AI value shows up in the hiring plan before it shows up in the P&L.
Over the last eighteen months I have worked on AI, architecture and technology strategy across five companies, some full-time, some fractional. Three results keep coming back.
The sceptic has a fair case. Vendor productivity claims are marketing. METR's controlled study of experienced developers working in their own repositories found they were slower with AI tools, while believing they were faster. And "we did it with fewer people" is an anecdote with no control group.
Mine are anecdotes too. The baselines are estimates from earlier projects, the scopes differ, and the teams were senior. Those are the confounders, stated up front.
It is as much a productivity story as it is a team-shape story. Productivity measures how fast a person works. Team shape measures what the organisation had to fund. Only the second reaches the P&L.
The compression was uneven. It landed on well-specified work: schemas, pipelines, APIs, feature engineering, model training. That work has a known shape, abundant prior examples, and outputs a machine can test. Models perform well on well-trodden ground, and most data engineering is well-trodden ground.
Three things changed in the teams that remained.
Everyone had to be high-agency. With fewer hands, each person owns a whole slice: decides, builds, verifies. There is no layer translating specifications into tickets, because there is nobody to hand the tickets to. Agency mattered more than years of experience.
Verification became the work. Generation is cheap. Knowing whether the output is right is the cost. The teams that held quality spent their effort on automated tests, review separated from build, and guardrails at execution.
The constraint moved. Once data work compresses, the slow parts are the ones that need agreement: what the product should do, who owns the outcome, which process has to change. Those resist compression because they are made of people deciding things.
That raises the obvious objection: if every seat needs agency, where do juniors learn? My answer was high-agency interns from university. I coached them directly, and I taught them how to be coached by the model as they built. I value coaching immensely, having learnt a great deal from the CTOs and CEOs who agreed to coach me, so passing it on has been one of the most rewarding parts of this work.
The junior pipeline survives, in a new shape. The apprenticeship now has two coaches: a person for judgement, a model for the hours in between. The future belongs to the very bright young people who learn to work that way.
Report AI value as the plan you didn't need to fund. Put the team you would have hired next to the team you did, hold scope constant, and publish the confounders beside the number.
Hours saved are a claim. The same scope delivered by a smaller team is a result, and it is the one a board can check.