Forty minutes into a conversation about organizational change, a senior executive raised her hand and asked the question that stopped the room.
She led a team of ten people. Each person now relied on several AI agents, some for research, others for customer insight, campaign planning, reporting, and the steady work of keeping a function running.
“So am I managing ten people,” she asked, “or am I managing something closer to sixty entities?”
The room went quiet. It was not the polite silence of executives waiting for the next slide. It was the deeper quiet of people recognizing a question they had been carrying for some time without knowing how to name it.
Her team still consisted of ten people. But the amount of work being produced through AI agents had expanded, and she remained responsible for what her function delivered.
That question led eight colleagues and me into the research that became The Human-Agent Orchestrator.
We wanted to understand one thing:
What changes for a leader when AI agents begin carrying out part of the work?
The answer reaches well beyond the technology.
Managers are becoming responsible for work they did not personally perform and that no employee may have performed. They must decide what to assign to an AI agent, explain what the work requires, examine the result, and know when it should not be used.
Most have received little preparation for any of this.
The management system has not caught up
For years, the workplace conversation about AI centered on the tools themselves.
Could they make research faster? Could they prepare a first draft? Could they reduce the time spent on reports or routine requests?
Those questions still matter. But they do not tell a manager how to remain responsible for work carried out through an AI agent.
An agent can prepare a customer analysis in minutes. The manager still needs to know whether the analysis used the right information and whether its conclusions fit the situation. If the work reaches a customer, affects an employee, or informs a consequential decision, someone must be able to explain why it was accepted.
This is where the current management system begins to show its age.
Job descriptions tell managers which people and business results they are responsible for. Approval processes establish who can authorize a decision. Reporting lines show where an issue should move when someone needs help.
These arrangements were not written for a workplace in which AI agents can carry out a growing share of the work.
The organization may approve the technology without deciding who will manage its use inside the workflow. Employees begin using it. Managers inherit the results. Responsibility moves before the formal management structure does.
This is why one conclusion from our research has stayed with me:
We are the last generation to manage only humans.
Managers will continue to lead people. They will also become responsible for directing and reviewing work carried out through AI agents.
The agents may change quickly. Human responsibility for the outcome cannot remain vague.
When the technology works
The first stage of adoption often looks encouraging.
Messages are drafted faster. Meetings are summarized. Research that once took hours can be assembled quickly. Reports appear sooner, and small teams find that they can take on work they could not previously manage.
Then the harder questions begin.
Who checked the reasoning? Does the person reviewing the work know enough to recognize a subtle error? When can an AI agent proceed without review, and when should the work stop until someone examines it? If the result is wrong, who had the authority to prevent it from being used?
The tool may work exactly as designed while the organization remains uncertain about how its output should be managed.
That uncertainty often appears late. The technology is introduced through a pilot or made available to employees. Initial training explains what the system can do. Far less attention goes to the manager who must decide how the output fits into actual work.
A capable system does not settle the management question. In many cases, it makes that question harder.
The faster an agent produces work, the more material a manager may have to review. The more confidently the system presents an answer, the more important it becomes to know what evidence sits behind it. The further an automated process can proceed, the more clearly someone must define where it stops.
These responsibilities belong inside the design of the work. They should not be left for individual managers to discover after something goes wrong.
The promotion almost no one asked for
A person who begins directing work through AI agents can take on management responsibility before anyone recognizes that the role has changed.
They may need to define the assignment, supply the necessary context, explain what a good result looks like, and decide whether the answer can be trusted. They may have to correct the work or stop the agent when the situation falls outside its limits.
This can happen even when no one reports to them.
That is an important change in professional work. Management has traditionally been associated with responsibility for other people. AI agents loosen that connection. An individual contributor may now manage a growing volume of agent-produced work without the authority, training, or recognition usually associated with management.
Some experienced professionals adapt quickly because they already know how to delegate.
They have learned to write a useful brief. They recognize when an answer sounds polished but rests on a weak assumption. Years of reviewing other people’s work have taught them where errors tend to hide.
Less experienced employees may be asked to do the same before they have developed those instincts.
An AI agent can produce a sophisticated first draft for someone who has never created one independently. That person may be expected to review the work without having learned what a strong draft requires or where it is likely to fail.
The agent completed the task. The employee now carries a level of responsibility they may not yet be ready to hold.
This is the promotion almost no one asked for.
It will affect more than managers. Professionals throughout an organization will need to direct work they did not perform, even when their titles remain unchanged.
Organizations need to recognize when this responsibility has entered a role. Otherwise, employees will be judged on their ability to oversee AI-produced work without being taught how to do it.
Trust changes when the work changes
Managers know that trust between people develops over time.
A colleague remembers what happened with a customer six months ago. They know why a particular approach failed in the past. They may notice that the instructions do not fit the circumstances and ask a question before continuing.
Much of that context is never written into the assignment.
An AI agent receives what people make available to it. It may produce convincing work without the organizational history or professional experience that would cause someone to hesitate.
For that reason, confidence in agent-produced work must come from the way the work is managed.
The manager needs to understand what information the agent used and the limits under which it operated. The result may need testing before it can be relied upon. Work with serious consequences may require a named reviewer who has the authority to reject it.
