Aishwarya Reganti and Kiriti Badam — having shipped AI features at OpenAI, Google, Amazon, Databricks, and Kumo, and consulted with 50+ companies — argue that PMs are dragging traditional-software muscle into a domain where it breaks. Their framework: Continuous Calibration / Continuous Development (CC/CD).
The two assumptions that trip up most teams:
- AI products are inherently non-deterministic. 'AI models are trained to generate plausible responses based on patterns, not to follow fixed rules. The same request can produce different results depending on phrasing, context, or even a different model.' You're no longer designing predictable user flows; you're designing for likely behavior.
- Every AI product negotiates a tradeoff between agency and control. Too much agency too fast erodes trust; too much control kills the magic. 'Every new capability shifts control further away from humans.'
The loop:
- Development (CD) — scope the problem, design architecture, set up evaluations to keep non-determinism in check. 'Start with features that are low-agency and high-control, then gradually move up as the system proves it can handle more.'
- Deploy — not a finish line. A transition.
- Calibration (CC) — observe real behavior, figure out what broke, ship targeted fixes. Every cycle earns the system a bit more agency.
The failure mode they keep seeing: a great demo, glowing early reviews, then 'a system that can't scale or sustain' — and user trust quietly erodes before the team realizes it's happening.
For a mid-level PM, the practical translation: stop writing PRDs that end at launch, start with the smallest agency surface you can defend, build evals before shipping, and treat post-launch calibration as the actual job, not a bug-fix backlog.