The Two-Person Company: Building With AI Agents Instead of Headcount

What actually changed

The cost of a competent first draft of almost anything fell to approximately zero, and the cost of routing work between steps fell with it. Neither of those is the same as the cost of being right falling, which is the distinction most writing on this subject skips.

What that unlocks is specific rather than general. In businesses that are largely digital, repeatable and instrumentable, a very small team can now operate at a volume that previously required a department. In businesses that depend on physical operations, regulated judgement or deep relationships, the change is real but marginal.

What agents replace: coordination, not craft

The hours that disappear are not the skilled ones. They are the connective tissue: chasing status, reformatting the same information for a different audience, triaging an inbox, summarising a call, moving a record from one system to another, checking whether something changed.

This matters because coordination is where small companies actually spend their time, and because it is the work that traditionally justified the first three hires. A company that removes coordination overhead does not become a company with fewer skills — it becomes a company where the same two people spend a much higher proportion of the week on work that requires them.

The work that does not compress

• Deciding what to do. Stage 01 judgement — which market, why now, what to refuse — does not compress at all, and it is the scarcest input in a studio’s economics.
• Anything with accountability attached. If a person or a regulator needs someone to answer for it, a person has to own it.
• Judgement that depends on undocumented context. The reason this client gets an exception and that one does not.
• Original information. First-party data, real experiments, things that happened. This is also exactly what makes content citable rather than restated — the point of answer-shaped content.
• Relationships. Nobody wants a durable commercial relationship with a queue.

Where agents fail

Predictably, and the pattern is not difficulty. Agents handle difficult work well. They fail where failure is hard to detect.

The verification tax

Every agent that produces work also produces review, and reviewing fluent output is harder than reviewing obviously rough output — because the errors do not announce themselves. This is the cost that never appears in the productivity claim.

Definition

Verification taxThe time cost of checking machine-generated output. Because errors in generated work are fluent rather than obvious, review is slower than for human work — and any honest measure of automation leverage subtracts verification time from time saved.

The practical rule that follows: automate tasks whose output is cheap to verify. A generated draft you were going to edit anyway is nearly free to check. A generated financial summary, a legal paragraph or a customer-facing claim is not, and the automation can easily cost more than it saves while feeling like progress.

The second-order effect is worse and less discussed: a team that verifies badly for six months accumulates errors it cannot locate, because it never knew which outputs were wrong. Cheap verification is not a nice property. It is the selection criterion.

How we actually run it

1. Instrument before automating. If a process is not measured, an agent will make it faster and nobody will know whether it got better. This is stage 05 following stage 04 in the operating system for exactly this reason.
2. Automate the shape, not the judgement. Agents assemble, extract, route, monitor and draft. People decide, commit and take responsibility.
3. Keep chains short. Two or three steps with a checkable output at each, rather than an elegant eight-step pipeline whose third step quietly went wrong in March.
4. Write the context down. Most agent failures are context failures, and the document that fixes them is the same one that makes the business transferable — the preparation described in the succession piece.
5. Review the automations quarterly. Automations rot. A workflow built against a system that changed six months ago is now producing confident nonsense on a schedule.

What it means for hiring

Not “do not hire”. Hire later, fewer, and for different reasons. The roles that used to be first — coordinator, junior analyst, first support hire — are precisely the ones that compress, so the first real hire moves toward someone with judgement and broad scope who can own an outcome rather than a queue.

There is an honest cost to that which the genre never mentions: the traditional first roles were also how people entered an industry and learned it. A company that never hires a junior is a company that never trains one, and that is a real trade the sector has not yet reckoned with.

How we think about it at Thinktier

Automation is stage 05 rather than a philosophy, and it exists to raise the ceiling on how many companies a small team can genuinely operate — the constraint that decides whether a portfolio compounds or quietly decays.

The test we apply before automating anything is deliberately unglamorous: can we verify the output in less time than the task takes? If not, the work stays human, however tempting the demo was. Most of what has actually worked for us is unremarkable — pipeline hygiene, reporting, first-draft production, monitoring — and unremarkable is the point. The interesting part of a company should be the part a person is doing.