This is not distrust of the technology. It is ordinary management applied to a new source of work.
NIST’s AI Risk Management Framework calls for organizations to define who is responsible for monitoring AI systems, responding to incidents, and managing the risks created by their use. It places responsibility for those decisions with named people and organizational functions, including senior leadership.
The OECD makes a related point. Organizations need safeguards that preserve human agency and provide an appropriate level of oversight, with accountability assigned according to the person’s role and ability to act.
Those principles become real in everyday work. Someone has to decide whether an AI-produced recommendation is ready to use. Someone has to know when the work should stop. Someone must answer for the decision that follows.
From supervision to orchestration
This change is moving management from supervision towards orchestration.
Supervision has often centered on the person carrying out the task. A manager could observe the work, discuss how it was done, and help the employee improve.
When AI agents perform parts of the work, managers may see only the request and the finished output. The important decisions can occur in between: what the agent was told, which information it received, what it was permitted to do, and where a person was expected to intervene.
The manager is now responsible for the arrangement around the work as well as the result.
One of the clearest explanations of this came from an interview we conducted for the book with a symphony conductor.
The part the audience sees, the hands, the podium, and the performance, is the smallest part of the conductor’s job. Most of the work happens before the first note. The conductor studies the score, clarifies the intent, understands what each section contributes, and decides where precision matters most.
A conductor does not play every instrument.
A conductor creates coherence.
That idea stayed with me because it describes an important part of the manager’s work now. The manager does not need to perform every task. But they do need to understand the purpose of the work well enough to organize it, recognize when something is out of balance, and decide what needs human attention.
The book calls this human-agent orchestration.
I use Human-AI Management for the management system that supports it: leading people and managing AI agents while keeping accountability for outcomes clearly human.
What people inside organizations are asking
The management discussion can easily become absorbed by systems, controls, and operating processes.
The people doing the work are having another conversation.
They are looking at tasks that once demonstrated their skill and seeing an AI agent complete them in minutes. Some are relieved to lose repetitive work. Others are trying to understand what will happen to the experience they spent years building.
They are asking:
What happens to my work?
What should I learn now?
Will my judgment still matter?
What becomes of my experience when the method changes?
These are practical career questions.
For many professionals, doing the work was how they learned to recognize when something was wrong. It was also how a manager learned that they could be trusted with something more difficult.
If a junior employee no longer prepares the first analysis, where will they learn which assumptions usually fail? If an experienced professional moves from producing work to reviewing a much larger volume of AI output, will the role still give them room to think?
The answers will differ across professions. Leaders still need to ask the questions.
A company can introduce AI successfully and leave its people uncertain about what the change means for them. Employees may use the new systems while becoming less certain about their own value. They may complete the required training without seeing how their work or development will change.
Training people to use the system is only part of the work.
Managers also need to explain what responsibility now sits with the person, what expertise the organization still needs from them, and how they will continue to learn when some of the old learning tasks move to AI.
When values become part of the work
Organizations often describe accountability as a core value.
AI agents force that value into the operation of the work.
If an agent prepares material that reaches a customer, the organization must decide who approves it. If the agent acts beyond the expected case, someone needs the authority to stop the process and examine what happened.
Accountability cannot be assigned to the agent.
It belongs with the person or organizational function that has the knowledge and authority to act. That person must know that the responsibility is theirs, and the surrounding process must give them a real opportunity to exercise it.
A name placed at the end of a workflow is not enough.
A manager cannot be accountable for an outcome if they cannot see how the work was produced, change the conditions under which the agent operates, or prevent the result from being used. Responsibility without authority is only a label.
This is where leadership becomes visible.
It appears in the decisions about what an agent may do and where a person must remain involved. It appears in whether managers have enough time to examine the work rather than approve it under pressure. It appears in what happens when an employee raises a concern about a system that appears to be working as designed.
Values enter the workflow through decisions like these.
The responsibility is already here
Many organizations are still writing their AI strategy while employees are already using agents in everyday work.
Managers cannot wait for every policy, role description, and operating model to be finished. They are making decisions now about which work can be delegated, what needs review, and whether the result is good enough to use.
Leadership teams need to see that responsibility clearly.
Managers require preparation grounded in the work they actually oversee. They need to understand the agents being used in their functions, the limits placed around them, and the decisions for which they remain answerable.
Employees taking on agent-management responsibilities need the same clarity. If the organization expects someone to review sophisticated work, that person needs the experience and authority required to challenge it.
Senior leaders must also decide where these responsibilities sit across the organization. Technology teams can establish systems and controls. Risk and legal functions can set requirements. But the business leader who owns the work cannot hand away responsibility for the outcome.
We are the last generation to manage only humans because this responsibility is already entering the manager’s role.
The organization chart may still look the same. The manager may still lead ten people. Yet more of the work under that manager’s authority may now be researched, drafted, analyzed, or carried out through AI agents.
The question raised by that executive is therefore one every leadership team should now be asking:
Who manages the agents, and who is accountable for their outcomes?
Technology may set the pace. Leadership sets the direction.
