How to Manage Human-AI Teams in the Age of AI Agents: A Training Guide for Modern Managers

We hand things over to the AI, yet in the end it stays oddly unclear who actually owns the outcome.

That was the remark Mr Moriwaki(not his real name), a sales-planning manager at one operating company, let slip to a learning-and-development colleague after the weekly meeting. Until six months earlier, the company had been happily using AI agents as a handy tool for drafting documents and summarising minutes, but it had never really thought through where those agents sat within goal-setting, delegation and the appraisal system.

Team members were unsure how far to trust the AI’s output, while managers wrestled with where to draw the line between work to leave to people and work to leave to the AI. HR, for its part, found that its conventional management training assumed teams made up solely of people, and so could not quite settle on how to teach the running of a team that now included AI agents.

Today, off the back of a three-month pilot programme for twelve managers, the number of teams that spell out the AI agent’s role in their regular meetings has risen from two departments to seven. In this article we set out, through the lens of management training for the age of AI agents, how to design roles, shape communication and build psychological safety so that a human–AI team can be run well.

The aim is a hybrid team that treats AI not as a stand-in for a subordinate but as a member that widens the team’s capabilities. Training alone, mind you, will not dissolve every management challenge of the AI era. Only when you add operating rules, a rethink of appraisal and steady, on-the-ground reflection do you arrive at a way of running a team that can be repeated reliably.

How a manager’s job changes in the age of AI agents

Once an AI agent joins the workplace, a manager’s job stops being merely a matter of “mastering the AI”. By an AI agent we mean an AI system that, in line with the user’s instructions and the business purpose at hand, supports work with a degree of autonomy: organising information, drafting text, summarising, pulling out tasks and suggesting the next action.

Managers have traditionally watched the volume of people’s work, their progress, their results and their motivation. But once an AI agent is writing the minutes, drafting customer replies and marshalling the key points of a report, a manager also has to keep an eye on how the work is being divided between people and AI.

Suppose a team member has an AI agent draft a proposal. What the manager needs to check is not simply how polished that draft is.

• Were the background facts handed to the AI the right ones?
• Is a person supplying the genuine understanding of the customer?
• Who checked the figures and the proper names?
• Where does final responsibility for the judgement sit?

Ignore those questions and judge only whether the AI saved time, and the team may well drift in the wrong direction, for AI use can climb while the quality of judgement and the depth of customer understanding quietly thin out. A manager in the age of AI needs not just the skill to use an AI agent but the ability to design how people and AI work together.

Treating an AI agent not as a tool but as a member of the team

In many companies, generative AI and AI agents first arrive as an “efficiency tool”: tidying up emails, summarising minutes, knocking together a first draft of a document.

There is real value in that stage, of course. But once an AI agent is used across a broader business context, treating it as a mere tool makes it hard to run well.

Consider, for instance, an AI agent that reads through a sales meeting, pulls out the next actions and drafts a set of tasks for each person. Here the AI is no longer a simple document-writing tool; it has become something that shapes what the team does after the meeting.

For that reason, a manager needs to hold questions of the following kind.

• How far should we delegate to the AI agent?
• Which points must a human always check?
• Who turns the tasks the AI proposes into formal instructions?
• Does the AI’s output leave members feeling anxious or watched?
• How do we treat the gains from using AI as gains for the team as a whole?

Treating an AI agent as part of the team does not mean treating it like a person. Rather, it means being clear about its role, as something with strengths and limits quite different from a human’s.

This matters from an AI-governance standpoint too. To use AI safely and appropriately, an organisation has to get straight its scope of use, lines of responsibility, explainability, and education and literacy. For an authoritative reference on principles for the trustworthy use of AI, see the OECD AI Principles, which set out internationally agreed principles for trustworthy AI.

Role design is the first thing management training should tackle

The first topic to tackle in management training for the age of AI agents is role design.

In teams where AI is not working well, it is common to find that what to leave to the AI has been left to each individual. One member hands almost every draft to the AI; another barely touches it; a third shares the AI’s output internally without checking it properly. In that state, the team’s quality never settles.

For role design, it helps to sort the work into four broad categories.

Gathering information

Searching past material, summarising minutes and classifying enquiries are areas that suit an AI agent well.

Producing first drafts

Emails, the outline of a proposal, a draft of training material, a set of FAQs — having the AI produce the first version frees up people’s time to think.

