Plan backwards from the evidence
Most roadmaps list topics. Hiring managers ask for evidence: a feature you shipped with a model in it, an evaluation you designed, a launch you stopped. Plan backwards from that. Each level below ends in a checkpoint you can do out loud, because that is what the follow-up questions will test.
Level 1: foundation
Learn what a token is and why it governs both cost and context, how a model generates text one token at a time, what retrieval-augmented generation does, and what a single call costs on a current rate card.
Then price something. As of October 2026, Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens. A support reply that reads 3,000 tokens and writes 400 costs $0.006 plus $0.004: about one cent. Doing that sum aloud in an interview is rarer than it should be.
Checkpoint: explain why a model invents a citation, and name two product changes that reduce it, without using the word hallucination as the explanation.
Level 2: practitioner
Here AI PM work stops being reading. Build a golden set of 50 to 100 real examples for one task, write a yes-or-no pass rule for each, and grade outputs by hand before automating anything. Learn how a model used as a judge is checked against human labels, and why a team should report that agreement instead of trusting the judge by default.
Add the economics. Cost per successful task divides total spend by the tasks that actually succeeded. It is the number that exposes a cheap model that fails often, and the one finance will ask for.
Checkpoint: set a launch bar for one feature with a pass rate, a sample size, a segment floor and a cost ceiling.
Level 3: senior
Senior AI PM work is systems and money. Learn when an agent is worth its unreliability and when a fixed workflow does the job, why tool permissions belong in code rather than in the prompt, how routing and prompt caching cut cost, and how governance works when the EU AI Act, US state laws and India's DPDP rules all touch one product.
Checkpoint: take a feature that loses money on heavy users and present a plan that restores margin without lowering the quality bar, with the arithmetic.
Level 4: group and principal
At this level you choose which bets to fund, when to stop them and what the organisation shares. That means model strategy (build, buy or fine-tune), a common evaluation platform, a portfolio with kill criteria agreed in advance, and a regulatory roadmap that plans for known dates instead of reacting to them.
Checkpoint: describe a bet you stopped, the evidence that stopped it, and the date you had set before you started.
Getting evidence without the title
You do not need AI in your job title to have AI evidence. Add a model-backed feature to the product you already own and evaluate it properly. Build a prototype that fixes a real workflow for your own team and measure it on real inputs. Or publish an evaluation of a public product: twenty prompts, a rubric, your scores and what you would change.
Each of these produces a story with numbers in it. One honest story like that carries more weight in a hiring discussion than any certificate.
Sources and further reading
Questions about this topic
For an experienced PM, focused preparation on the first two levels takes weeks rather than months. The slower part is producing one piece of real evidence, a shipped feature or a measured prototype, that you can talk about in interviews.