The unit of work is changing

Most founders are trained to translate work into roles.

If product requirements are unclear, hire a product manager. If code review is slow, add senior engineering capacity. If decks take too long, find a designer or operator who can package the thinking.

That model still makes sense. But AI agents are starting to change the first question a founder asks.

The question is no longer: who should own this?

Increasingly, it is: what workflow are we trying to run?

The unit of work is changing

The most useful shift from AI agents is not that they behave like perfect employees.

Repeated knowledge work can now be broken into smaller, callable workflows.

In software and product work, that looks surprisingly ordinary: commit messages, pull request reviews, bug analysis, product requirement drafts, idea-to-outline flows, outline-to-deck creation. None of these tasks replaces a full person. But each one removes a small recurring loop that previously consumed attention, coordination, and context switching.

The one-person company is often described as a founder with a set of AI employees. I think that framing is too simplistic. A more accurate version is a founder with an operating system: a set of structured workflows that can be triggered, reviewed, improved, and composed over time.

Why this is not just hype

McKinsey’s 2025 State of AI survey found that 62 per cent of respondents said their organisations were at least experimenting with AI agents, while 23 per cent said they were scaling an agentic AI system somewhere in the enterprise. At the same time, nearly two-thirds had not yet begun scaling AI across the enterprise.

That gap is not an awareness problem. It is an operationalisation problem.

Gartner made a similar point, predicting that 33 per cent of enterprise software applications will include agentic AI by 2028, while also warning that many agentic AI projects will be cancelled when value is unclear or implementation is too costly.

Capability alone is not enough. The hard part is turning capability into a workflow people can trust.

The wrong frame is replacing people

One mistake is to ask which junior role an AI agent can replace. The question is tempting because it maps new technology onto an old cost structure, but it often leads to bad design.

A junior employee does not only complete tasks. They learn context, notice exceptions, ask questions, escalate uncertainty, and build judgment. An agent does not automatically do those things because it can produce an output.

A better question is: which repeatable parts of the work can be made more structured?

• Can the agent produce the first draft?
• Can it summarise the issue?
• Can it check the diff against a known standard?
• Can it prepare the review packet before a human makes the decision?

That is where the economics start to change. The goal is not to pretend a company needs no people. The goal is to spend less human attention on procedural work, so more attention can go toward judgment, taste, relationships, and accountability.

The founder’s job becomes system design

If workflows become the unit of leverage, the founder’s job changes.

It becomes less about assigning tasks one by one and more about designing the conditions under which work can move reliably. That means defining the input, the expected output, the boundary of the agent’s authority, the review point, and the escalation path.

For example, an agent can help analyse a bug report. But should it create a ticket, assign priority, propose a fix, or touch the codebase? Those are different levels of delegation. These questions sound operational, but they are strategic. They decide whether agents make a team faster or create more output that someone else has to clean up.

The lean company still needs judgment

AI compresses execution faster than it replaces judgment. That means a small team can now do more. But only if the people in that team are clear about what should and should not be delegated.

The most reliable agent workflows are narrow. They have a clear job, clear context, and a clear definition of done. Broad autonomy sounds impressive in a demo. Narrow reliability is what survives in production.

This is also where trust enters the picture. Teams do not adopt agents because they are powerful. They adopt systems they can inspect, predict, and override. Logs, review points, consistent outputs, and the ability to see why something happened all matter. Without those pieces, an agent becomes another source of uncertainty.

What changes from here

The lean, agent-powered company of the future will not simply be a tiny company doing the work of a large one. It will be organised differently.

Instead of building only around roles, founders will build around repeatable workflows. Instead of asking whether a task needs a person, they will ask what part of the task requires human judgment. That changes what hiring is for.

The most valuable people in these companies will be the ones who can define work clearly, recognise quality, make trade-offs, and design systems that improve over time.

My practical observation is simple: before asking whether an AI agent can do a job, write down the workflow.

If the workflow is vague, the agent will expose that vagueness. If the decision boundary is unclear, the agent will make the ambiguity more expensive. If nobody knows what good looks like, faster output will not help.

But if the workflow is clear, repeatable, and reviewable, the economics of building a company start to change. Not because one person can do everything. Because one person can start assembling the system through which more work gets done.