Loading Runway...
Loading Runway...
The deep version. Four sections: what AI actually changes about the PM job, how to lead AI-feature work without lying to yourself about capability, how to keep judgment as your moat, and how to run the org around AI products. Ends with a 20-question scenario assessment that moves your AI Market Fit Score.
Course · Card 1 of 14
Most AI-for-PM content teaches you prompts. This teaches you what changes about the decisions you own. The honest position: an LLM compresses the production layer of your job (drafts, summaries, first-pass analysis) and leaves the judgment layer exposed and more valuable. The PMs who lose are the ones who can't tell which layer a given task is. By the end you will be able to.
Section 1 · What AI changes
Take any PM task and split it: the SYNTHESIS layer (assemble known inputs into an artifact) vs the COMMITMENT layer (decide what's true / what we'll do / what we'll stop). The 2026 Anthropic Economic Index shows synthesis-heavy PM tasks are where deployed AI has moved fastest. Commitment-layer tasks have barely moved — not because models can't generate a decision, but because nobody can be accountable for one they didn't make.
Synthesis is being commoditised. Commitment is being concentrated.
Section 1 · What AI changes
The 'what' (scope, surface, structure) is the synthesis layer — an LLM gets it ~70% right and rising. The 'why' (the bet, the user truth you're staking the quarter on) and the 'how we'll know' (the metric you'll be wrong in public about) are commitment. A PRD that is 100% AI-drafted is a PRD where nobody made the bet. Reviewers feel that even when they can't name it.
Section 2 · Leading AI features
The single most expensive PM error in 2026 is shipping a roadmap off a demo. A model doing a task in a benchmark (capability) is not the same as it doing the task in your product at your error bar under your latency and cost (deployment), which is not the same as it changing the user's outcome (impact). Stanford HAI 2026 Q1: the median enterprise AI feature that demoed well had a 9-month gap to reliable production. Plan to that gap or it plans you.
Section 2 · Leading AI features
For an AI feature, the eval set IS the spec. If you cannot describe the 50 cases the feature must get right and the 10 it must never get wrong, you have not specified the feature — you've described a vibe. Write the eval before the prompt. The eval is the commitment-layer artifact; the prompt is synthesis. PMs who own the eval own the feature; PMs who own the prompt own nothing.
Section 2 · Leading AI features
A deterministic feature fails predictably. An AI feature fails plausibly — confidently wrong, in the user's voice. Your job is to decide the blast radius of a wrong answer before launch: is this a feature where wrong-but-confident costs a typo, or a lawsuit? The consequence-stakes of the task decides how much human-in-the-loop you owe, and that is a product decision, not an ML one.
Section 3 · Judgment as moat
When drafting gets cheap, the bottleneck moves to deciding which draft is right and being accountable for it. That's not a consolation prize — it's the part of the job with the highest consequence-stakes and lowest verifiability-by-output, which is exactly the profile that resists automation (the Runway task model). The PMs compounding in 2026 spend the time AI freed on more bets, sharper, faster — not on watching AI draft.
Section 3 · Judgment as moat
An LLM is trained on the public internet's product decisions — the average. Your edge is the proprietary, unlogged dataset of why your specific users churned, what your CEO actually means by 'enterprise-ready', which past bet failed and why. That context never entered training. Feeding it deliberately into how you use AI (not just asking cold) is the difference between AI-as-intern and AI-as-amplifier.
Section 3 · Judgment as moat
The reason a VP greenlights your roadmap is not the deck — it's that you've been right enough before to be trusted with being wrong now. MIT Sloan's 2026 cross-functional study found AI-generated recommendations were adopted at less than half the rate of identical recommendations carried by a trusted human. The relationship is the moat. AI can write the update; it cannot be the person who owns the miss.
Section 4 · Running the org
AI features blur the PM/eng/data line: the eval set is product, the prompt is eng, the failure analysis is data, and all three change weekly. The teams that ship treat the eval as a shared, PM-owned artifact reviewed like a metric, not a one-time doc. The teams that stall argue about whose job the prompt is. Decide ownership of the eval explicitly in week one.
Section 4 · Running the org
You're now planning against a dependency (model capability) that improves on someone else's schedule. Two failure modes: assuming today's capability is permanent (you under-build), or assuming the demo's capability is today's (you over-promise). The discipline: separate the roadmap into 'works at current reliable capability' and 'works if the next jump lands' — and never let the second fund headcount.
Section 4 · Running the org
Engagement lies for AI features — users poke a novelty then leave. The honest metric pair: task-completion (did the user get the outcome they came for) and trust-retention (do they come back to the AI path vs route around it). A feature with rising usage and falling trust-retention is pre-churn, not traction. Defining that pair is a commitment-layer act and it's yours.
Course · Card 13 of 14
Every section was one idea: AI collapsed the synthesis layer of product management and concentrated value into the commitment layer — bets, evals, failure design, trust, metric definition, capability judgment. Your career math is simple: move your hours toward commitment, use AI to evacuate synthesis, and make the bet visibly yours.
Course · Card 14 of 14
20 scenario questions. Not recall — judgment. Each puts you in a decision and asks which move a strong AI-era PM makes and why. Pass bar is 70%, the same system-recorded bar as a module — passing moves your AI Market Fit Score and lands in your Artifacts as a course completion. There is no partial credit for the comfortable answer.