Making judgements

The final proposal to a customer, appraisal comments, hiring decisions, the interpretation of contract terms — these are areas a human should own. The AI can help marshal the issues and build a comparison table, but it cannot be the one who makes the final call.

Building relationships

One-to-ones with a team member, building trust with a customer, forging agreement between departments — here a human’s grasp of context and reading of feeling are what count. The AI can help you prepare, but it is no substitute for the conversation itself.

In management training, it works well to start by taking stock of your own team’s work against these four categories.

Judge goal-setting by how it fed into team results, not by whether AI was used

Once AI agents are in place, managers are tempted to watch how much the AI was used, since the number of uses, the count of documents generated and the hours saved are all easy figures to capture.

But make the sheer amount of AI use the goal, and the team starts treating using it as the point in itself. What matters is how that use fed through to the team’s results.

For a sales team, say, it is not enough that the AI shortened the time spent writing proposals. Did that, in turn, free up more time for understanding the customer? Did the quality of the proposals rise? Did follow-up after a meeting get quicker? You need to look at that connection to results.

For an HR team, the point is not that the AI produced the training material, but whether participants came away with knowledge they can use on the job and whether managers could put it to work in conversations with their people.

An AI agent is not the goal itself but a means of reaching the goal. In management training, that premise is worth restating again and again.

In delegation, separate what the AI is left to do from what a person approves

In the age of AI agents, delegation is not only about delegating to people. It also takes in designing just how far you delegate to the AI agent.

The crucial thing here is not to confuse letting the AI do something with letting the AI decide it.

The AI can marshal the material for a decision, but it cannot shoulder responsibility. So a manager has to be clear about who signs off on whatever the AI produces.

Settling on rules of the following kind within the team tends to keep things on an even keel.

• Anything that leaves the building is always checked by a person
• Figures, proper names and contract terms are checked against the original source
• Appraisal comments written by the AI are never used as-is
• Where the AI is unsure, have it flag the item as “needs checking”
• Topics where the judgement is genuinely hard are handled in a meeting of people, not by the AI

Delegation is not about letting the AI off the lead. It is about designing where you delegate and where you draw the line — a point that sits neatly with the question of managing the risks of generative AI and keeping humans in oversight. For an authoritative reference, the NIST AI Risk Management Framework sets out a framework for managing AI risk across an organisation.

In appraisal, look separately at attitude to AI and at quality of judgement

As AI agents spread, the appraisal system feels the effects too.

Should the member who churns out a great deal with the AI be rated more highly? Does the member who works carefully without it end up undervalued? How should we rate the knack of editing the AI’s output well? Push AI use forward without answering these, and a sense of unfairness takes root on the ground.

Management training need not overhaul the appraisal system overnight. It is important, though, to hold at least the following perspectives.

Look at the link to results

The point is not whether someone used AI but how they connected it to results. Heavy use counts for nothing if a lack of checking has dragged the quality down.

Look at the skill of instructing the AI

The ability to supply the right background, make the constraints clear and draw out the output in a form you can compare is becoming part of the core skill set.

Look at the skill of reviewing AI output

You need the ability not to be carried along by plausible-sounding phrasing, and to separate fact from inference from the unverified.

Look at how well it is shared with others

Someone who does not merely get results with the AI on their own but shares effective prompts and operating rules with the team lifts productivity across the whole organisation.

In the appraisal system of the AI era, you cannot do without a view that rewards not only the person who can produce things quickly but the person who asks the right questions, verifies, and gives back to the team.

In a human–AI team, designing communication matters all the more

When AI agents come in, team communication grows more efficient, yet differences in understanding can become harder to spot.

Leaving the minutes and the tidying of tasks to the AI makes sharing quicker. But if only the AI’s text circulates while human-to-human checking and background explanation drop away, things no one has actually agreed can start being treated as settled decisions.

If, say, only an AI-generated task list is shared in an internal chat such as Slack, members may be left wondering whether it is a formal instruction, who set the priorities and whether they are allowed to change it.

For that reason, managers need to label communication that involves an AI agent.

• This is a draft written by the AI
• This is a decision the manager has checked
• This is a starting point for discussion
• This is an unverified hypothesis
• This is a point to be decided at the next meeting

Labels as simple as these cut a great deal of confusion out of the team.

And while a summary from an AI agent is handy, there are always remarks that did not make the cut and a sense of unease that does not. In one-to-ones, appraisal meetings and customer negotiations especially, the feeling and context that never quite turn into words are what matter.

