AGENTIC PRODUCT MANAGEMENT PLAYBOOK

Purpose
Use this playbook to roll out AI agents that (1) continuously research customer + market signal, (2) generate PRD/spec drafts, and (3) act as copilots across meetings and execution—without creating security, trust, or workflow chaos.

1) Choose the first workflow (pick ONE)
- Weekly Voice-of-Customer (VOC) brief
- Competitive monitoring + pricing/release-note digest
- PRD “spec compile” (draft + citations)
- Meeting-to-actions copilot (decisions, owners, deadlines)
Selection criteria: high frequency, repeatable structure, clear success metric, accessible data sources.

2) Standardize your inputs (avoid garbage-in)
- Define a single taxonomy for: customer segments, churn reasons, request categories, severity levels.
- Identify 3–5 canonical sources of truth (example):
  * Product analytics: Amplitude/Mixpanel dashboards
  * Support: Zendesk/Intercom tags
  * CRM: Salesforce fields (ARR, segment, stage)
  * Work tracking: Jira/Linear epics and labels
  * Knowledge base: Confluence/Notion PRDs + policies
- Create a “definition sheet” (1 page): KPI definitions, event names, time windows, and segment rules.

3) Set agent output requirements (non-negotiables)
- Every factual claim must include a citation (link to dashboard, ticket, doc, or transcript).
- Include a “Known Unknowns” section (what data is missing).
- Separate: Observations (facts) vs Recommendations (judgment).
- Use your company’s PRD template (Problem, Goals, Non-goals, User stories, Acceptance criteria, Risks).

4) Permissions + governance guardrails
- Start read-only. Draft into a sandbox (Notion/Confluence draft space).
- Restrict tool access by default (least privilege):
  * Allow: search/read, summarization, draft creation
  * Block: sending external emails, deleting tickets, changing production configs
- Require human approval for any publish/send action.
- Turn on audit logging where possible (who ran the agent, what it accessed, what it wrote).
- Block PII where feasible (redaction rules; do-not-ingest sources like HR docs).

5) Evaluation (treat it like production software)
Create a small “golden set” (30–50 historical cases) and score the agent weekly:
- Citation coverage rate (% of claims linked to sources)
- Precision for requirement extraction (did it capture the real need?)
- Spec completeness (did it fill all required sections?)
- Stakeholder rating (1–5 from Eng/Design/Support)
Define a rollback rule: if scores drop >10% after a model/prompt change, revert.

6) Rollout plan (60 days)
Week 1–2: Grounding + taxonomy + sandbox drafts
Week 3–4: Ship the first agent output on a fixed cadence (weekly VOC or PRD drafts)
Week 5–6: Add limited write permissions (create Jira tickets, assign owners)
Week 7–8: Expand to a second workflow + formalize evals + governance

7) ROI tracking (simple scoreboard)
Track before/after metrics:
- Cycle time: idea → ready-for-eng (target: -20% to -30%)
- Hours saved per PM per week (target: 2–5 hours)
- Rework rate (bugs/defects traced to unclear requirements)
- Meeting load (number of alignment meetings per initiative)

PRD COMPILER TEMPLATE (copy/paste)
1) Problem (with citations)
2) Goals (measurable KPI target + timeframe)
3) Non-goals
4) User stories (persona + job-to-be-done)
5) Acceptance criteria (testable)
6) Risks + mitigations
7) Open questions + owners
8) Evidence appendix (tickets, quotes, charts)

If you implement only one rule: “No citation, no trust.” That single constraint turns an LLM from a fluent guesser into an accountable product teammate.