Microsoft: why faster AI coding did not speed up the team

There is a useful follow-up to my posts on spec-driven development and AWS AI-DLC. On September 3, 2026, Microsoft Digital, the company's internal IT organization, shared its experience adopting SDD. It opens with an uncomfortable admission.

Developers learned to use AI tools and became faster. But, according to the initiative's leader, team productivity did not improve. The authors trace this to intent getting lost as work passes between participants. Everyone accelerated their part of the process; agreeing on what to build did not become easier.

They reorganized development around a living specification: business goals, user scenarios, edge cases, and acceptance criteria, versioned in the repository and updated with the product. Before writing specs, the team agrees on a constitution covering architectural principles, security requirements, and constraints.

GitHub Spec Kit supports their workflow: define the problem → clarify questions → plan → break work into tasks → check alignment → implement and test. They recommend small, focused specs whose outputs are easier to review and refine incrementally.

Roles change too. PMs own the specification throughout development; engineers spend more time on requirements, plans, and validating generated output. A familiar theme: AI pushes forward the work that used to wait until the first “we actually meant something else.”

However, this article provides no before-and-after measurements or control group. It describes Microsoft Digital's experience, not evidence that SDD accelerates every team.

This brings me back to the discussion of AI metrics: measure the entire journey from task to accepted outcome, including alignment, review, and rework. Otherwise, code can be written very quickly while useful changes still take a long time to ship.

#AI4SDLC #AI #Engineering #Management #DevTools