Your Leader Is Basically a Relay
Last month I had dinner with a friend. He leads a team of twenty at a big company, and his company is pushing an AI-first transformation. He was a little rattled.
"You know what I actually do every day? In the morning I read Feishu and translate the decisions from above into something the people below can understand. In the afternoon I read the weekly reports and translate the progress from below into the format the people above want to see. In between there are a few meetings where I coordinate whether team A or team B goes first."
He paused, then said something that stuck with me:
"I suddenly realized recently that an agent could probably do all of this."
He was right. Coordinating schedules? An agent can align both sides' specs and dependencies and automatically surface a conflict resolution. Summarizing weekly reports? An agent pulls status directly from execution data, faster and more accurately than a human. Reporting upward? If executives can query the data themselves, look at a dashboard, or even ask an agent directly, where is the value of that middle translation layer?
This is not my friend's problem. He is a very responsible person. The problem is that many leader positions are themselves built around information asymmetry. In the past, information was a scarce resource—the top did not know what the bottom was doing, the bottom did not know what the top was thinking, and cross-team, nobody knew what the other side was busy with. So you needed a layer of people whose job was to move information around, translate it, and summarize it.
Now AI has knocked down that asymmetry. Information has become readily available. The leader position that existed by moving information around has lost its foundation.
The Leaders Who Disappear Never Should Have Existed
This sounds like I am saying AI is going to wipe out leaders. It is not. What AI wipes out is one specific kind of leader: the ones whose value rests on information gaps.
The relay type—adds no value, just reports up and relays down. Their core output is "message delivered."
The report-aggregator type—their main job is collecting and summarizing subordinates' weekly reports, turning ten people's work into a three-page deck. Their core output is "information aggregated."
The blind-commander type—no real judgment, makes decisions through positional power, because "I'm the boss so listen to me." Their core output is "decision made."
These three kinds of leaders have one thing in common: if you remove them, information flows faster.
In the past, it was hard to tell whether a leader was genuinely valuable or just moving information around. Because the flow of information itself had a cost—you had to hold meetings, write emails, prepare reports—so having someone dedicated to that looked like a reasonable division of labor.
What AI changes is not the value of leaders but that it exposes which leaders never had value. When everyone can access information directly and invoke capabilities directly, the "information relay" value of the middle layer drops to zero. The leaders who existed on information gaps have had their fig leaf pulled away.
But notice, this is not the same as "AI replaces leaders." Those people were never leaders. They just happened to be sitting in a leader's seat.
The Leaders Who Truly Survive Become a Different Species
So after those people are gone, do we still need leaders?
Yes. But the definition of a leader has to be rewritten.
I have been thinking about this a lot, and I eventually landed on a word: orchestrator.
Why do we need an orchestrator? Because the sum of local optima rarely equals the global optimum.
An agent can make every module and every task locally optimal—but that "optimum" is defined within the boundary it can see. The best solution for module A might double the maintenance cost of module B. This week's optimal schedule might drive next month's architectural evolution into a dead end. The optimal delivery cadence of three teams, each on its own, might combine into an integration disaster.
An agent cannot see these things. Not because it is not smart enough, but because an agent's field of view is inherently local—it is bounded by the prompt, the scope of the spec, the context of the task. It will not spontaneously look up to see what the neighboring team is doing, nor will it wonder whether a shift in requirements three months out will make today's solution obsolete.
This is work that requires a person. A person who is not sunk into the execution of any single module and can look at the whole system from above.
That is the orchestrator.
A Leader's Job Did Not Shrink, It Got Heavier
Reading this far, you might ask: so from now on a leader just handles orchestration? What about managing people? That is gone?
Quite the opposite. Managing people should not only continue, it should get more attention than before. It is just that the time previously eaten by coordination work can now be given back to what truly matters.
Let me break this down. A leader's job in the AI era splits into three layers.
Layer one: the burden AI takes away.
The things mentioned earlier: aggregating reports, coordinating schedules, relaying information, routing approvals. AI can do these faster and more accurately, and leaders no longer need to spend time on them. This is not stripping power away; it is removing burden.
Layer two: the core that was always there.
Observing and coaching employees—someone has been off lately, is he stuck on something? Shaping team culture—are people willing to tell the truth, and when something goes wrong do they point fingers or carry it together? Mediating conflict—when two people disagree on an approach, it is not resolved by having an AI declare a winner; you have to find the difference in assumptions behind the disagreement. Hiring judgment—whether a candidate is technically strong enough and fits the team; an AI can screen resumes but cannot make the final call. Performance feedback—telling someone what they did well and where they need to improve is on a different order of difficulty from writing code.
