The economics of AI-first development
Seat prices are the smallest line item. Token economics, review cost, and the real cost of a bad rollout, with the numbers from ours.
When leadership asks what AI-first development costs, they usually mean seat prices. Seat prices are the smallest line item. Here is how the economics actually worked for a 30-engineer organization.
Tokens, not seats
During evaluation we ran the same real tickets through Claude, Cursor, and Codex and logged the tokens each one burned to reach a mergeable result. The cheapest seat was not the cheapest tool. A tool that needs three retries to get a change right costs more per merged PR than a pricier one that gets it in one pass, and it costs engineer attention on top of that. Token economics at team scale was one of our four evaluation criteria for exactly this reason.
Review is where the cost moved
Generation got cheap, so review got expensive. Our first month produced pull requests two to three times the size of the previous year's average, and review time per PR rose to match. The fix was structural: PR size caps enforced in the merge rules, AI-assisted first-pass review to catch the obvious issues before a human looked, and a norm that the author summarizes what the tool did and what they changed. Review cost came back down, and defect escape rate stayed flat.
The pilot paid for the year
Three projects, estimated at 12 to 18 engineer-months, delivered in six weeks by three engineers. Take the conservative end, 12 engineer-months, and subtract the 4.5 engineer-months actually spent. That is 7.5 engineer-months returned in the first six weeks, which covered the tooling budget for the organization for the year before the rollout started.
The cost nobody budgets
A failed rollout costs more than the tools. It burns the team's willingness to try again, and the second attempt starts from below zero. Guardrails, manager adoption, and honest expectation setting are not overhead on the transformation. They are the cheapest insurance you can buy against having to do it twice.
What I would tell a CFO
Budget for tokens, not seats. Budget for review capacity in month one. Expect the pilot to pay for the program, and expect the range of gains by work type to land between 50% and 90%. Then hold your managers accountable for using the tools themselves, because that is the variable that decides whether the numbers hold.