How product managers can lead AI initiatives

Executive overview

Product managers need to move from "building products" to "solving the right problems" when working with AI. Rather than jumping at the technology, identify real user pain points first, then determine if AI is the smart solution. As every PM becomes an AI PM, success requires partnering effectively with research scientists who understand model training, managing uncertainty in research timelines, and collecting quality data—all while maintaining strategic focus on outcomes that matter.

Core insight: AI PM success depends on solving problems first, not adopting technology for its own sake.

The AI PM mindset shift

• Avoid the "shiny object trap"—only pursue AI when addressing a genuine user pain point
• AI PMs solve problems; general PMs build products—a fundamental reframe
• Future-ready PMs must get comfortable with research scientists as core team partners
• Research involves uncertainty; launches may take 6+ months with no guarantee of success
• Every PM should become an AI PM because personalization and automation now apply across industries

How to start without overengineering

• Don't use AI for your MVP—fake the AI experience with Figma prototypes to test market demand
• Use no-code tools (e.g., AutoML) to train custom models without coding expertise
• Start with adjacent products at your company that succeeded with AI; use those as proof of concept
• Use generative tools (ChatGPT) daily: rewriting mission statements, ideating user segments, generating feature ideas
• Begin by identifying data you already collect and asking what can be improved with it

Data and model fundamentals for PMs

• Think of a model like training a child's brain—repeated examples of patterns until it recognizes them
• Models output probabilities, not certainties; a photo classifier might be 70% confident it's a cat
• Data requirements depend entirely on the task (15–20 images for simple image recognition; thousands for voice recognition)
• Getting quality, diverse data is harder than using public datasets; companies that collect their own data build stronger products
• The training process outputs weights and patterns that the model uses to make predictions on new inputs

Real-world AI applications

• Google Glass translation feature: real-time transcription and translation between languages, breaking communication barriers
• Wind turbine maintenance: AI-powered drone imagery reduced inspection time from three weeks to hours
• X-ray analysis and loan models: teams with zero coding experience built functioning AI products in three weeks using no-code tools

Building buy-in for AI investments

• Reference adjacent wins: show leadership a successful AI product already launched in your company
• Propose small, reversible bets with clear rollback plans and capped downside risk
• For long-term model improvements, clarify early with hiring managers how progress is measured in research orgs (different from typical product launches)
• Trust compounds over time; failing fast and learning is only possible in cultures that welcome experimentation

Challenges unique to AI product management

1. Research uncertainty: results may not match hypotheses; you must motivate teams through setbacks
2. Direction changes: pivoting mid-project is common; requires strong leadership communication
3. Data scarcity: you may need to creatively source or synthesize data, even collecting it manually
4. Career progression differs: fewer launches in research orgs; clarify success metrics upfront

Resources and learning paths

1. Newsletters: MIT Technology Review, TLDR (AI is being "sprinkled in" everywhere)
2. Research papers: arXiv.org for cutting-edge work; follow researchers publishing real-time innovations
3. Coding fundamentals: Coursera (Stanford's "Introduction to AI"), Career Foundry, General Assembly, Coding Dojo
4. Learning modality matters: choose offline courses or community-based cohorts depending on your discipline
5. Hands-on pairing accelerates learning; pair with another PM or engineer trying to learn together

Marily's AI PM course structure (Maven, three weeks)

• Week one: product development lifecycle differences and idea generation (users don't know what they want until you show them)
• Week two: productionization, collaborating with research scientists, influence without authority
• Week three: career navigation, interviewing, resume building
• Nine workshops total with an end-to-end AI product build as capstone
• Notable student projects: loan models analyzing X-rays, recommender systems with three weeks of learning

What PMs should learn alongside tools

• Learn to code basics—even if no-code tools exist, understanding fundamentals changes your confidence
• Analogy: learning piano basics unlocks the ability to create music, not just follow sheet music
• Understanding how tools work prevents blind reliance and opens new problem-solving angles
• Shadow your company's research scientists for one hour weekly to absorb domain context

Hype and reality check

• ChatGPT is both overhyped and underhyped: it's transforming work, not eliminating jobs
• Writers feared job loss; in reality, ChatGPT enhances writing productivity
• Lesser-known AI applications: lie detection, security, ethical safeguards, accuracy improvements in healthcare
• Tech moves fast; diversify information sources to avoid tunnel vision on one tool

The future PM stack

• All future products will need personalized experiences and recommender systems
• Automation and technological progress require AI-centric thinking in every sector
• PMs freed from tedious work (PRDs, boilerplate) can focus on strategy and zero-to-one problems
• AI will unlock new areas of product management we haven't yet imagined