AI cannot do these things. They are not "soft skills"—they are the reason the leader role exists. And in the AI era, their importance only grows. Why? Because when execution is accelerated, team composition gets more complex (people plus agents), and the pace of change goes up, people's states become easier to overlook. In the past you could naturally observe team culture in weekly meetings, scheduling discussions, and daily collaboration; now those touchpoints have been compressed by AI. If you do not deliberately carve out space, you might only discover after something breaks that someone has been burned out for three months.
So in the AI era a leader is not done with managing people—they have to manage people more deliberately. In the past you could observe team state in the gaps of coordination work; now that coordination is done by AI, observation and coaching require you to schedule time and create touchpoints on purpose.
Layer three: the new layer—orchestration.
This layer is something leaders did not need to do much of before—or rather, only a few senior people needed to. But now it is becoming part of every leader's job.
Set constraints instead of issuing instructions. Do not say "do it this way," say "this problem must be solved, these boundaries cannot be crossed, the rest is your call." Both agents and engineers need enough autonomy; the fewer and clearer the constraints, the higher the execution efficiency.
Find global optimization opportunities instead of watching local efficiency. When you are no longer consumed by the execution details of some module, you can finally see the duplicated work between A and B, the infrastructure C and D could share, the fact that what module E keeps solving is not a technical problem but a requirements-definition problem. The people heads-down in execution cannot see these things; if you cannot see them either, then no one sees them.
Allocate resources instead of hoarding them. Old-style leaders had an instinct: hold on to people. "This person reports to me" mattered more than "what this person should be doing." An orchestrator flips this—people do not report to you, but you must be able to judge which task should go to whom, which task should go to an agent, and which task should be put on hold.
Align context instead of re-translating information. Translation converts language A to language B; the amount of information stays the same. Alignment ensures everyone understands the same thing the same way—what the goal is, where the boundaries are, what success looks like. This is not moving information around; it is compressing and focusing it.
Add these three layers together and a leader's job actually got heavier, not lighter. The coordination burden was taken away, but attention to people has to be more deliberate, plus there is a new layer of global orchestration. Those who only orchestrate and do not care about people will find they are scheduling a system that spins in place; those who only care about people and do not orchestrate will find they took great care of the team but drifted off course.
The Hardest Part Is Not Layer Two, It Is Layer Three
Of the three layers, most leaders were already doing layer two (observing and coaching people); it just needs to be more deliberate now. The truly hard one is layer three—the global perspective orchestration requires.
The problem is, where does the global perspective come from?
How did leaders in the past acquire a global perspective? Mostly by grinding through it. Stay in an organization long enough, see enough projects fail, argue with enough teams, and you slowly develop an intuition for "what will go wrong." The process is slow, but it at least closes the loop: execute → collide → get burned → form judgment.
In the AI era, that path is being cut too.
If a leader no longer participates in execution firsthand, no longer argues over concrete details with people, no longer pays the price for a wrong decision—then what does he rely on to form global judgment? If all he sees is "team A's delivery slipped by 3 days" on a dashboard, without knowing the real reason behind the slip (not enough people? requirements changed three times? the agent keeps failing on some edge case?), his scheduling becomes another form of blind commanding.
Stepping back from execution is meant to gain a global perspective, but fully detaching from execution turns that perspective into a castle in the air.
This is not a problem you can easily route around. The direction I can think of is: an orchestrator cannot be entirely hands-off from execution but should execute selectively—not by writing code, but by chasing the causal chain of key decisions. Why did this requirement change three times? Why did that cross-team conflict take three weeks to surface? Every time a systematic surprise appears, the orchestrator should personally dig into the root cause, not just read an incident report.
This is really the same proposition as the previous post. Newcomers train their judgment by reviewing agent output plus a senior's retrospective; leaders maintain the feel of global judgment by tracing the causal chain of key decisions. Two roles, the same predicament: once execution is outsourced to AI, where does judgment come from?
How the Organization Should Adjust
If the analysis above holds, there are a few moves an organization cannot avoid:
First, dismantle the leader's relay work, but give the freed-up time back to the team in a structured way. If your report agent can write reports, your scheduling agent can coordinate, and your sync agent can keep information aligned—then let the agents do it. But if the freed-up time just gets filled with more meetings, dismantling the relay is pointless. That time should be explicitly allocated to coaching, 1:1s, team retrospectives, and cross-team alignment. In the past you could observe team state in the gaps of coordination work; now that coordination is taken away, observation and coaching require you to deliberately create touchpoints. If the organization only removes the burden without building the structure, leaders will end up in a vacuum: "no longer handling relay but also not tending to the team."
