AI for Product Management

1. 01

   Product fundamentals

   Beginner

   The craft AI sits on top of. A PM who cannot frame a problem does not become useful by adding a model to it.

   • SVPG — Articles Article Silicon Valley Product Group Marty Cagan and colleagues 8h Free Start with the pieces on product discovery and empowered teams. This is the closest thing the discipline has to a canon.
   • Shape Up Book 37signals Ryan Singer 6h Free Free in full. Whether or not you adopt six-week cycles, the chapters on shaping and appetite change how you scope work.
   • Product Talk — Continuous discovery Article Product Talk Teresa Torres 5h Free Opportunity solution trees and the habit of weekly customer contact. The most practical discovery method in general use.
2. 02

   AI product thinking

   Intermediate

   What changes when the system is probabilistic: the feature can be wrong, and the design has to hold that gracefully.

   • People + AI Guidebook Documentation Google PAIR 5h Free The best free material on designing for a system that is confidently wrong sometimes. The chapters on mental models and errors are essential.
   • HAX Toolkit — Guidelines for Human-AI Interaction Documentation Microsoft Research 4h Free Eighteen research-backed guidelines with concrete examples. Use it as a design review checklist for any AI surface.
   • Prompt Engineering Guide Documentation DAIR.AI 5h Free Vendor-neutral and kept current. You do not need to write the prompts, but you need to know what is and is not achievable with them.
   • Stanford AI Index Report Paper Stanford HAI 4h Free Where capability, cost and adoption actually stand, with sources. The antidote to strategy built on press releases.

   Build this

   A failure-mode map for one AI feature

   Take a real AI feature — yours or a competitor's. Enumerate every way it can be wrong: confidently wrong, subtly wrong, wrong for one group of users only, refusing when it should answer.

   For each, decide what the interface does about it. This document is what separates a shipped AI feature from a demo.

   Done when: A failure-mode table with a designed response for each row.

3. 03

   AI-assisted discovery and research

   Intermediate

   Using models to move faster through synthesis and analysis — without letting them invent your findings.

   • NN/g — AI and UX research Article Nielsen Norman Group 3h Free Evidence-based rather than enthusiastic. Read it for where AI-assisted synthesis measurably helps and where it quietly does not.
   • GOV.UK Service Manual — User research Documentation GOV.UK 4h Free The baseline method. Know how research is done properly before deciding which parts to accelerate.

   Build this

   Synthesis, twice

   Take twenty real pieces of customer feedback. Synthesise them by hand into themes. Then do it again with a model, independently.

   Compare. What did the model miss, invent, or over-weight? Write that up — it is the most useful thing you can know before trusting AI synthesis on a real decision.

   Done when: Two syntheses and a written comparison of where they diverged.

4. 04

   Writing specs for AI features

   Intermediate

   A specification for a probabilistic system needs things a normal PRD does not: acceptance criteria that are statistical, and a defined behaviour for being wrong.

   • Google — Technical Writing Courses Course Google 4h Free Most rejected specs are rejected for ambiguity, not for the idea. This fixes the ambiguity.
   • Model Cards for Model Reporting Paper arXiv Mitchell et al. 1h Free The documentation standard for what a model does, on whom, and how well. Borrow its structure for your own specs.

   Build this

   A PRD engineering could actually build from

   Write the full specification for one AI feature: the user problem, the non-AI baseline you are beating, acceptance criteria expressed as measurable quality thresholds, the fallback when the model fails, and the data you will need.

   Then give it to an engineer and ask what is missing. Their first three questions are your rewrite.

   Done when: A reviewed PRD, plus the questions an engineer asked and how you answered them.

5. 05

   Experimentation

   Advanced

   Deciding whether the thing worked. Most disagreements about AI features are really disagreements about measurement.

   • How Not to Run an A/B Test Article Evan Miller Evan Miller 30 min Free Why stopping a test when it looks significant is the most common and most expensive mistake in product analytics.
   • Online controlled experiments — practical guide Paper exp-platform.com Ronny Kohavi and colleagues 5h Free Papers from the people who built experimentation at Microsoft and Amazon. Dense, and the most credible source there is.
   • Seeing Theory Interactive Brown University 3h Free If confidence intervals are still vague, spend an evening here before arguing about a test result again.
6. 06

   Evaluation and measuring outcomes

   Advanced

   The discipline that decides whether an AI product improves: a repeatable evaluation set, owned by the product team rather than the model provider.

   • Your AI Product Needs Evals Article hamel.dev Hamel Husain 2h Free The clearest practical writing on building an evaluation loop for an LLM product. Read it before your first prompt change goes to production.
   • HELM — Holistic Evaluation of Language Models Documentation Stanford CRFM 3h Free How the research community measures models across many axes at once. Useful for understanding why a single benchmark number means little.
   • NIST AI Risk Management Framework Documentation NIST 4h Free The vocabulary your legal and security colleagues will use. Knowing it makes you the PM who can get an AI feature approved.

   Build this

   An evaluation set you own

   For the feature you specified, hand-write 50 test cases with expected behaviour — including the awkward ones: ambiguous inputs, out-of-scope requests, and cases where refusing is the right answer.

   Score the current implementation against it, then change one prompt and score again. That table is how AI product decisions should be made, and almost nobody has one.

   Done when: A 50-case evaluation set, two scored runs, and what the numbers changed your mind about.

Nothing matches that combination in this path yet.