AI playbook for the head of product — discovery, specs, releases

If you're a head of product reading 12 user-interview transcripts a quarter and still feeling behind on signal — the issue isn't your discovery cadence. It's that 77% of the AI work helping product teams is invisible, scattered across PMs' personal accounts, and never makes it into your roadmap.

Why product is where AI signal goes to die

Stanford's 2025 work on enterprise AI surfaced a pattern: ~77% of AI usage in organizations is invisible — done on personal accounts, in unsanctioned tools, with no propagation back to the team. In product, that means a PM might have a brilliant ChatGPT thread synthesizing 30 customer interviews, and you, the head of product, never see it.

The head of product's job isn't to ban that — it's to bring the work into shared light. That requires sanctioned tools and a couple of structural rituals. Not a tool RFP.

Definition: Discovery — the structured search for signal about what users actually need, before deciding what to build. Distinct from delivery.

The 90-day product AI playbook (six plays)

Discovery — Play 1: interview synthesis with traceable quotes

Have every PM upload their interview transcripts (sanctioned tool, EU-region storage if you're in Europe) and run a fixed synthesis prompt: "Pull 5 themes. For each theme: 3 verbatim quotes with timestamps, the count of distinct interviewees raising it, and one disconfirming quote." The disconfirming quote is the magic — it forces the model to look for evidence against the theme, which is where most "AI summary" outputs fail.

You are a discovery analyst. From these N transcripts, extract:
- 5 themes (max)
- For each: 3 verbatim quotes (with interviewee initials + timestamp)
- For each: count of distinct interviewees raising the theme
- For each: 1 disconfirming quote OR "no disconfirming evidence found"
DO NOT invent quotes. DO NOT paraphrase quotes.
TRANSCRIPTS: {{transcripts}}

Discovery — Play 2: jobs-to-be-done extraction

Same corpus, different lens: extract verb-noun-context tuples (e.g., "draft a renewal email | when a customer is silent for 30 days"). Hand the JTBD list back to the PM and the designer. The output is the input for the next sprint's discovery doubles-down.

Tool tip (Course for Business): In our 6-week program product PMs run Shoulder-to-Shoulder sessions on their own live transcripts. We've seen first-week wins where a PM walks away with a synthesis prompt that runs against their actual research backlog by Friday. The principle is Augment, don't replace — the PM still owns the theme, AI handles the mechanical extraction. See course.aiadvisoryboard.me/business.

Specs — Play 3: PRD draft from discovery + JTBD

Once the synthesis is in, AI drafts a PRD skeleton: problem statement, target user, top 3 user stories, top 3 risks, out-of-scope. The PM rewrites it. The benefit isn't speed; it's that the first draft is no longer a blank page, and the PM is critiquing rather than generating.

Specs — Play 4: requirement-gap interrogation

Take any spec — yours or one you inherited — and run an "interrogator prompt": "List 10 questions a senior engineer would ask before estimating this work. Group by ambiguity, missing data contract, and missing failure-mode definition." This catches more spec defects than any peer-review template, because the LLM is shameless about asking obvious questions you'd hesitate to.

Releases — Play 5: release notes that humans read

The classic failure: engineering writes 40 commit-summaries, marketing rewrites them as "exciting new features", and customers learn nothing. Replace this with a structured AI pipeline: git log → segment by user-visible vs internal → for each user-visible change, write 1-line "what changed" + 1-line "why you might care". A PM signs off; marketing styles. Cycle time on release notes drops from 4 hours to 30 minutes.

Releases — Play 6: incident-postmortem first-draft

When something breaks in production, the postmortem dies because nobody has 4 hours. Have the on-call paste the timeline + Slack threads into a sanctioned tool with a fixed template: "Timeline, contributing factors, what worked, what didn't, 3 action items with owners." Engineering edits. Postmortem-completion rate roughly doubles.

Team scan (what AI champions report after week 1)

• Each PM ships at least one synthesis prompt against real transcripts.
• 5+ JTBD tuples surface that the team had not articulated before.
• 1-2 specs get rewritten because the interrogator prompt found gaps.
• Release notes cycle time drops noticeably (commonly ~50%).
• 1 incident postmortem that would have been skipped gets written.
• The head of product reads weekly verbatim quotes from real users for the first time in months.
• A previously-shadow PM AI workflow is brought into the sanctioned stack.
• Engineering reports fewer "spec ambiguity" tickets back to product.
• 1 PM admits they had been using personal-account AI for spec drafting and now switches.
• 1 designer joins the discovery synthesis loop unprompted because the output is finally usable.

Tool tip (Course for Business): Product orgs we work with run with an AI Champions (1:15-20) ratio across PMs, designers, and engineering managers. For a 100-person product+eng group that means 5-7 champions — typically 1 senior PM, 1 design lead, and 2-3 engineering managers. They run weekly clinics. The bridge from "scattered prompts" to "shared playbooks" happens in week 3-4, not week 1. course.aiadvisoryboard.me/business.

Micro-case (what changes after 7-14 days)

A typical 220-person SaaS company with 4 PMs runs the playbook as follows. Week 1: synthesis + JTBD across last quarter's interviews; the head of product discovers two themes the team had been undercounting. Week 2: PRD-draft and interrogator prompts roll into the active sprint; one in-flight spec is rewritten before engineering picks it up, saving an estimated 2-3 days of rework. Week 3: release notes pipeline goes live; marketing stops complaining. Week 4: first AI-drafted postmortem ships. By day 14 the head of product has clearer signal, less spec rework, and faster release comms — without adding headcount or cycle pressure.

Note on this case: This example is illustrative — based on typical patterns we observe with companies of 30-500 employees, not a single named client. Specific numbers are rounded approximations of common ranges, not guarantees.

FAQ

Should we use AI to write user stories from scratch?

You can, but the value is much lower than using AI to interrogate a draft. Generation is cheap and produces generic stories. Interrogation forces specificity. Default to interrogation.

What about Cursor / Copilot for PMs prototyping ideas?

Useful for spike-prototypes and Figma-to-code experiments. Not where the head of product's leverage lives. That's the head of engineering's playbook.

How do we keep PMs from over-trusting the synthesis output?

Mandate the disconfirming-quote field, and have PMs cite at least 2 quotes by interviewee initial in any roadmap doc. Citations create accountability; vague summaries don't.

Are there compliance concerns with uploading customer transcripts?

Yes — handle in a sanctioned, region-appropriate tool (EU for EU customers), strip PII at upload, and have a written DPA with the vendor. The 46% shadow-AI figure is mostly people taking shortcuts here. Make the right path the easy path.

Does this overlap with your daily-management product?

Slightly — the daily-management OS surfaces what every team did, including product, at the company level. The playbook above is what a head of product does inside their own function. Use both, but for different decisions.

Conclusion

A head of product who runs these six plays in 90 days has compressed discovery, sharpened specs, and accelerated releases — without buying anything. The hard part is structural, not technical: making the sanctioned path more convenient than personal-account workarounds.

If you want every product team member to ship their first AI automation in five days — book a 30-min call and we'll map your team's first week: course.aiadvisoryboard.me/business.