Product & Strategy

research plan → interview guide → transcripts → coded themes → insight report → opportunity backlog

• Draft
  Research planning & guide drafting — AI drafts interview guides and screener questions from a stated learning goal, flags leading questions · low stakes, easily verified by reading, and the PM must own what gets asked
• Automate
  Recruiting & scheduling — agent screens panel candidates against criteria, sequences outreach, books slots · rule-based, high-volume, trivially reversible if a bad fit slips through
• Assist
  Conducting interviews — AI takes live notes and suggests follow-up probes; the human runs the conversation · rapport and probing are relationship work, and the interview itself is the context AI lacks
• Draft
  Transcription & thematic coding — AI transcribes, tags quotes against a codebook, clusters themes across dozens of sessions · high volume and checkable by spot-reading quotes in context, but AI over-clusters and flattens dissenting voices, so a human must re-anchor themes
• Assist
  Insight synthesis & prioritization — deciding which findings change the roadmap · verdicts here are irreversible bets with no ground truth to verify against; AI can pressure-test the argument, not make it

problem brief → PRD draft → edge-case/open-questions log → reviewed spec → engineering-ready tickets

• Assist
  Problem framing — writing the one-page "why this, why now" with sizing and evidence · this is the highest-stakes judgment in the chain and the evidence lives in conversations AI wasn't in
• Draft
  Structure & first draft — AI expands the framing into a full PRD against the team's template, pulling in linked research and prior specs · verifiable by the author line-by-line, repetitive format, and the PM keeps ownership through editing
• Automate
  Edge-case and failure-mode generation — AI enumerates states, permissions, error paths, and abuse cases the draft missed · generation is cheap, every suggestion is independently checkable, and a bad suggestion costs one deletion
• Avoid
  Stakeholder review & tradeoff resolution — negotiating scope with design, eng, legal · relationship-driven, politically loaded, and AI in the loop erodes the trust the negotiation runs on
• Draft
  Ticket decomposition — AI breaks the approved spec into stories with acceptance criteria · mechanical and verifiable against the spec, but decomposition choices affect sequencing so eng leads must edit

opportunity backlog → scoring sheet → draft roadmap → tradeoff memo → committed plan → comms deck

• Automate
  Backlog consolidation — agent sweeps tickets, sales asks, support themes, and research insights into one deduplicated opportunity list · pure aggregation at high volume, and errors surface immediately when owners review their items
• Draft
  Scoring & sizing — AI drafts RICE/impact-effort scores with cited reasoning per item · the arithmetic is checkable but the impact estimates encode strategy, so humans must own every score they keep
• Avoid
  Tradeoff decisions — choosing what to cut and what to fund · irreversible resource allocation with no verifiable right answer; AI-generated rankings create false objectivity that shuts down the argument the team needs to have
• Assist
  Scenario modeling — AI generates "what if we cut X / added a team / slipped Y" roadmap variants · useful for exploring the space fast, but capacity assumptions are soft and each scenario needs human sanity-checking
• Draft
  Roadmap communication — AI tailors the committed plan into exec, sales, and customer-facing versions · repetitive reformatting with the committed plan as ground truth, but external promises carry stakes so a human signs off

monitoring feed → change log → competitor teardown → battlecard → strategy implications memo

• Automate
  Signal monitoring — agent watches competitor changelogs, pricing pages, job postings, filings, and app-store reviews, and files structured change alerts · high-volume, repetitive, and every alert links to a checkable source
• Draft
  Teardown drafting — AI compiles a feature/pricing/positioning teardown from the collected signals · claims are verifiable against sources, but AI fills gaps with plausible fabrications, so every unlinked claim gets cut
• Draft
  Battlecard maintenance — AI updates sales battlecards when the change log shifts · templated and source-grounded, but a wrong claim ends up in a customer's ear, so sales-facing edits get reviewed
• Draft
  Win/loss interview analysis — AI codes win/loss call transcripts for recurring competitive themes · same verifiability profile as research coding; spot-check quotes against calls
• Assist
  Strategic implications — deciding what the competitor's move means for your bets · low-volume, unverifiable, and the answer depends on private context about your own strategy; AI is a sparring partner only

positioning doc → messaging matrix → launch plan → asset kit (announcement, docs, enablement, FAQ) → launch retro

• Assist
  Positioning & messaging — choosing the audience, the alternative you're positioned against, and the claim you'll lead with · one irreversible public choice with no verification path; AI generates options and stress-tests, humans choose
• Draft
  Launch plan assembly — AI drafts the run-of-show, owner matrix, and dependency checklist from the team's launch template · templated and internally verifiable, repeated every launch
• Draft
  Asset production — AI produces the blog post, email variants, docs updates, demo script, and sales one-pager from the approved messaging doc · high-volume derivative writing with the messaging doc as ground truth; humans edit for claims and voice
• Avoid
  Legal/claims review — verifying public claims about performance, pricing, and compliance · regulatory exposure and reputational stakes; AI can pre-flag risky phrasing but sign-off is human by policy
• Automate
  Enablement & FAQ — AI drafts internal FAQ and objection-handling from the spec and battlecards, then answers field questions as a launch-scoped bot · internal audience, source-grounded, and wrong answers are cheap to correct
• Draft
  Launch retro — AI compiles metrics, field feedback, and timeline slips into a retro doc · aggregation is mechanical, but the "what we'd do differently" section is the point and must be human

metrics readout → feedback digest → hypothesis backlog → experiment briefs → ship/kill decisions → changelog

• Automate
  Feedback aggregation — agent clusters support tickets, reviews, NPS verbatims, and sales notes into a weekly digest with links to raw items · high-volume classification, every cluster is checkable by clicking through, misfiles are harmless
• Draft
  Metrics anomaly triage — AI flags adoption dips and funnel breaks and drafts first-pass "what changed" narratives · flags are checkable against dashboards, but AI narratives assert causes confidently, so treat every explanation as a hypothesis
• Draft
  Hypothesis generation — AI proposes experiments from the feedback digest and funnel data, formatted as testable briefs · cheap to generate and verify for testability; humans pick what's worth a slot
• Assist
  Experiment design review — checking that the test can actually answer the question (power, metric choice, duration) · subtle-error territory where mistakes silently invalidate results; AI checklists help but a human owns the design
• Assist
  Ship/kill decisions — reading results and deciding · results are rarely clean, the decision is semi-irreversible, and it prices in strategy context AI doesn't hold
• Automate
  Changelog & comms — AI drafts release notes and closes the loop with customers who reported the issue · templated, low-stakes, verifiable against the ticket, and users love the closed loop