"AI products aren't deterministic. They require continuous calibration, not just A/B tests."

product-strategy

product-discovery

product-architecture

product-delivery

product-leadership

• Building AI agents that act on behalf of users
• Adding LLM-powered features to existing products
• Designing human-AI interaction patterns
• Deciding how much autonomy to give AI
• Setting up eval strategies and calibration loops
• Managing the "agency-control tradeoff"

Credit: Aishwarya Goel & Kiriti Gavini

┌─────────────────────────────────────────────────────────────────┐
│                        CCCD LOOP                                │
│                                                                 │
│    CALIBRATE → CONFIDENCE → CONTINUOUS DISCOVERY → CALIBRATE   │
│         ↓           ↓              ↓                 ↓         │
│     Eval and    Build user    Observe AI       Update evals    │
│     adjust AI    trust over   interactions     and models      │
│     behavior     time         at scale                         │
└─────────────────────────────────────────────────────────────────┘

Level 1 → Build trust → Level 2 → Demonstrate reliability → Level 3 → ...

• Irreversible actions (delete, send, purchase)
• High-stakes decisions (financial, legal, health)
• Novel situations outside training distribution
• Actions affecting third parties

• "Which parts of this task feel tedious vs. require your judgment?"
• "If AI made an error here, what would the consequences be?"
• "How would you want to verify AI's work?"

• "What would AI need to demonstrate before you'd trust it to [action]?"
• "Have you used AI tools before? What built or broke your trust?"
• "Would you prefer AI to do more but occasionally err, or do less perfectly?"

• "What kinds of errors are annoying vs. damaging?"
• "How quickly do you need to catch and fix AI mistakes?"
• "What's your 'undo' option if AI gets it wrong?"

OBSERVE → IDENTIFY → CALIBRATE → VALIDATE → DEPLOY
   ↑                                           │
   └───────────────────────────────────────────┘

• Eval metrics below threshold
• New failure pattern identified
• User feedback trend (negative)
• Model update available
• New use case discovered

templates/

• agency-assessment.md

  — Determine appropriate agency level

• eval-strategy.md

  — Design eval suite for AI feature

• calibration-plan.md

  — Set up continuous calibration

1. Assess agency level: "What agency level should [AI feature] have?"
2. Design agency progression: "Create a graduation path from assist to autonomous for [feature]"
3. Identify failure modes: "What could go wrong with [AI feature]? How do we mitigate?"
4. Design eval strategy: "Design an eval suite for [AI feature]"
5. Plan calibration: "Create a calibration plan for [AI feature]"
6. Adapt discovery: "What AI-specific questions should I ask in discovery for [use case]?"
7. Design confidence building: "How should [AI feature] show its reasoning?"
8. Plan AI rollout: "Create a staged rollout plan for [AI feature]"
9. Set AI metrics: "What metrics should we track for [AI feature]?"
10. Review AI brief: "Critique this solution brief for AI considerations"

• Agency level defined — Clear level for this feature
• Graduation criteria set — How we'll earn higher autonomy
• Failure modes mapped — Know what can go wrong
• Evals in place — Automated quality checks
• Human evals scheduled — Subjective quality review
• Calibration loop running — Continuous improvement process
• Confidence mechanisms built — Users can verify AI work
• Guardrails active — Prevent harmful outputs
• Rollout staged — More cautious than traditional features
• Override tracking — Learning from user corrections

• Aishwarya Goel & Kiriti Gavini — CCCD Loop, Agency-Control Trade-off
• Anthropic — Constitutional AI, RLHF approaches
• OpenAI — Eval best practices
• Google DeepMind — AI safety frameworks