The workflow is the moat

A 30-developer project, scoped at 12–18 months. Delivered in 76 days.

That’s not a vendor claim — it’s the Amazon Bedrock team’s own inference engine, one of the experiments Swami Sivasubramanian published from hundreds of Amazon engineering teams working AI-natively. The median team saw 4.5x productivity. Some passed 10x.

Here’s what I find most useful in that data: the gains didn’t come from giving engineers AI tools. Teams that bolted AI onto their existing workflow saw modest improvements. Teams that redesigned the workflow around AI — specs as source of truth, agents working in parallel, humans concentrated on judgment calls — saw the multiples.

As Swami put it: “Frontier teams are not just using AI to code faster. They’re redesigning how software gets built.”

This is exactly why embedding matters in enterprise AI. You cannot ship a workflow redesign in a zip file. Someone has to sit with your team, restructure how work flows, and prove it on your actual backlog. That’s what forward deployed engineers are for — they arrive with the redesigned workflow already lived-in, because Amazon ran it on itself first.

The tooling is table stakes now. The workflow is the moat.