AI is not only how you build — it is increasingly what you build. The CTO and CPO need one shared view of AI in the product.
Division of labour
- The CPO owns which problems AI should solve for customers, the experience, pricing and success metrics.
- The CTO owns feasibility, architecture, the platform (gateway, retrieval, evals, observability), cost, security and reliability.
- Together they own the quality bar: what accuracy, latency and failure behaviour is acceptable before launch.
Questions to settle together
- Value: which customer outcome improves, and how will we measure it?
- Quality: what does the eval set look like, and what score is good enough to ship?
- Failure modes: what happens when the model is wrong — fallback, human handoff, disclosure?
- Unit economics: inference cost per user or per action at expected volume, and the pricing that covers it.
- Dependency: which model providers do we rely on, and what happens if one changes terms or capacity?
Vendor dependency is a product risk
In June 2025 the AI coding company Windsurf said Anthropic had cut off nearly all of its direct capacity to Claude 3.x models with less than five days' notice, amid reports that Windsurf was being acquired by OpenAI (TechCrunch, June 3, 2025). Windsurf's users saw short-term availability issues; some switched tools or brought their own API keys. An AI product that depends on one provider inherits that provider's business decisions.
Mitigations: a gateway that can route to a second model, evals that tell you quickly whether an alternative is good enough, contracts with notice periods, and a product design that degrades gracefully.
Operating rhythm
A joint monthly AI product review: quality and eval trends, cost per unit, incidents, provider changes and roadmap. Shared metrics stop "the model is great" and "the feature is too expensive" from being separate conversations.