AI Changes the Delivery System — and the Product Itself

We accelerated code. Not the system.

AI can generate software faster than teams can review, understand, integrate, and recover it. That does not automatically make delivery faster — it can simply move the bottleneck.

But that is only the first change.

When AI moves from the developer’s toolchain into the runtime product itself, teams face a different engineering problem. Consequential behavior may now depend partly on probabilistic model judgment. Requirements become boundaries around acceptable behavior, one green test is no longer sufficient evidence, and release decisions must account for evaluation uncertainty, human authority, fallback, and operational control.

This talk connects both sides of the shift.

We will examine recent evidence on AI-assisted development and separate code activity from delivery outcomes such as stability, maintainability, review capacity, recovery, and team understanding. Then we will look at what teams can actually measure and what emerging countermeasures — spec-driven development, agentic review, repository intelligence, generated documentation, stronger knowledge systems, and disposable implementation — are worth testing.

Finally, we will move inside the product itself: behavioral requirements, evaluation evidence, release criteria, and control loops for systems that rely on model judgment at runtime.

This is not a proposal for a new “correct AI SDLC.” Many of these practices are still hypotheses. The practical question is how to preserve engineering rigor when machine capability changes faster than the surrounding engineering system can adapt.