How do you manage a team of AI agents and people?

Write down who does what, run the team on a fixed rhythm, and manage the agents through evidence rather than impressions. The manager decides which work sits with agents and which with people, reviews held queues, exceptions and feedback daily and weekly, and changes an agent's authority only when its record supports it. People shift towards judgement and review.

How is managing agents different from managing people?

You do not motivate, coach or appraise an agent. You set its bounds, read its record, and change its brief, sources or depth when the evidence says so. Its performance is visible in a way a person's is not: every action, hold and refusal is filed.

The people change too. Much of their work becomes reviewing the agent's drafts, resolving what it hands over and handling the cases that need judgement. That is skilled work, and a manager who treats it as clerical will lose the people best at it. Which skills matter is covered in what skills a team needs to run AI agents.

What rhythm should the manager run?

The daily look is short. It answers one question: is anything stuck or going wrong? The weekly review is where improvement happens.

What are the steps, and who does each?

1. Manager: write the role map. Which tasks sit with which agent at what depth, and which with which people. Evidence: the role map, approved by the process owner. Our guide on deciding what stays with people when an agent joins a team helps with this.
2. Manager: set up the hand-over queues. Each with named people and cover. Evidence: queue list and rota.
3. Agent operations lead: build the daily view. Queue size and age, holds, refusals, errors. Evidence: the daily view.
4. Manager: run the rhythm in the table above. Evidence: the logs and notes.
5. Manager and process owner: propose depth changes when evidence supports them. Evidence: change requests with data.
6. Manager: develop people. Training in review, exception handling and spotting agent errors. Evidence: training records.

What does this look like in practice?

Take a lender's collections team of twelve people and two agents. One agent prepares arrears reminders and payment plan proposals; the other reconciles incoming payments against accounts.

The manager's role map keeps all contact with customers flagged as vulnerable, all disputes and all plans over six months with people. Each morning she checks the held queue and the reconciliation breaks from overnight, and assigns them. Each Friday she reviews the reason codes on rejected drafts. One week the codes show officers editing plan proposals for self-employed customers, because the agent was not using the income evidence those customers send. A source change is requested, tested and released.

Three months in, two officers who were best at spotting the agent's errors become the team's specialists for complex plans. The reconciliation agent's evidence supports executing matches at depth 3, and the manager proposes the change with the data. The way work moves between agents and people in her team is the subject of handing work between AI agents and people.

Which controls help the manager?

• A written role map, so nobody argues about who does what.
• One case record shared by agents and people.
• Reason codes on every hold and rejection.
• Authority changes only on evidence, with the process owner and risk agreeing.
• A stop switch the manager can use without asking anyone.

The rhythm builds on the stages in deploying an AI agent: train, shadow, certify, widen: widening an agent's authority is a management decision made on evidence.