The AI PM concept map
107 concepts across 12 tracks, drawn with the 274 links between them. 127 of those links (46%) cross tracks. Those crossings matter most: the ideas that decide AI PM interviews, such as cost per task or golden sets, run through several rounds at once. Hover a concept to see only what it touches. Click to read it.
same trackcrosses tracks free membersbigger dot = more connections107 shown
Where the map is dense
The most-connected concepts are the ones the rest of the library keeps reaching for, which makes them the highest-leverage things to be solid on: Golden Sets: Building and Maintaining the Eval Data That Decides Launches (19), The Tool Execution Envelope: How Agents Act Safely on Real Systems (13), The Cost of a Wrong Answer: Error Cost and Recoverability as AI Design Inputs (11), Cost per Successful Task: The Unit Economics Number for AI Products (11), and The AI Fit Test: Does This Job Actually Need a Model? (9). If you are deciding where to spend a week, start with a hub rather than a leaf.
Every concept, by track
AI Product Discovery9
- The Cost of a Wrong Answer: Error Cost and Recoverability as AI Design Inputs11
- The AI Fit Test: Does This Job Actually Need a Model?9
- Sizing an AI Opportunity: From Task Volume to Value You Can Defend7
- Writing an AI PRD: An Eval-First Spec With Failure Modes, Thresholds and Kill Criteria6
- Capability vs Reliability: Why AI Demos Mislead and How to Plan Past Them5
- Deterministic Plus Probabilistic: Designing Workflows With Rules, Models and People5
- The AI Product Lifecycle: What Evidence Earns the Next Investment?4
- Customer Segmentation: Choose Whose Problem the AI Product Solves4
- AI Product Archetypes: Feature PM, Platform PM and Transformation PM3
Designing for Uncertainty7
- Autonomy Levels for AI Features: From Suggest to Delegate7
- Designing for Uncertainty: UX Patterns for AI That Is Sometimes Wrong6
- Measuring Trust in AI Products: Behaviour, Calibration and Validation6
- Discoverability and First Use: Getting Users From an Empty Box to a Task That Works4
- Risk and Reversibility: The Action Matrix for AI Features3
- Human Handoffs: Escalating From AI to People Without Losing the Customer3
- Perceived Latency: Making AI Features Feel Fast Before They Are Fast3
Strategy & Moats9
- Pricing AI Products: Seats, Usage, Credits and Outcomes7
- Platform Risk and Lock-In: Depending on Model Providers Without Being Trapped6
- Wrapper Gross Margin: The Unit Economics of AI Applications5
- Data Flywheels: Which Ones Are Real and Which Are Slideware5
- From Pilot to Production: Why Enterprise AI Pilots Stall and How to Get Them Into Use5
- Feedback Data: Learn From What Users Saw, Not Just What They Clicked5
- The Moat Ladder: Five Rungs of Defensibility for AI Products4
- Build, Buy or Fine-Tune: Choosing Between APIs, Open Models and Your Own Model4
- AI Product Distribution: Choose a Route to Customers That Can Pay4
How LLMs Work, for PMs11
- Context Engineering: Deciding What the Model Sees on Every Call6
- RAG Explained: Retrieval-Augmented Generation From Ingestion to Cited Answer5
- Fine-Tuning and LoRA: What Training on Your Examples Changes, and What It Costs5
- Classical ML for AI PMs: Thresholds, Generalization and Production Data5
- SFT, RLHF and DPO: Choosing the Behavior You Train For5
- Tokens and Context Windows: The Units of AI Cost, Speed and Memory4
- How LLMs Generate Text: Next-Token Prediction, Sampling, Prefill and Decode4
- Transformers, Attention and Model Weights: What Changes When an LLM Answers4
- Embeddings and Vector Search: How Semantic Retrieval Works and Where It Fails3
- RAG vs Fine-Tuning: Choosing Between Retrieval and Training (and When to Use Both)3
- Reasoning Models: Trading Latency and Cost for Accuracy on Hard Problems3
AI Architecture Patterns7
- Permission-Aware RAG: Keeping Retrieval Inside What Each User May See7
- Model Gateway and Routing: One Front Door for Every Model Call6
- Model Migration: Moving a Production Feature to a New Model Without Regressions6
- Latency Budgets for AI Features: Allocating Milliseconds Across the Request Path5
- Prompt and Prefix Caching: Why Prompt Layout Is a Cost Decision4
- Document AI Pipelines: Turning PDFs, Scans and Forms Into Data You Can Trust4
- Fallbacks and Graceful Degradation for AI Features3
The Agent Lifecycle10
- The Tool Execution Envelope: How Agents Act Safely on Real Systems13
- Workflow vs Agent: Choosing the Least Autonomy That Does the Job7
- Reliability Compounding: Why Multi-Step Agents Fail More Than Their Steps Suggest5
- Choosing Agent Orchestration: Raw SDK, LangChain or LangGraph5
- Single vs Multi-Agent Systems: When Splitting an Agent Earns Its Cost4
- MCP and A2A: The Protocols Connecting Agents to Tools and to Each Other4
