What actually changed in software delivery, 2023–2026

Three years into the generative-AI cycle, the honest question for anyone running delivery is not "is AI real?" — it plainly is — but "where has it actually changed the economics of building software, and where is the change still a promise?" This report separates the two.

Where the needle moved

The clearest, most durable gains are in the first draft of everything: code, tests, migrations, documentation, and the boring glue that used to eat senior time. Teams that have instrumented their own delivery report faster cycle times on well-scoped, well-tested work — not because the model is a better engineer, but because it removes the activation energy of starting.

The second real gain is comprehension: reading an unfamiliar codebase, explaining a failure, or turning a regulation into a checklist. These are retrieval-and-summarise tasks, and they are exactly what the current generation is good at.

Where it hasn't

Adoption has run far ahead of value. The uncomfortable pattern we see across estates is that the tools land, individual productivity feels higher, and yet lead time to a governed, shippable change is unchanged — because the constraint was never typing. It was decision-making, review, approval, and the evidence a regulated organisation needs before it can deploy.

AI does not remove that constraint. Left ungoverned, it makes it worse: more code, more options, more claims, and no more capacity to decide between them or to prove afterwards what was decided and why.

The through-line

The organisations getting compounding value have one thing in common — they treated adoption as a governance problem, not a tooling problem. They made decisions, evidence and runtime assurance first-class, so that going faster did not mean losing the thread. That is the whole thesis of this firm, and it is why we publish the evidence for these claims rather than asserting them.