Imagine a CEO tells me: “AI agents are now capable enough that I can remove half my management layer.”

My direct response would be: you’re going to be broke in six months.

That may sound dramatic, but the premise is crazy. AI is already disrupting how work gets produced. It can retrieve and structure information, analyse large amounts of data, generate code, create drafts and automate repetitive processes at extraordinary speed. I use LLMs every day, and I automate as much as I responsibly can.

But management is not the production of information.

Management is deciding what matters, understanding people, connecting work across a business, making choices under uncertainty and remaining accountable for the result. It is also knowing when the numbers are incomplete because something human is happening underneath them.

AI can make a good manager faster. It can also reveal very quickly when someone was never managing clearly in the first place.

AI can calculate the odds. Management plays the table.

Poker is a useful analogy.

An AI system can calculate the odds of the cards that may hit the deck. A manager also has to play the players: confidence, fear, incentives, history and the emotions in the room.

Those forces affect business outcomes enormously, yet they are rarely captured properly in a dashboard or meeting transcript.

I remember a situation where the company’s data, structure and history all pointed toward staying the course. The organization knew how to operate in a particular way, not only in that country but across the group. The formal evidence supported doing more of what had worked before.

But the room was not aligned. It was afraid.

The team believed change was necessary, but management was present and nobody felt confident enough to say it. Silence could easily have been interpreted as agreement. It was not agreement; it was self-protection.

Normally, I want the team to bring ideas forward. In that moment, leadership meant speaking first. I introduced the idea, supported it with argument and precedent, and proposed a measured way to test it. Once the silence was broken, the organization could discuss what people had already been sensing.

We eventually did something that had never been done in the history of the group, and it paid off.

The point is not that leaders should always override the established playbook. The point is that they need to recognize when the apparent consensus is false. A model reviewing the minutes might conclude that everyone agreed. A manager has to understand why nobody disagreed.

AI punishes ambiguous management

One of the best things LLMs have done for me is force me to communicate more clearly.

You cannot leave important assumptions unstated. You cannot assume the model has assembled the history in your head, understood which constraint matters most or inferred the result you actually want. Even with structured documentation, project folders and specialized agents, you need to be precise.

If the instruction is poor, the output usually exposes it quickly. I can lose an entire morning because I failed to provide the right context or supervise the work closely enough. That is frustrating, but also educational.

Human teams suffer from the same ambiguity. They simply hide it better.

An employee may receive an unclear brief, say they understand it and then run with their own interpretation. The manager assumes everything is fine. The misunderstanding only becomes visible days or weeks later, once the work returns.

With an LLM, that feedback loop is compressed. Poor instruction becomes visibly poor output almost immediately. The model forces the manager to confront a basic question: did I communicate clearly, or did I merely assume I had?

That is how AI reveals management.

A task is not an outcome

Consider a product instruction such as: “Add another button to the payments menu.”

That tells someone what to build. It does not explain the problem being solved.

A proper managerial brief would be closer to this: investors need to understand the payment options available when they invest, the difference between using fiat currency and stablecoins, and how each route works through a card or a wallet. The experience must make those choices clear while remaining coherent with the rest of the platform.

Now the designer, developer or AI agent understands both what needs to happen and why it matters.

There should still be room for their own sauce. The manager does not need to prescribe whether the solution is a button, a pop-up or another interaction before the work has even begun. Good delegation defines the intended human outcome, the important constraints and the standard of success. It leaves appropriate freedom over how to achieve it.

The weak manager delegates activity. The good manager communicates purpose.

The human is the integration layer

The best results I have achieved with AI have come from breaking complex work into smaller components and using specialized agents to execute them. I currently work with an AI setup of eleven agents, including one whose sole job is to attack the code and try to break it.

That team can produce a fantastic amount of work. But it cannot simply be left to run.

My job is to be the bridge: define the architecture, decide how the work is divided, supply the right context, supervise the output and bring the pieces back together. In software development, I can allow LLMs to do a large share of the coding. I cannot delegate the responsibility for whether the product makes sense, remains coherent and solves the intended problem.

This is where the analogy between managing agents and managing people works. Both need context, defined roles, standards, handoffs and oversight.

But there is also a major difference. People carry lived experience, relationships and institutional memory. An AI agent works from the context made available to it. A system can preserve documents and retrieve earlier information, but somebody must still determine which history matters and how it should influence the present task.

Agents produce components. Management creates coherence.

The agent can produce. It cannot release.

AI is not perfect, and a manager cannot treat its confidence as proof.

Models can make assumptions, move outside the requested scope or claim that something was verified when it was not. Sometimes a bad output is clearly the consequence of a bad instruction. If I ask for a button without giving visual or product constraints and receive a terrible neon-green button, that failure is mine.

Other times, the model takes an unjustified shortcut. It answers from prior context instead of inspecting the current code, or presents an inference as confirmation. Those failures require explicit rules, validation and supervision.

My clearest rule is visual first. Anything an investor or client will see must be shown to me and approved before it is coded or reaches production. The agent can propose and produce; it cannot release.

The same rigor applies to security and software quality. Code needs to be robust, audited and tested at an enterprise level. That is why one of my agents exists specifically to attack what the others build.

But even with multiple layers of AI review, the final responsibility remains human. An agent should never be trusted to certify the reliability of its own work without independent verification.

AI enhances judgment. It does not manufacture it.

I am optimistic about LLMs. They help me work faster, analyse more information and conceive better products. I do not see them as tools that will simply take all work away.

If AI puts immediate pressure anywhere, I expect it to be on junior production work that is repetitive, structured and easier to verify. Experienced managers and decision-makers should become more capable because of it — but only if they learn to direct it properly.

The management layer should change. Roles built mainly around moving information, compiling reports or coordinating activity without adding judgment will be challenged. That is healthy.

But removing management because agents can generate work confuses output with leadership.

The companies that benefit most from AI will not be the ones that eliminate human responsibility. They will be the ones that combine machine speed with human context, psychology, judgment and accountability.

AI does not remove the need to manage.

It makes the quality of management impossible to hide.