Chapter 1 What do PMs actually do?
Architect? had it right: professions are defined by their apprenticeships (Crites and Ward 2014).
PM had a bad apprenticeship. “Watch me write docs and go to meetings.”
AI killed that apprenticeship. Good. It was lying.
The grunt work — sizing tickets, triaging bugs, formatting specs, pulling funnel data — used to be how you earned your way into judgment. Survive the tedium, collect enough reps, eventually someone hands you a real decision.
That ladder is gone. AI does the grunt work (ProductBoard 2024). There is no tedium left to survive. You get thrown into judgment on day one.
Which means you don’t have two years to figure out if you like it. You need to know now.
1.1 The three questions
You won’t spend your day in Figma or SQL. You will spend it answering:
- What problem are we actually solving?
- Is it worth solving now, with these resources, for these users?
- How will we know if we solved it?
Everything else — the specs, the prototypes, the dashboards — is scaffolding built to answer those questions. AI is very good at scaffolding. It is not good at the questions. It will generate twenty answers to question three before you finish writing question one.
Your job is the questions.
1.2 What that looks like by context
The questions are the same everywhere. The stakes, speed, and cost of getting them wrong are not.
Startup (fewer than 50 people)
There is no PM team. There is you, a founder who has opinions about everything, and engineers who will build whatever you agree to in the next fifteen minutes. Question one is existential — wrong answer and the company doesn’t exist in six months. AI will help you generate hypotheses and prototype answers before lunch. It will not tell you which hypothesis is worth betting the company on.
The job here is saying “stop” as often as “go.” Most early PM work is deletion.
Scale-up (50 to 500 people)
You have a roadmap, a design team, a data function, and a quarterly planning cycle that produces documents everyone reads and no one follows. Question two dominates — there are now more things worth building than you can build, and the cost of picking wrong is losing a quarter, not a company. AI can generate ten prioritized backlogs from your customer data by tomorrow morning. Picking which one to execute is still on you.
The job here is saying “not yet” with enough rigor that it stays said.
Mega-corp (5,000+ people)
Question one was answered years ago. You are stewarding a product with users who depend on it and stakeholders who have staked their careers on it. Question three becomes the hardest — measuring whether you actually solved anything at the scale and accuracy that an engineering organization of hundreds requires. AI can query the data warehouse. It cannot tell you whether the metric moved because of your feature or because of the seasonal cohort effect your analytics team buried in a footnote last March.
The job here is maintaining clarity in a system that generates fog by design.
1.3 The artifacts that still matter
A PM’s work shows up in three durable forms, regardless of context:
- Decisions — documented, with alternatives considered and rejected. AI does not generate this. It generates options. You collapse them.
- Narratives — why this, why now, why us. PRD, memo, one-pager, five-slide deck. AI drafts these. You add the part the model cannot see: the scar tissue, the political constraint, the thing that will blow up if left unnamed.
- Systems — how the team measures, learns, and repeats. This is the one most PMs underinvest in, and the one AI is least equipped to replace.
AI makes #1 faster to populate and slower to trust. It makes #2 faster to draft and easier to outsource — until you’re in a room explaining it to a skeptic and the model isn’t there. It barely touches #3.
1.4 Try this
Exercise: Pick a product you use daily. Write three sentences:
- What is broken about it, and for whom?
- Is fixing it worth the effort right now, given everything else the team could be doing?
- What would you measure in thirty days to know if you fixed it?
Constraint: no new engineering for v1. You can only change copy, defaults, or the order of operations. Then find two people who disagree with your sentence 2 and update the doc.
If you found yourself obsessing over sentence 2 and couldn’t stop iterating, that’s worth noting. If you wanted to skip straight to the solution, note that too. The disposition you brought to that exercise is the disposition you’ll bring to the job.
In Chapter 2 we’ll look at what traits make the questions tractable — and which ones AI quietly makes more important.