AI Product Manager is a work process, not a personality
Open a jobs board and you can feel the confusion. Companies want an AI Product Manager. They cannot describe the work. The posting is a regular PM role with "ChatGPT" in the requirements, or it is a science-fiction list that no one human has ever done.
I run a community and jobs platform for product people, so I see both sides every week. Candidates who can talk about models and cannot spec a fallback. Hiring managers who want "AI" on the roadmap and have no one who owns what the agent decides.
We are treating a new occupation as a personality.
Fluency is not the skill
Being good with the tools is table stakes now, the way being good with Jira used to be. It is not a differentiator and it should not be a title.
The work I actually hire for, and the work I am trying to credential, looks like this.
Find the problem worth pointing an AI at, which means having the spine to say no to nine demo-friendly ideas. Write a spec that includes the unhappy path: what the system does when it is unsure, what it must never do, where a human stays in the loop. Sit with the people who will live with the output, not just the people who will clap at the prototype. Lead a mixed room of engineers, data, clinical or ops, and sales without letting the model become the strategy. Measure an outcome a finance person would recognize. Own the miss in public.
That is product management with a probabilistic component. It is not a new species. It is a higher bar on the parts of the job that were always the job.
Occupations have a work process
This is why I am sponsoring a U.S. Department of Labor registered apprenticeship for an AI Product Manager occupation, filed with the DOL and with Florida. Not because a credential makes someone good. Because the occupation needs a named work process or we will keep hiring vibes.
The process I want in the world is unromantic. AI fundamentals so you cannot be snowed. Discovery. Specs with guardrails. How to lead an ML-adjacent team without pretending you are the researcher. How to take an AI feature to market. How to design agents that are small enough to test. Responsible AI as operating practice, not a slide. Measurement that survives contact with a P&L.
Twelve months of doing that next to someone who has shipped, not a weekend of prompt school.
If that sounds like ordinary craft, good. Ordinary craft is what is missing.
What I would tell a PM trying to become one
Stop collecting model news. A leaderboard score is not a product decision. Pick one real job in a real workflow and take it from demo to something a user keeps. Write the allowlist. Write the fallback. Instrument the outcome. Sit in the room when it fails.
Do that twice and you are more hireable than the person who can recite the last six model releases.
The title will keep spreading. The companies that get value will ignore the title and look for the work process. I would rather train that process in the open than watch another year of postings that mean nothing.
If you are a product leader trying to define this role on your team, or a PM trying to do the actual job instead of the costume, start here.