Second, span of management can grow, but depth of management cannot get shallower. After AI lowers coordination costs, the number of people a leader can cover can indeed increase. But coverage is not the same as understanding. What you need to understand is not what each person does every day, but the decision logic of each direction, where the key risks are, and what unresolved tensions exist between teams. A larger span is an efficiency gain; unchanged depth is a quality floor.
Third, decision rights have to be made explicit. The biggest trap of flattening is cutting layers without redistributing decision rights. On the surface everyone is equal; in reality, whoever is loudest, most senior, or closest to the boss calls the shots. This "shadow hierarchy" is worse than a formal one, because it cannot be questioned, cannot be optimized, cannot be held accountable. A good orchestrator does not hoard every decision but makes the ownership of each class of decision crystal clear: which decisions the agent makes itself, which the engineer decides, which must escalate to the orchestrator.
Fourth, the way leaders are evaluated has to shift from "team output" to "system improvement." In the past, you evaluated a leader by how many features the team shipped. Going forward, agents are shipping features too, and it becomes increasingly hard to separate how much of the output came from agents and how much from the leader. A more reliable evaluation standard is: what systematic improvement did this leader leave behind? Did they make the team's next delivery more reliable? Did they find and eliminate duplicated work across teams? Did they distill the root cause of an incident into a long-term organizational constraint?
Things I Have Not Figured Out
One uncomfortable question: if a leader's core value is the global perspective, and AI's global data-analysis ability keeps getting stronger—then where exactly is the unique value of a leader's global perspective?
The answer I can give myself is: data tells you what is happening, but not what should happen. AI can say "team A and team B have a scheduling conflict," but it cannot say "whether this conflict is worth resolving, or whether we should let them collide, and after the collision both sides will understand their boundaries more clearly." The latter is judgment, values, and the willingness to bear consequences. AI does not have these, and should not.
But whether this is the endgame, I am not sure.
Another question: will the orchestrator role also turn out to be a transitional form? If one day agents can not only execute but also do global scheduling and optimization—then what is left for the orchestrator? Maybe by that day the orchestrator retreats further to the position of "goal setter": not scheduling concrete resources, not designing concrete systems, only defining what is good and what is not allowed. But what role that is, and what it is called, I cannot say clearly now.
There is also a more practical question: in today's organizations, how many leaders are willing to do layer three? "Managing people" comes with a sense of power—people report to you, you can approve their leave, you can decide their performance. But layer three is lonely—you no longer own anyone, you are just designing a system that lets everyone work better. This work has no sense of power, only a sense of responsibility. Not every leader is willing to make that shift.
In one sentence: AI did not make leaders disappear, it merely zeroed out the ones who fed off information gaps; the leaders who survive have a lighter coordination burden, but their attention to people has to be more deliberate, plus a new layer of global orchestration—the leader role did not get lighter, it got heavier.
References
- Fortune (2025/08/07), "AI is already changing the corporate org chart" — a systematic account of the Great Flattening trend
- AI-Native Transformation Framework, "The Flatter Hierarchy" — decomposition of five middle-management functions: coordination, aggregation, escalation, and strategic breakdown are absorbed by AI, while people development is retained
- BCG, interview with Nick South — first use of "orchestration layer" to describe management in the AI era
- McKinsey, "Building leaders in the age of AI" — core leadership capabilities in the AI era: strategy, judgment, creative thinking, values-driven leadership
- HBR (2025/07), "How AI Is Redefining Managerial Roles" — leaders shift from decision maker to sense maker
- Deloitte Insights (2025), "What's the future of management?" — the problem of shadow leaders and ambiguous decision rights in the absence of formal titles
- HBS/INSEAD Working Paper 26-090 — quantitative data on YC startups: AI-native companies have 25% fewer people and ~15% fewer managers
- LinkedIn/Lize de Kock, "Balancing The Great Flattening with the Need for Leadership in the AI Era" — 42% reduction in middle-management roles, 80% change success for middle leaders vs 20% for senior leaders
- Tsinghua University School of Economics and Management, "Symbiotic Leadership: A New Leadership Paradigm for the Deep Integration of Humans and AI"