4 min read

Career

Most PMs Trying to Become AI PMs Are Learning the Wrong Things

The courses teach model theory. The hiring screen asks what you have shipped with a model inside it. Here is a build-first plan for moving from PM to AI PM, with the timeline I think is realistic.

how to become an AI product managerPM to AI PMAI product management skills

Most PMs Trying to Become AI PMs Are Learning the Wrong Things · Products Decoded
By Debashis Nayak · Products Decoded23 Mar 2026

A lot of PMs who want an AI PM role are collecting certificates the way kids collect Pokémon cards. Course after course, every theoretical concept covered, and zero real-world experience shipped. I think that is the wrong investment, and I will say why.

Hiring teams are screening for something the courses do not produce. They want to know what you have built.

What are PMs learning that hiring teams do not care about?

Model theory, mostly. How transformers work, how training differs from inference, what a loss function is. Some of this vocabulary matters. You do need to know inference versus training, fine-tuning versus prompting, and what a vector database is for. But that is a weekend of reading plus some hands-on time, and it is where most people stop and call it progress.

What I have seen separate the hired from the ghosted is much more practical. In 2026 the AI PM job looks different from a PM job in 2022, and the difference shows up in three places:

  • Hands-on familiarity with the stack. Prompting, RAG, agents and evals. Talking intelligently about models is no longer enough.
  • Ownership of the feedback loop. PMs now design the evaluation framework. You define what good looks like, instead of waiting for a data scientist to hand you a metric.
  • Systems thinking. Teams are shipping agentic AI now, meaning software that writes code, browses the web and acts with some autonomy. That calls for thinking like a system designer, not a feature owner.

You cannot get any of those from a certificate. You get them from breaking something.

What does “shipped something with a model in it” mean?

It means a working prototype that has met real users or at least real inputs, and a document explaining the trade-offs you made. My roadmap suggested project ideas like a GenAI resume assistant, a RAG-powered knowledge base for enterprise docs, or a customer feedback sentiment analyzer. The topic matters less than the fact that you wrestled with latency, brittle prompts and weird user behavior.

I should say where I am coming from. I build Astrika and Jobtune, and both have an AI layer inside them. I will not dress that up as a credential, but it is the kind of thing a hiring team can open and poke at. If you want to see what a shipped AI product looks like from the outside, you can try Astrika or Jobtune.

A prototype also gives you something to say in an interview. “I learned about RAG” invites a polite nod. “I built a RAG knowledge base and here is what broke” invites a follow-up question, which is the whole point of the conversation.

What should the order be?

The roadmap I wrote on Medium put about 12 weeks of fundamentals before the build. Looking back at it, I would flip that. This is a judgment call, and I could be wrong for some people, but here is how I think about it now: start the build early, and let the building tell you which fundamentals to study.

Here is the full plan, with my edits.

PhaseWeeksWhat to do
PM fundamentals1 to 8Write one full PRD for a non-AI product to prove you understand the product loop. Read Inspired by Marty Cagan.
Practical AI basics9 to 12Learn the vocabulary. Karpathy’s Neural Networks: Zero to Hero and DeepLearning.AI short courses. Spend hands-on time in the OpenAI Playground and Claude.
Probabilistic intuitionIn parallelDeconstruct one AI product a week, such as Perplexity or Notion AI. Find the fallback UI when the model fails. Guess at the system prompt.
Build and ship a tiny AI product13 to 18Working prototype plus a PRD of your design trade-offs.
LLM ops and infrastructure19 to 20Latency and cost trade-offs. Is a model that is smarter worth 3x the cost and 2x the lag? Have a framework for hallucinations and bias.
Get visible and apply earlyThroughoutPost PRDs and product breakdowns publicly. Join builder communities. Apply before you feel ready.

The total is about 4 to 5 months at 10 to 15 hours a week. My edit: do not wait until week 13 to build. Start a small thing in week 3, and let it be bad.

Why does “probabilistic intuition” matter so much?

AI features do not give the same output twice. That changes how users trust the product and what the interface has to do when the model is wrong. The cheapest exercise I know is to open an AI product you use, make it fail on purpose, and look at what the interface does. Does it admit uncertainty? Does it offer a retry? Does it just sound confident?

Hallucinations and bias belong in this category as well. They are core constraints of the product you are managing, and the interviewer who asks about them wants to hear a framework, not a shrug.

Should I apply before I am ready?

Yes, and early. Fail a few interviews at the start. It is the fastest way I know to learn what companies are actually asking for, and it costs less than a fourth certificate. Post your work in public while you do it, because skills nobody can see do not get you hired.

A checklist you can start this week

  1. Pick one tiny AI product idea and open a blank project today.
  2. Write a one-page PRD for it, trade-offs included.
  3. Add a way to measure whether the output is any good (even ten test cases and a pass/fail column).
  4. Break it, then write down what broke.
  5. Post the write-up where hiring managers will find it.

I still have a question I cannot answer: how many of the PMs reading this already built something and did not count it? My guess is quite a few.