AI Product Canvas

AI Product Canvas Skill

Define AI products with the same rigour as any product decision — but with additional layers for data, model, evaluation, and responsible AI. This canvas prevents the most common AI product failure: building a technically impressive feature that doesn't solve a real problem.

AI Product Anti-Patterns to Check First

Before building, flag if any of these apply:

• ❌ "We should add AI to [existing feature]" — with no user problem defined
• ❌ Accuracy target undefined before build begins
• ❌ No plan for what happens when the model is wrong
• ❌ User-facing AI output with no human review or fallback
• ❌ Training data not audited for bias or quality
• ❌ No evaluation metric — "we'll know it when we see it"

---

AI Product Canvas Output Format

AI Product Canvas — [Feature Name] — [Date]

PM Owner: [Name] ML/AI Lead: [Name] Status: Discovery / Design / Build / Evaluation / Live

---

1. Problem Definition

User problem being solved:

[What specific situation is the user in? What job are they trying to get done?]

Why AI?

[What makes this problem require AI vs a deterministic solution? If the answer is "because we can," stop here.]

Success for the user looks like:

[What outcome does the user experience when the AI feature is working well?]

---

2. AI Approach

Task type:

• Classification
• Generation (text, image, code)
• Summarisation / extraction
• Recommendation
• Search / retrieval
• Prediction / forecasting
• Conversation / agent

Model approach:

• LLM API (GPT-4, Claude, Gemini, etc.) — specify: [Model name + version]
• Fine-tuned model on own data
• Custom model trained from scratch
• RAG (retrieval-augmented generation)
• Embedding + vector search

Rationale for chosen approach: [Why this, not alternatives]

---

3. Data Requirements

Data gaps: [What's missing and plan to get it] Privacy considerations: [Any PII in training or inference data] Data ownership: [Do we own this data? Can we use it for training?]

---

4. Evaluation Framework

Primary metric: [The number that defines success — accuracy, F1, BLEU, user rating, task completion rate] Minimum acceptable threshold: [Below X, the feature does not ship] Human evaluation plan: [How will humans review model outputs? Sampling rate? Review panel?]

---

5. User Experience Design

How is AI output presented?

• Direct output shown to user (high trust required)
• AI-assisted with user confirmation
• Suggestion user can accept/reject
• Background action with audit log

Confidence and uncertainty handling:

• What happens when confidence is low? [Show alternative, ask for clarification, fallback to manual]
• How is uncertainty communicated to the user? [UI pattern]

Fallback plan:

• If the model fails or returns an error: [Specific fallback behaviour]
• If accuracy degrades below threshold: [Kill switch or graceful degradation plan]

---

6. Responsible AI Checklist

• Bias audit completed on training data
• Demographic fairness evaluated (does performance differ by user group?)
• Hallucination / confabulation risk assessed and mitigated
• User can see and correct AI output
• Opt-out mechanism exists (can user disable the AI feature?)
• Output provenance visible when relevant (does user know AI generated this?)
• PII not used in ways user didn't consent to
• Regulatory review completed (GDPR, AI Act, sector-specific)
• Model cards / documentation completed

---