Every job is shifting from managing people to managing AI agents, and marketing is getting there first. That sentence is now said so often it has stopped meaning anything. Worse, it is usually said in a way that gets the job exactly wrong.
The common version goes like this. Agents are cheap junior staff. You delegate to them the way you delegated to people, only faster and at greater volume. The manager becomes a kind of foreman, pointing a larger crew at a larger pile of work.
MIT Sloan professor Kate Kellogg, whose research this essay builds on, points somewhere more demanding. Her studies of consultants at Boston Consulting Group, and of AI agents deployed in hospitals, keep reaching the same conclusion: the hard part is not the prompting but the managing. As agency moves from humans to machines, she argues, there is "a real increase in the importance of governance and infrastructure" to control and support those agents. In other words, managing agents is not delegation at scale. It is a different craft, closer to editing, architecture and quality control than to people management.
And the skill at its centre is one most marketers have never had to name: the supervision of ambiguity.
Billions of agents, and who they work for
The reason this applies to every job, not just marketing, is scale. We are not heading towards a world with a few clever assistants. We are heading towards one with billions of agents, and an AI for almost everything you can imagine.
This is not my speculation. At the MIT Media Lab, Professor Ramesh Raskar leads Project NANDA, which is building the plumbing for what he calls an Internet of AI Agents. The project imagines billions of specialised agents that find each other, verify each other, and work together, learning, earning and transacting on our behalf. Their research describes an internet that may soon host billions to trillions of autonomous agents, negotiating and delegating in milliseconds. Raskar's own framing is memorable: "Right now, we are in the mainframe era of AI." The PC era, with an agent on every desk and then in every pocket, is what comes next.
Follow that logic and the shape of the AI market changes. A person will not have one agent. They will have several: one for money, one for health, one for work, one for the household. An organisation will run hundreds. And it will not stop at people and companies. Cities will have agents. So will villages, communities, cooperatives and associations, each acting for the collective it represents.
That is where AI is most likely to expand. The customer of the AI economy is not just the individual human. It is every entity that can be represented by an agent. And every one of those agents needs someone to brief it, supervise it, and catch it when it goes wrong.
For marketers there is a second implication, which deserves its own essay. If entities are represented by agents, then increasingly the first reader of your pitch, your pricing page and your proposal will not be a person. It will be an agent shortlisting on someone's behalf.
People absorb ambiguity. Agents multiply it.
Here is the thing nobody tells you about a vague brief. When you hand one to a good person, it mostly works. Not because the brief was good, but because the person quietly repaired it.
They asked a question in the corridor. They remembered what the CEO said in last month's review. They noticed the deadline clashed with a product launch and flagged it. Human organisations run on an enormous, invisible layer of gap-filling. Every manager has been subsidised by it for their entire career without ever seeing the invoice.
Agents do not provide that subsidy. Give an agent an ambiguous brief and it will not stall, sulk or ask for a meeting. It will resolve the ambiguity itself, silently, in whatever direction seems most plausible. Then it will hand you something polished, confident and subtly wrong.
Now run that brief across ten agents in parallel. You do not get one wrong answer. You get ten, each wrong in a slightly different direction, each beautifully formatted. The ambiguity you left in the brief has not been absorbed. It has been multiplied.
This is the inversion that matters. With people, the risk of a bad brief is that work stalls. With agents, the risk is that work proceeds. Fluent output is the most dangerous kind of error, because it does not look like one.
Kellogg's own research found a sharper version of this. When BCG consultants tried to push back on the AI's answers, it didn't concede. It became more persuasive instead, talking people out of positions that were correct. The human in the loop, she warns, is now contested ground.
Five skills nobody hired you for
If the job is supervising ambiguity, what does it actually consist of? In my own work running marketing across several brands, largely through AI workflows, five capabilities keep separating the output that ships from the output that embarrasses.
1. Decomposition. Knowing which work can run in parallel and which cannot. Competitor research, twenty ad variants and five persona-specific landing pages can fan out happily. Positioning cannot. Positioning is a chain of judgments, each depending on the last, and parallelising it produces five positions instead of one. The first skill is knowing where the fan-out begins.
2. Specification. The old creative brief, returned with no tolerance for vagueness. A brief that worked when a colleague could ask you a question will fail when nobody can. The new brief states what good looks like, what is forbidden, and shows an example of each. If you cannot write it down, you have not decided it yet.
3. Verification design. You cannot read everything ten agents produce, so you must decide in advance where to look. Some outputs are cheap to spot-check. Some are better critiqued by a second agent. And some, such as claims, pricing, customer names and anything with legal exposure, are only ever approved by a human. Deciding where the checkpoints go is now a strategic act, not an administrative one.
4. Merge judgment. Parallel work drifts. Five agents writing for five segments will invent five slightly different brand voices and three contradictory proof points. None of them is wrong in isolation. The error lives only in the combination, which means only someone looking at the whole can catch it. This is the new editorial job.
5. Restraint. Knowing when not to parallelise at all. When capacity is free, the temptation is to generate fifty options because you can. But fifty options is not a strategy, it is a delegation of the choice back to yourself, at a worse moment and with more noise. Sometimes one careful pass beats ten fast ones.
The three-line problem
One caution, from my own team today. A colleague handed a task to AI, and the code it wrote broke. The fix was to read it and delete two or three lines. Instead, he kept asking the AI to repair its own work. The issue is still open.
Kellogg's research has a name for this. In her study of consultants at Boston Consulting Group, some became self-automators, offloading the whole task to AI. She contrasts them with centaurs, who keep AI on a tight leash guided by their own expertise, and keep building that expertise as they go.
AI produces a great deal of work, and some of its mistakes it cannot fix itself. Someone has to be able to read the code, check the number, or hear the false sentence. A manager who can only re-prompt is not supervising the agent. Delegation without the ability to intervene is not management. It is dependence.
Sources
