Headcount Is Not a Unit of Work

A claim that AI performed the work of a large engineering team compresses several different questions into one dramatic number. Was the comparison based on lines written, tickets closed, hours avoided, features released, or software still working after customers touched it? Each answer describes a different machine. Without the unit, the multiplier is a mood.

This does not make the underlying change fake. Agents can execute repeatable delivery work at a scale a person cannot match. They do not sleep, wait for a free afternoon, or forget the standard between runs. But raw generation is the easiest part to multiply. The expensive question is how much generated work advances through specification, review, verification, release, and production without being rebuilt by people later.

Headcount is an input to an old operating model. Using it as the output measure for a new one keeps the old model in charge of the story.

If the capacity is real, there is a mechanism. Work enters in a form the system can act on. Agents receive bounded context and permissions. Standards are enforced on every change. Independent checks decide what advances. Failed work returns with evidence instead of disappearing into a meeting. Production behavior becomes an input to the next run.

That is a software factory. It is not one model producing a heroic pile of code. It is a governed route that converts intent into accepted software repeatedly, including when the preferred model changes or the person who built the first route leaves.

Ask to see that route. Ask where work is rejected, who owns the exception, and whether the evidence can be reproduced somewhere the generating agent did not control. A large capacity claim with no visible acceptance system usually means humans are still doing the integration, cleanup, and risk absorption off the books.

The labor consequence arrives when the route becomes executable. Once a factory can take a defined class of work from request through proof, the people who manually carried that class of work are no longer required at the same scale. Ticket translation, routine implementation, repeated testing, release coordination, and status reconstruction do not remain jobs merely because the model still makes mistakes. The factory exists to catch and route those mistakes mechanically.

Human judgment remains at the consequential boundaries: choosing what deserves to exist, defining the standards, accepting risk, and deciding whether the result is right. That is not a promise that every current role survives with a more strategic title. Fewer people can direct more productive machinery, and budgets will eventually price the difference.

Our position is that companies should say this plainly. Hiding displacement behind an impressive headcount analogy turns an operating change into theater. The honest question is which work has become machinery and which decisions still require an accountable person.

A useful factory measure begins with accepted outcomes. How much requested work reached production? How much passed independent verification on the first route? How often did production evidence send it back? What did retries, model calls, tools, and human intervention cost? How long did the result remain inside its reliability and policy boundaries?

Those measures reveal whether automation is compounding or merely moving labor. A fast generator followed by a human repair queue is not autonomous delivery. A cheap model whose work burns more verification and retry capacity may be the expensive worker. A release count that ignores rollback and production correction rewards the factory for creating its own demand.

Put the meter across the whole route. The point is not to find one flattering number. It is to identify the constraint, improve the machinery, and retain that improvement for every later run.

Start with one bounded class of work. Define the destination in plain language, the evidence required to advance, and the conditions that force a stop. Record every handoff automatically. Separate the agent that creates the change from the system that judges it. Feed failures back as durable rules rather than reminders somebody must carry.

Then remove the human coordination the route no longer needs. Do not preserve a manual approval, status meeting, or ticket-moving role just to make the transition feel familiar. Keep people where judgment changes the destination or owns a consequence. Put repeatable execution into the factory.

AI will do work that once required large engineering teams. We believe that direction is already set. But the winner will not be the company with the loudest headcount equivalent. It will be the company whose smaller group can direct a governed system, prove what survived, and run it again without rebuilding the story around another heroic number.