Everyone's racing to ship AI features.

Spent 3 years watching teams nail the demo, then crash in production.

The pattern? They build AI like traditional software.

But AI products break every assumption:

• Non-deterministic behavior (same input ≠ same output)

• Agency-control tradeoff (more autonomy = less visibility)

Aishwarya Naresh Reganti and Kiriti Badam developed CC/CD after 50+ AI implementations at OpenAI, Google, Amazon.

Their framework: Start high-control, earn agency gradually.

v1: Route tickets (observable, reversible)

v2: Suggest responses (human review)

v3: Auto-resolve (earned autonomy)

Most teams jump straight to v3. Then spend months debugging a black box.

What AI capability are you rushing to ship that should start smaller?

Full framework: