AI in Product Discovery | Product Discovery Framework

The human owner remains accountable.

Use an organization-approved AI service and remove confidential, personal, financial, security, credential, or client data unless that service is approved for it.

Keep original notes and sources. AI summaries and generated text are drafts, not evidence or approval.

Verify benchmark claims against the original source and record its URL, publication date, and access date.

A named BA, Product Manager, Designer, technical reviewer, or business owner remains accountable for the final decision and sign-off.

Choose AI for the work it can safely accelerate.

Benchmarking

AI-assisted benchmarking

Use AI to accelerate the collection and comparison of internal and external examples without treating an AI summary as the source.

Find comparable products, processes, interfaces, standards, and internal prior work.
Compare examples against explicit, consistent criteria.
Separate observed evidence from interpretation and generated suggestions.
Record and verify the original source, publication date, and access date.

Note-taking

AI-assisted note-taking

Use AI after approved interviews, workshops, observations, and review sessions to structure evidence while keeping the original record.

Separate direct observations, participant statements, decisions, assumptions, risks, open questions, disagreements, and actions.
Keep participant attribution and the original evidence traceable.
Review every summary before sharing it.
Never allow AI to invent quotations or imply agreement where people disagreed.

Documentation

AI-assisted documentation

Use AI to create the first draft of discovery artifacts from reviewed evidence. The accountable owner must validate the result.

Convert approved notes into the required template and audience-appropriate language.
Identify missing sections, unanswered questions, and conflicting rules.
Keep facts, assumptions, decisions, risks, and open questions distinct.
Mark generated material as draft until the required human review is complete.

Prototyping

AI-assisted prototyping

Use Codex or Claude Code to turn a defined discovery question into the smallest working experience that users and stakeholders can evaluate.

Provide the users, problem, journey, rules, states, design references, synthetic data, and specific question to answer.
Include realistic happy paths, errors, empty states, permissions, and important edge cases.
Host the reviewable prototype on Vercel and preserve meaningful test versions.
Use Supabase Free only when authentication, persistence, storage, or multi-user behaviour improves the test.

Build only what the discovery question needs.

Use a coding agent to create the experience, Vercel to share it, and Supabase Free only when a lightweight database materially improves the test.

Codex Claude Code Vercel plans Supabase plans

1. Define the decision

State what the prototype must help the team learn or decide. Do not begin with an open-ended request to build the whole product.

2. Define the boundary

List the users, journeys, screens, rules, states, devices, and edge cases that matter, together with what must remain outside the prototype.

3. Provide design direction

Give Codex or Claude Code the approved Figma references, design system, brand guidance, content, accessibility expectations, and synthetic sample data.

4. Build with a coding agent

Use Codex or Claude Code to create only the functionality needed for the discovery question. Include realistic happy paths, errors, empty states, permissions, and important edge cases.

5. Deploy to Vercel

Publish a stable review URL and use preview deployments to preserve meaningful iterations. Confirm that the selected Vercel plan permits the intended personal, client, or commercial use.

6. Add Supabase only when needed

Use the Supabase Free plan when authentication, persistence, storage, or multi-user behaviour materially improves the test. Use synthetic or approved test data and treat the service as prototype infrastructure, not production.

7. Test and retain evidence

Record the version tested, participants, observed behaviour, findings, decisions, changes, and unresolved questions. The prototype is evidence, not the final requirement.

Keep evolving prototypes credible and consistent.

Keep evolving prototypes close to the product

For products that will use several prototypes over time, obtain the existing Figma design system and integrate its components, design tokens, typography, colours, spacing, icons, and interaction patterns into the prototype code. Ask Codex or Claude Code to reuse these foundations and document exceptions.

Use a professional sharing address

If a suitable domain is available through a legitimate promotion for approximately $1–$2, consider it for a client-facing prototype. Check renewal price, privacy charges, ownership, and transfer conditions; register it in an organization-controlled account; and display a clear Prototype or Preview label.

Apply the four areas differently in each playbook.

Open an activity's AI Assist tab for the contextual workflow.

Tactical Discovery

Benchmarking: Compare the request with previous tickets, existing components, established rules, and similar changes.
Note-taking: Turn focused requester and SME conversations into confirmed rules, edge cases, disagreements, and open questions.
Documentation: Draft the triage note, requirements, regression checklist, user story, and acceptance criteria from reviewed evidence.
Prototyping: Create a small before-and-after interface or clickable interaction only when the requested change is visually ambiguous.

Operational Discovery

Benchmarking: Compare processes across departments, branches, field teams, policies, and relevant operating-model patterns.
Note-taking: Summarize interviews and workshops by stakeholder group while preserving variations, exceptions, and disagreements.
Documentation: Draft current-state, governance, dependency, target-operating-model, roadmap, and executive artifacts.
Prototyping: Model future workflows, role-based workspaces, approval paths, exceptions, and service blueprints before committing to screens.

Mass Product Discovery

Benchmarking: Compare onboarding, activation, accessibility, trust, support, retention, and failure experiences across relevant products.
Note-taking: Cluster evidence by user segment while retaining participant identifiers, original observations, and contradictory findings.
Documentation: Draft segments, journeys, event plans, success metrics, trust requirements, rollout plans, and the final recommendation.
Prototyping: Build and test alternative mobile journeys, localized content, accessibility, slow-network behaviour, and recovery states.

Scaled Solution / Platform and SaaS Discovery

Benchmarking: Compare configuration, onboarding, administration, permissions, integrations, analytics, pricing, packaging, support, and tenancy.
Note-taking: Keep buyer, governance owner, administrator, and daily-user evidence separate in notes and decisions.
Documentation: Draft product boundaries, tenant and access models, operations, pricing, architecture readiness, MVP, pilot, and recommendation artifacts.
Prototyping: Create role-specific buyer, administrator, configuration, onboarding, and daily-use experiences using the approved design system.

Integration Discovery

Benchmarking: Compare official API documentation, authentication, integration patterns, limits, error standards, and reconciliation approaches.
Note-taking: Turn technical workshops into decisions, ownership, dependencies, conflicts, unknowns, actions, and approvals.
Documentation: Draft and check the SSOT matrix, data dictionary, NFRs, error rules, reconciliation logic, and sign-off package.
Prototyping: Create sanitized mock integrations, payloads, sequence visualizations, failure simulations, and reconciliation scenarios.

Data Discovery

Benchmarking: Compare metric definitions, reporting conventions, dashboards, alerts, and decision-support patterns without losing sight of the decision.
Note-taking: Separate agreed definitions, disputed definitions, assumptions, data gaps, decisions, and owners.
Documentation: Draft the business question, metric dictionary, lineage, quality log, assumptions, output requirements, and validation record.
Prototyping: Create dashboards, tables, alerts, filters, and drill-downs with synthetic data, then test whether users can make the intended decision.

AI & Machine Learning Discovery

Benchmarking: Compare the proposed model with the current process, a non-AI baseline, relevant capabilities, routes, and documented limitations.
Note-taking: Capture failure concerns, acceptable-error discussions, correction behaviour, escalation needs, and unresolved evidence.
Documentation: Draft readiness, boundaries, inputs, permissions, evaluation, safety, human review, monitoring, and final recommendation artifacts.
Prototyping: Build a constrained prompt, model workflow, or thin vertical slice and test evaluation, failure, adversarial, escalation, and rollback cases.