Leading teams that work with AI agents

0:00 Leading teams that work with AI agents

Imagine this. It's Monday morning and your customer support team has a new colleague. It works twenty-four hours a day, handles hundreds of tickets, never gets tired, and occasionally, with total confidence, tells a customer something completely wrong. That colleague is an AI agent. More and more teams are working alongside them. The question for you as a leader isn't whether the technology is impressive. It's this: what work do we give it, with how much freedom, and who's accountable when it gets it wrong? In this lecture, you'll learn autonomy levels for agents, five leadership responsibilities, the governance basics, and how to write an agent charter.

0:47 Why it matters

Why does this matter now? Because agents are moving from demos into everyday work. They're built into mainstream office suites and customer-service platforms, and teams can build their own with automation tools. An assistant only suggests. An agent acts: it updates records, routes tickets, sends messages, runs a process. That's powerful, and it changes the leader's job. When an agent makes a mistake, it can repeat it hundreds of times before anyone notices. The management skills you already have, delegation, quality control and building trust, are exactly what make agents safe and useful.

1:27 Autonomy levels for agents

Here's the concept. Delegate to an agent the way you'd delegate to a new starter, with levels of autonomy. Level one, assist: the agent drafts or summarises, and a person decides and acts. Level two, recommend: the agent proposes an action with reasons, and a person approves each one. Level three, act with review: the agent acts, and people sample the results and can undo them. Level four, act within limits: the agent acts inside hard limits, such as a value cap, and people watch the metrics. Here's the key idea. Anything irreversible, customer-facing, legally significant, or about people's jobs and pay, stays at level one or two.

2:14 The junior doctor

Here's an analogy. Think of a hospital bringing in a brilliant new junior doctor. On day one, they don't operate alone. They take histories and suggest diagnoses, and a senior doctor signs off. As they show good judgement, they're trusted with more, but some decisions always need a consultant, and every patient has a named responsible clinician. Agents are like that junior doctor, but faster, more literal, and without the instinct to ask when something feels off. So you need the sign-offs, the named owner and the gradual trust even more.

2:54 Five responsibilities

Now five leadership responsibilities. One: a named human owner for every agent workflow, who defines the task, checks quality and answers for results. The agent can be responsible for a step. Only a person can be accountable. Two: clear boundaries, meaning what data the agent can see, what systems it can change, and what it must never do, with least-privilege access. Three: quality measurement, meaning sampling outputs and tracking errors, escalations, rework and complaints, not just volume. Four: psychological safety, so people feel safe saying the agent got this wrong. And five: skills and roles, planning training and talking honestly about how work will change.

3:40 Psychological safety and AI errors

Let's dwell on psychological safety, because it's where many teams quietly fail. Amy Edmondson's research on psychological safety describes teams where people feel able to speak up about mistakes and concerns without fear. With AI, there's an extra pressure: nobody wants to look like the person who's against innovation. So when the agent routes a ticket wrongly, people quietly fix it and move on. The error never gets logged and the agent never improves. Make reporting easy with a shared AI error log, and publicly thank the first person who uses it.

4:20 Worked example 1: Sana in Lahore (illustrative)

A simple worked example. Sana leads a six-person support team for an e-commerce brand in Lahore. The company introduces an agent that drafts replies and routes tickets. Sana sets refunds at level two: the agent recommends, a person approves every one. Routing goes to level three: the agent acts, and a team member reviews twenty routed tickets a day. She names Omar as the owner and starts an AI error log. After a month, routing accuracy is strong, so she moves routing to level four with limits. Refunds stay at level two. Money and trust aren't worth the risk.

5:03 Worked example 2: Dubai agency (illustrative)

Now a realistic business scenario, illustrative. A marketing agency in Dubai, with thirty staff, builds an agent that pulls campaign data every Monday, drafts client performance reports, and emails them. In week three, the agent sends a client a report containing another client's figures, because two data sources had similar names. The managing director, Rania, pauses the agent the same day and apologises to both clients. Then she fixes the system, not the blame. She writes an agent charter: a named owner per client report, no external emails without human approval, data access limited per client, and a weekly quality sample. The agent resumes at level two.

5:50 Governance basics

A word on governance. Keep it proportionate, but don't skip it. In the EU, the AI Act's AI literacy obligation has applied since the second of February, twenty twenty-five. Organisations that provide or deploy AI systems must take measures so staff who deal with them have sufficient AI literacy. Frameworks such as the NIST AI Risk Management Framework and the ISO standard for AI management systems give larger organisations a structured approach. And data protection law applies whenever agents touch personal data. So involve your data protection lead before connecting an agent to customer or employee records. This is general guidance, not legal advice.

6:35 Watch me: an agent charter

Watch me do it. I write an agent charter for an illustrative workflow: an agent that tags and summarises incoming sales enquiries. Purpose: route enquiries to the right account manager within an hour. Owner: our sales operations lead. Autonomy: level three, act with review. Allowed: read enquiries, add tags, write a summary. Never allowed: reply to customers or change deal values. Data: enquiry text only; our data protection lead has been consulted. Quality: sample fifteen a day; an error is a wrong tag or a missing key fact. Escalation: anyone can pause it from the team channel. Review monthly.

7:18 Measuring success

How do you measure whether it's working? Compare before and after on four things. Time saved. Quality: errors and, crucially, hidden rework, because time saved by the agent can reappear as time spent fixing its output. Customer outcomes, such as complaints and satisfaction. And team experience: do people trust it, and do they feel their work has become more or less meaningful? If time saved comes with rising complaints or hidden rework, the workflow isn't working yet. Adjust the autonomy level down, not just the prompt.

7:55 Common mistakes

Common mistakes. Giving an agent more freedom than you'd give a new starter. No named owner, so errors fall between teams. Measuring volume instead of quality. Punishing people who report AI mistakes. Connecting agents to personal or client data without a data protection review. Forgetting to tell customers when they need to know they're dealing with AI. And staying silent about what agents mean for people's roles, which fills the gap with rumours and resistance.

8:28 Recap and try this now

Let's recap. Delegate to agents like a fast, literal new starter, and match autonomy to risk, starting low. Keep irreversible, legal and people-related actions at assist or recommend. Give every workflow a named owner, clear boundaries, quality sampling and a pause button. Make it safe to report AI errors. And stay on top of AI literacy and data protection duties. Your try this now: list every AI assistant or agent your team uses and, for each, name the owner, the autonomy level and one quality metric. Then write a full charter for the riskiest one. Next, we'll master asynchronous communication.