AI product strategy

AI product strategy is not "add a chatbot." It is choosing which jobs tolerate probabilistic systems, what quality bar makes those jobs shippable, how inference cost hits margins, how you defend the product when models commoditize, and how you price so power users do not bankrupt you.

The interaction patterns live in AI product UX. This page is the decision layer. Treat the model like a powerful, unreliable junior employee: hire it for the right tasks, measure its work, put guardrails on irreversible actions, and do not build the company's identity on which vendor badge it wears this quarter.

As of when this was written, foundation models improve quickly and prices fall in fits. Strategy that bets only on a clever prompt expires. Strategy that bets on workflow, data, distribution and eval discipline compounds.

Concepts in this section

Read them in order.

At work

Engineers on AI features should explain the quality bar, the COGS risk and the positioning in one minute. Interview and design-review signal: "Here's the verifier, here's what we measure before auto-send, here's how a 10× power user hits our margin, here's why this isn't a wrapper." If you cannot answer those four, you are not ready to flip the feature flag — no matter how good the demo looks on stage.

A $30 seat with 200 agent runs/day at $0.05 each is ~$300/month COGS. That is a product problem (caps, credits, cheaper models), not an infra surprise. Write the eval bar in the PRD so prompt tweaks are measured, not argued.

Do not hard-code today's failure modes as permanent UX, and do not wait for a hypothetical future model. Ship the copilot now. Gate the agent behind quality bars.

Common mistakes

• Adding a chatbot to a job that needs rules. If search or a rules engine solves 90%, start there.
• No verifier, no bar, auto-send on. That is how you buy an incident.
• Unlimited seats on bursty agents. One power user inverts gross margin.
• Competing on the prompt. Wrapper risk. Defend with workflow and data.

Key takeaways

• Choose AI jobs with high value, tolerable errors and verifiers.
• Position honestly: copilot vs agent vs automation implies different UX and liability.
• Write eval bars into the PRD. Watch inference COGS per outcome. Price work, not tokens.
• Defend with workflow, data, integrations and distribution.

Further learning

YC-adjacent founders and investors debate AI startup pricing and value capture — useful framing for "wrapper vs product."

The New Business of AI (a16z)At a technical level, artificial intelligence seems to be the future of software. AI is showing remarkable progress on a range of difficult computer science problems, and the job of software developers – who now…Andreessen Horowitz You are not a model. Don't price per token (a16z)Charging per token is becoming the default for AI companies. For most AI applications, it is a mistake.Andreessen Horowitz AI Engineering (Chip Huyen, O'Reilly)oreilly.com