Drop-off looked like a UX problem. It wasn't. How segmenting by risk instead of redesigning the screen lifted onboarding completion ~25% with zero increase in fraud.
Razorpay's CPO described an internal tool where AI agents draft PRDs, specs, and UAT plans in parallel. I tried to build a version of it. The idea held up better than the implementation.
The production work is disappearing fast. What's left is judgement, and judgement is getting scrutinised harder than ever.