Manjunath Sindagi says AI product managers need eval loops, not PRDs

Manjunath Sindagi, writing in The AI Craft, argues that product management built on deterministic logic breaks down once systems generate probabilistic outputs rather than execute fixed rules. He replaces the traditional PRD with four competencies: judging when a problem needs AI at all, designing retrieval pipelines that ground models in proprietary data, building automated evaluation harnesses against curated 'golden datasets', and designing interfaces for outputs that can be wrong with confidence. The piece frames failure itself differently: instead of a stack trace pointing to broken code, a system can hallucinate a refund policy while functioning exactly as built, which is why Sindagi insists PMs need routing thresholds that escalate low-confidence outputs to human review rather than shipping them.