The Double Diamond Is Broken for AI Products. Here's What I Use Instead.
2026-01-29 · 12 min read · Janaina Maia
I love the Double Diamond. Discover → Define → Develop → Deliver. Clean, logical, widely understood. I've used it for years.
But it wasn't designed for AI products. And when I tried to force-fit it, things broke in ways I didn't expect.
Traditional product design assumes deterministic systems. You design an interaction, build it, and it works the same way every time. AI products are probabilistic. Same input, different outputs. Quality depends on data, model version, context, and factors that change over time.
After wrestling with this mismatch on multiple products, I built something I call the AI Diamond.
Where the Double Diamond Breaks
Discovery doesn't include data. Traditional discovery focuses on user needs. For AI products, you also need to discover what data exists, what quality it is, and what AI can realistically do with it. The most elegant design is worthless if the data doesn't support it.
Definition can't be static. In traditional products, you define the problem once and build it. With AI, model capabilities change. What was impossible last quarter might be feasible now. Your definition needs versioning.
Development requires experimentation. You can't fully spec an AI feature in Figma and hand it off. You need to prototype with real models and real data. Design and development blur together.
Delivery is never done. AI products change behavior over time as models update and data evolves. Delivery is the beginning of continuous monitoring, not the end of a project.
The AI Diamond Framework
Two additions to the Double Diamond: a Data Discovery track parallel to User Discovery, and a Monitoring Loop that follows Delivery.
Phase 1: Dual Discovery
Run two tracks simultaneously:
User track: Needs, pain points, workflows, mental models. Same as always.
Data & AI track: What data is available? What AI capabilities are mature? What's been tried before? What are the known failure modes? Talk to data scientists as key stakeholders.
The output: a Feasibility Map — user needs mapped against AI/data readiness. Immediately shows where high-value, high-feasibility opportunities exist.
Phase 2: Adaptive Definition
Define the problem, but add AI-specific elements:
- Accuracy requirements — a specific number, not "as accurate as possible"
- Failure mode design — how the feature behaves at each confidence level
- Autonomy level — copilot? coach? supervisor?
- Human-in-the-loop spec — where and how do humans oversee?
- Ethics review — during definition, not after launch
And version your definitions. Date them. Track changes. An AI feature defined in January may need redefinition by March.
Phase 3: Experimental Development
Replace "develop" with iterative experimentation:
- Test AI capabilities with real data, not synthetic examples
- Wizard of Oz testing before the model is production-ready
- Test with a range of inputs including edge cases and data gaps
- Verify confidence calibration — if it says 85% confident, is it right 85% of the time?
- User test with actual AI outputs, not mocked data
If you're testing with mocked AI outputs, you're testing a fantasy. Use real outputs or don't bother.
Phase 4: Monitored Delivery
Ship, then monitor continuously:
- Track real-world accuracy against targets
- Watch user behavior — override rates, engagement with explanations
- Capture corrections and feed them back to model improvement
- Schedule quarterly reviews against original design intent
- Treat model updates as design changes — because they are
You don't need to overhaul your entire process. Layer these additions onto what you already do. Add data discovery to existing research. Add accuracy requirements to existing specs. Add real-model testing to existing prototyping. Add monitoring to existing post-launch measurement.
The AI Diamond isn't a replacement for the Double Diamond. It's an upgrade. And honestly, once you've used it, going back feels like trying to design a conversation without knowing who you're talking to.