Using the information the AI has tidied while not cutting too many corners on human-to-human checking is what communication in a hybrid team requires.

Without psychological safety, AI use will not take hold

Bringing in AI agents is a positive change for members, but it can breed anxiety too.

• Will my job be replaced by the AI?
• If I cannot use the AI well, will my rating slip?
• Will the records the AI keeps be used to watch what I say and do?

Leave such worries unaddressed while pressing on with AI, and people may look like they are using it on the surface while quietly resisting underneath.

Management training has to cover not only the skill of using AI but how to protect members’ psychological safety.

First, it is important to say plainly that the AI is not a surveillance device for judging members. Meeting summaries and task extraction, you need to explain, are there to get everyone on the same page, not to apportion blame.

Second, it matters to create a space where AI failures can be shared. The more a team can share not just the prompts that worked but the misfires, the wrong outputs and the missed checks, the more its use of AI tends to sharpen.

Third, the decision not to use AI deserves respect too. Not every task needs AI in it. Where human conversation or careful judgement is called for, deliberately leaving the AI out is itself a management decision.

To make AI use stick, you have to design not only for convenience but for how the anxiety is faced.

A design example for carrying management training through to practice

Management training sticks better when it is tied to the tools people actually use and to existing work processes, rather than left as lectures alone.

At the start of the programme, for instance, people learn the basics of AI agents and the rules for handling information, building a shared understanding of what AI is good at, where it falls short and which risks to watch.

Next come case exercises tailored to each manager. A sales manager can practise pulling next actions out of meeting notes and turning them into instructions; an HR manager can practise designing the questions to put to a team member from one-to-one notes; an IT manager can practise classifying internal enquiries and turning them into knowledge.

At this stage,a service such as Kanata, which our company provides, which brings AI chat, AI summarising, e-learning and project-level library management together in one environment, makes it easier to connect what was learned in training to day-to-day operation. Save the prompts, decision criteria and checklists used in training to a project library, for example, and the team can reuse them long after the course ends.

That said, you need not use Kanata specifically; combining an existing LMS, internal chat, knowledge-management tool and generative-AI tool works just as well. What matters is that the training does not end at “attended and done”, but lives on in a form people can use in next week’s meeting, in one-to-ones, in drafting documents and at the point of decision.

In the month after the course, each manager tries one AI-agent-driven improvement with their own team and shares the result. Designing in a cycle like that helps the leadership learning of the AI era take root on the ground.

Hands-on exercises worth including in management training

In management training for the age of AI agents, it is important to fold in hands-on exercises, not just lectures, because understanding something as knowledge rarely settles into day-to-day practice on its own.

Five exercises are worth recommending.

Leadership in the AI era shifts from supplying answers to designing questions

Managers have long been asked to supply answers drawn from experience, and judgement will of course remain important.

But as AI agents spread, what is asked of a manager, more and more, is the ability to design the question.

• What should we ask the AI?
• Which assumptions should we hand it?
• On what grounds should we compare?
• From which point on should a human judge?
• Which output should the team share?

The quality of questions like these is what shapes the team’s results.

The AI does not take the lead in a manager’s place. If anything, it lays bare the coarseness of a manager’s questions, the vagueness of the role design and the gaps in the appraisal criteria.

That is precisely why management training in the age of AI agents cannot be mere tool training. It needs to be designed as a place to relearn leadership, management and team design.

In summary: running a human–AI team is a new foundational skill managers must relearn

Bringing in AI agents does not shrink the manager’s role. If anything, the responsibility for design — combining the strengths of people and AI to deliver results as a team — grows larger.

It is not whether you use AI but which work you leave to it, which judgements people own and which results you appraise as a team. Whether you can design that is what sets managers apart in the AI era.

The topics for management training in the age of AI agents are not just how to operate the tool. They have to take in role design, goal-setting, delegation, the appraisal system, communication and psychological safety — the whole business of running a human–AI team.

In designing the programme, it helps to combine tools — AI chat, AI summarising, e-learning, knowledge management — to suit the purpose. A service such as Kanata, which handles learning, exercises and library management in one environment, is one option where you want to carry the training through to practice on the ground.

An AI agent is no cure-all, though. It cannot stand in for human judgement, conversation and responsibility. Which is exactly why a manager needs both the nerve to delegate to the AI and the responsibility a person must carry.

Management training from here on cannot stop at producing managers who can use AI. What matters is producing managers who can design teams where people and AI deliver results together.

Q&A