In 1895 the Ponemah textile mill installed electric motors next to the steam engines they replaced and connected them to the same shafts and belts. Power changed. The factory did not. The productivity gains of electricity took another thirty years, until companies redesigned the factory around what the new motor made possible.

Most companies are doing the same with AI today. They buy the tools, adoption numbers go up, and the way work flows from customer problem to shipped software stays exactly as it was in 2015. The result is the same code, produced a bit faster, by a process built for a different cost structure.

This talk is about the second move: redesigning how an established company builds software once code stops being the hardest part.

I lead AI Platform Engineering at BetterUp, a 15 year old company with 1000 employees and a real legacy estate, where we went from AI tooling adoption of 15% to 85% in a year, moved a quarter of infrastructure changes to agents, and, harder, changed who decides what gets built and how.

I will cover what we changed in the operating model rather than the toolchain, where the resistance really came from, the metrics that mattered to the CFO versus the ones engineers like, what we got wrong in the first six months, and why companies over ten years old should now treat rebuilding their core systems as a business decision, not a technical one.

Honest field notes, not predictions. For CEOs, business leaders and anyone who owns a P&L that depends on software being built.

You will leave with: a way to tell whether your company is getting the small local gain or the big one; a concrete operating model shift with the metrics that changed, the ones that did not, and the mistakes; and a way to decide when “never rewrite” stopped applying to you.