- Agent Memory: What an AI Assistant Should Remember, and What It Should Not4
- Computer-Use and Browser Agents: When an Agent Should Click Instead of Call an API4
- Prompts, Tools and Skills: Instructions Are Not Execution4
- Budgets and Loop Control for AI Agents: Caps, Stop Conditions and Graceful Failure3
Evals10
- Golden Sets: Building and Maintaining the Eval Data That Decides Launches19
- Regression Gates: How AI Changes Move From Prompt Edit to Production Safely9
- Binary Pass/Fail Judgments: Why They Beat 1-to-5 Scores for AI Evals6
- Validating LLM Judges: Catch Rate, False Flags and Held-Out Labels6
- Confidence Intervals for Eval Pass Rates: The Arithmetic Every AI PM Needs6
- Error Analysis First: Read Traces Before You Build Eval Infrastructure5
- Evaluating RAG: Retrieval Recall, Groundedness, Correctness and Online Signals5
- Trajectory Evals: Grading AI Agents on the Path as Well as the Outcome5
- The AI Testing Pyramid: Contracts, Model Behavior and Real User Outcomes4
- Agent Benchmark Evidence: What a Pass Rate Can Actually Tell You4
AI Unit Economics10
- Cost per Successful Task: The Unit Economics Number for AI Products11
- The AI Metric Tree: Business, Product, Quality, System and Risk Layers7
- Metric Definitions for AI: Resolution, AI Revenue, Active Users and Definition Drift7
- Experimenting With AI Features: A/B Tests, Holdouts and the Traps Specific to Models6
- Effective Cost With Human Fallback: Why Resolution Rate Beats Model Price5
- Inference Cost Levers: What to Cut First, What It Saves and When It Backfires4
- Latency Metrics for AI Products: TTFT, Tokens per Second, End-to-End and Tails4
- Loss Functions and Product Costs: When the More Accurate Model Loses4
- Estimating AI Product Demand Without Counting the Same Work Twice4
- Diagnosing Metric Drops: When Every Segment Improves and the Total Falls4
Security & Governance7
- Prompt Injection: Why External Content Must Never Be Treated as Instructions7
- The Confused Deputy: When an Authorised Agent Is Tricked Into Misusing Its Access5
- The EU AI Act for Product Managers: Roles, Risk Tiers and the Calendar as of October 20265
- Data Retention and Residency for AI Products: Five Questions That Get Blurred5
- OWASP LLM and Agent Risks: Turning the Top 10 Lists Into Product Requirements4
- Copyright and IP Risk in Generative AI Products: Inputs, Training, Outputs and Contracts4
- US and India AI Rules: A Fragmented US Patchwork and India's DPDP Framework, as of November 20263
Models & Platforms9
- The Model Lifecycle: From Data to Retirement, and What a Platform PM Owns7
- Quantization for PMs: Cheaper, Faster Models and Where the Quality Goes5
- Self-Host vs API: The Break-Even Most Spreadsheets Get Wrong5
- Synthetic Data and Distillation: Training Smaller Models on Generated Examples5
- GPU Economics for PMs: Training, Serving and the Cost of Idle Capacity5
- Data Readiness: Can This Dataset Support the Product You Promised?5
- Serving and Speculative Decoding: Throughput, Latency and Traffic Shape4
- Developer Platform PM for AI: APIs, Activation, Limits and Change Management4
- Launching a Model: Staged Access, Documentation, Safety Gates and Deprecation Windows3
Building with AI10
- Prototype to Production: The Path From an AI Demo to a Reliable Feature7
- BDD Acceptance Criteria for AI Features: Given/When/Then When Outputs Vary5
- AI Coding Agents for Product Managers: Claude Code, Codex, Cursor and Antigravity5
- Practising an AI PM Build Exercise: Scope, Test and Explain a Prototype4
- Spec-Driven Development: Writing Specs That Coding Agents Can Build From4
- Prototype, Pilot or Product: What Has Your Experiment Actually Proved?4
- Prompt-Matrix Evals: Comparing Prompts and Models on Real Examples in an Afternoon3
- AI Research Synthesis: Clustering Interviews and Tickets Without Losing Traceability3
- Natural Language to SQL, Safely: Semantic Layers, Permissions and Checkable Answers3
- Synthetic Personas: Useful for Preparation, Never Evidence of Demand3
Leading AI Teams8
- Kill Criteria: Time-Boxed AI Bets With Pre-Registered Thresholds8
- AI Roadmaps: Capability Bets, Reliability Milestones and Model-Swap Cadence6
- STAR-L Stories: Explain Your Decisions and the Evidence Behind Them5
- Capability vs Outcome Commitments: Promising What You Can Deliver With AI5
- Working With Research Teams: Aligning Offline Metrics, Product Outcomes and Timelines5
- Decision Memos for AI Bets: Writing One Page Leadership Can Act On4
- Evaluating an AI PM Role: Ask Who Can Change the Product4
- Culture Conversations at AI Labs: Motivation, Disagreement and Safety Trade-offs3
The map is for orientation. If you would rather be told what to do in order, the start-here path sequences the same material by background and stage, and the courses walk it front to back.