I've been estimating software projects for nearly three decades, so when AI started making some work dramatically faster, I paid close attention, because "some" is hard to estimate.

In controlled studies, developers finished certain tasks 55% faster with AI while other work came out 19% slower, and my own experience has back this up. It turnes out that AI accelerates work that is well-defined and well-guarded, and it punishes ambiguity: vague requirements become confident code that solves the wrong problem, missing tests let rework pile up quietly, and a shaky release pipeline throttles however fast the code arrives.

That means the fundamentals we've preached for twenty years (requirements gathering, automated testing, CI/CD) didn't get less important in the AI era. They became the thing your estimate depends on.

In this talk we'll look at the real numbers on what gets faster and what gets slower with AI, how to adjust your estimates for both, and how to use AI itself to close the fundamentals gaps that make your timelines unpredictable in the first place.