AI that makes one person faster can just as easily create a bottleneck for everyone downstream of them. The leverage I’m after isn’t personal productivity — it’s what changes when a whole team’s thinking is amplified together.
Treat it like onboarding, not like a tool
Bounded tasks first. Judgment observed. Scope expanded as trust is earned — not handed the keys on day one.
That’s the same sequence I’d use with a new teammate, and for the same reason: you don’t yet know where its judgment fails, and finding out on something that matters is expensive.
Context has to be explicit and shared
The failure mode isn’t the model. It’s context trapped in one person’s head or one person’s prompt history, which makes the output unreviewable by anyone else.
So: business logic gets defined before AI is asked to help build it, and the rules and skills that shape AI behavior are versioned alongside the code — not bolted on after, not living in a chat window.
What it looked like in production
A multi-agent development workflow across Claude, Codex and Gemini with explicit roles — a context-extraction agent, parallel Challenger and Risk agents, a red-green-refactor cycle gated by human approval.
It cut time to market from months to weeks on the work it touched (full case study →). The part that mattered more: it changed how the broader team worked with AI, not just how I did.