useful skill, but incomplete training.
Asking an agent to do something and knowing how to delegate work are not the same thing. As personal agents move into everyday enterprise work, that distinction becomes a practical adoption problem … not just a philosophical argument about whether software counts as a colleague.
A better prompt is not a delegation plan
Say you ask your agent to prepare a client update. It pulls information, writes a convincing summary, and tells you it’s done.
Did you explain which sources to trust? That last month’s forecast was superseded? That the agent can draft the email but cannot send it? And when you review it … you ARE reviewing it … are you checking the facts or just whether it sounds professional?
A polished update can still contain the wrong forecast, an unsupported commitment, or information that should never leave the company. None of those problems becomes less important because the paragraphs flow nicely.
These are management decisions. The chat window doesn’t make them disappear.
Visibility helps you investigate. Supervision helps you prevent.
When something goes wrong, the conversation quickly becomes: “We need more visibility into what agents are doing.” Sometimes we absolutely do. Agents introduce technical challenges around identity, access, monitoring, and control.
Existing application logs may capture a record change without capturing the full context behind it. Teams may need additional tracing to understand which tools were called, which sources informed an output, and what approval was involved.
But you can also have a beautiful audit trail of work nobody should have approved.
Logs can help reconstruct what happened. They cannot retroactively turn a vague assignment into a good one. You can solve the murder mystery with clues after it has happened … but the goal is to prevent it.
Technical observability and responsible delegation solve different parts of the problem. Neither excuses us from doing the other.
What agent-manager training actually looks like
This should be practical training with real assignments, not another presentation about how AI will change everything. People need to practice defining an outcome, limiting authority, setting checkpoints, verifying results, and correcting mistakes before handing over more responsibility.
For the client update, a useful assignment would explain:
- The outcome: Prepare a draft explaining progress, open issues, and decisions needed from the client.
- The trusted sources: Use the current approved project record and forecast. Identify conflicts instead of choosing whichever version makes the update easier to finish.
- The authority: Read the approved records and prepare the draft. Do not change project commitments or send anything externally.
- The evidence: Link important claims back to their sources and flag anything that could not be verified.
- The checkpoint: Return the draft to the responsible employee before it goes anywhere else.
That is more useful than “write a professional client update.” It also gives the reviewer something to check beyond whether the email has a respectable opening sentence.
The boundaries need system-enforced permissions behind them. If the assignment is draft-only, the agent should not have unrestricted sending authority just because everyone promised to be careful.
“Ask me if you need help” is not an escalation policy
It sounds reasonable. It also leaves the agent deciding what counts as needing help.
A more useful rule is specific: if the current forecast conflicts with the project record, stop and bring me the discrepancy. If the update requires a new delivery commitment, ask for approval. If you cannot verify a material claim, mark it unresolved rather than filling the gap.
Then test those situations. Give the agent conflicting documents and see whether the issue reaches the right person. Check whether it stops before the action, not whether it produces a thoughtful explanation afterward.
A training exercise should evaluate the employee too. Did they recognize the conflict? Did they inspect the supporting evidence? Did they give a correction that improves the next assignment?
The human is not automatically an effective control just because their name appears in an approval box.
More autonomy should follow evidence, not enthusiasm
Start with a bounded task. Review the result against clear standards. Correct the process when something fails. Expand authority where the evidence supports it.
An agent that produces reliable internal summaries has not automatically earned permission to email clients. An agent that handles one type of request well has not proved it can handle every exception around it.
None of this means standing over an agent’s shoulder all day. Good management is knowing what you can delegate, what evidence you need back, and where your judgment is still required.
The difference is speed. An agent can carry a bad assumption into hundreds of actions before your next check-in. Add multiple agents and handoffs, and the question becomes who owns the finished result … not just who gave the first instruction.
Put management skills into the personal-agent rollout
For leaders planning personal-agent adoption, training needs to sit alongside provisioning, permissions, and technical support. Give employees a repeatable way to assign work, a standard for reviewing it, and a clear path for handling exceptions.
Measure more than usage. Look at whether delegated work meets its acceptance criteria, how much correction it needs, and whether required escalations actually happen. A busy agent is not necessarily a well-managed one.
Delegation used to become your problem when you got promoted. With a personal AI agent, it can become your problem on day one.
We’re preparing everyone to have an agent. We should probably also prepare them to manage one … before the welcome package comes with production access.




