🎯Your AI Product Strategy
How to think about adding AI to your product without falling into the 'AI feature graveyard' trap.
Most companies' AI strategies are reactive — 'we need to add AI' — and produce features that don't get used. A real AI strategy starts from the customer job and uses AI as a means, not an end.
A great AI strategy answers four questions: (1) What customer job does AI uniquely enable? (2) What's our defensibility? (3) What's the cost / quality / latency target? (4) How do we measure success? AI is a means; the customer outcome is the end. Strategies that start with 'we'll add AI to X' fail; strategies that start with 'customers struggle with Y, AI can solve it' succeed.
The four questions
1. What customer job does AI uniquely enable? Don't add AI to features customers don't already use. Find the jobs where AI removes real friction:
- Synthesis (long documents → summary)
- Generation (blank page → draft)
- Search (keyword → semantic)
- Personalization (one-size → tailored)
- Decision support (data → recommendation)
If the job doesn't change qualitatively with AI, the AI feature won't get adopted.
2. What's our defensibility? This is the existential question for AI products. If you're a thin GPT wrapper, you have no moat. Defensibility comes from:
- Data. Proprietary data the model needs to be useful
- Distribution. Existing users you can deliver AI to
- Integration. Deep ties to user workflows
- Eval discipline. Real quality measurement that competitors don't have
- Multi-model routing. Picking the right model for each task
NOT from: 'we use Claude' / 'we use GPT-5' / 'we have a chat interface.'
3. Cost / quality / latency target? Pick the point in the 3-space. Customer support automation needs high quality (regulatory + brand risk), latency can be 5s (chat), cost matters at scale. Autocomplete needs sub-300ms latency, medium quality is fine, cost per call must be sub-cent. The product design follows from this pick.
4. How do we measure success? AI features fail to launch when there's no clear success metric. Define upfront:
- Adoption (% of users using the AI feature)
- Quality (eval score, user satisfaction, escalation rate)
- Business impact (revenue, retention, deflection)
The strategy doc
For an AI initiative, the strategy doc structure:
- The customer job. Specific, evidence-backed.
- Why AI changes this. What it enables that wasn't possible before.
- Defensibility analysis. Where our moat comes from. Why competitors can't fast-follow.
- Product approach. Specific UX, model choice, RAG/agent architecture.
- Cost model. Per-user, projected.
- Eval plan. What we'll measure, how, what 'good' looks like.
- Risks. Hallucination, cost, competitor catch-up.
The trap to avoid: feature-led AI
The pattern: 'let's add AI to feature X' → builds AI version → ships → 5% adoption → kill.
The reason it fails: feature X wasn't broken in a way AI could fix. AI was the answer to the wrong question.
The fix: start from customer jobs that are unsolved or expensive today. Add AI where it qualitatively changes the experience, not where it's a nice-to-have.
The 2026 reality
The frontier models are very good. Most AI products' constraint isn't 'the model isn't smart enough' — it's 'the product UX is wrong' or 'the strategy is unclear.' The PM job is to find the right product design and strategy; the model is a commodity.
Real-world examples
Notion AI added AI to an existing product workflow. The strategy was clear — 'AI helps you write faster in Notion' — and the integration was deep. The challenge: as standalone Claude and ChatGPT got better, the value of Notion's embedded AI relative to switching to a browser tab eroded.
Cursor's strategy was clearer: be the IDE for developers in the AI era. The wedge was the existing developer workflow (VS Code), AI capabilities were the differentiator, and the moat became distribution + product velocity once they had millions of developers.
Go deeper — recommended reading
Interview questions (1)
Q1Walk me through how you'd add AI to a B2B SaaS product (e.g., a CRM).ai-pmsenior▼
Don't start with 'where can we add AI.' Start with 'what jobs do our users do today that are painful or expensive?'
For a CRM, painful jobs include: writing follow-up emails, summarizing call notes, identifying deals at risk, prioritizing the day's outreach, prep before a meeting. Each is a candidate for AI.
I'd pick 2 to start, based on:
- Customer pain magnitude. Which is the most-complained-about?
- AI fit. Synthesis (call summaries), generation (email drafts), prediction (deals at risk) — all natural for AI.
- Defensibility. Jobs where our CRM data gives us an unfair advantage vs. a standalone GPT browser tab.
Then for each, define cost/quality/latency targets, eval plan, success metric. Ship one as a public beta, measure rigorously, iterate.
The strategic point: defensibility comes from our data + workflow integration. A salesperson won't switch to ChatGPT to draft a follow-up — but they will use our CRM's draft button. That's the moat. Build the UX that's better than the browser tab and we keep them.