Atiq Israk leads AI product at Chromatics (Kaizen, Anne, Pico) after shipping 15+ products across Ether and Navana. This is AI product strategy for 2026: moats from workflow, data, and trust—not from picking the newest foundation model.

Key Takeaways

• In 2026, models commoditize fast; durable AI products win on workflow embedding, proprietary context, and operator trust.
• Strategy starts with where manual work bleeds money or revenue—then asks if AI beats manual on an eval, not if the demo wows.
• Portfolio thinking beats single-feature bets: wedge, measure, compound (Kaizen pattern).

Why Doesn't Model Choice Matter Most?

Foundation models are converging on capability for many tasks. Hiring data shows PM roles emphasize RAG (8%), agentic workflows (8.5%), and observability (18%)—not "pick GPT vs Claude" (Axial Search, 2026). Your moat is what the model cannot download: your inventory graph, your restaurant tickets, your hospital workflows.

What Counts as an AI Moat in 2026?

Four moat types I look for before greenlighting AI bets:

1. Data loop — usage improves labels, labels improve retrieval (AssetIQ scan corrections).
2. Workflow lock-in — AI sits inside daily ops, not a sidebar chat (Neoshift shift flow).
3. Trust & audit — operators can verify outputs (inventory reconciliation before dashboards).
4. Cost structure — you can serve inference at margin in your market (thin-margin pricing).

How Should PMs Sequence AI Bets?

Use wedge sequencing from Kaizen: ship the smallest AI surface that doubles retention on one job-to-be-done before expanding. Brand voice before general assistant. Navbot: hours and reservations before open-domain chat.

Each wedge needs: named metric, eval gate, and kill criteria. See demo to production.

What Is the 2026 Portfolio View?

At Chromatics I treat AI as a portfolio:

• Kaizen — SMB wedge, retention metric.
• Anne — consumer trust/privacy differentiation.
• Pico — generative commerce with client-specific context.

Not every product gets the same AI depth. Capital and attention follow eval-proven outcomes.

How Do You Say No to AI Theater?

Kill or defer when:

• No labeled eval set and no owner for context maintenance.
• Metric is vanity (messages sent) not business (cost saved).
• Operators bypass the feature within two weeks.

That discipline is outcome-first strategy—not anti-AI.

Explore: case studies · frameworks · writing