An AI product manager does the same core job as any product manager, deciding what to build and making sure it works for users, but for features whose output is probabilistic: the feature is sometimes wrong, its quality has to be measured with evals rather than checked once, and cost and latency are product decisions. It is also already the norm in job ads: of 109 product manager postings we collected on 29 September 2026, 77 (71%) mentioned AI, LLMs, machine learning, agents or a model vendor.
What does an AI product manager do that a regular PM does not?
The familiar parts stay: discovery, prioritisation, working with engineering and design, launching, measuring. Three things change.
- Defining "good" before building. A normal feature either works or has a bug. An AI feature is right most of the time and wrong some of the time, so the spec has to say what "good enough" means, on which cases, measured how. In practice that means eval criteria in the spec: a set of example inputs, what a good answer looks like, and the pass rate that justifies launch.
- Trade-offs that did not exist before. A bigger model may be more accurate, slower and more expensive per request. A PM decides where on that curve the product sits, and when to build, buy or fine-tune.
- Guardrails as product decisions. What the feature refuses to do, when a person must approve its output, what it tells users about its limits. These used to sit with legal or engineering; on an AI feature they shape the experience, so the PM owns them.
After launch, the PM reads production traces as well as dashboards: the actual inputs users sent and what the model did with them. That is where the next version of the spec comes from.
What do product manager job ads ask for in 2026?
Our sample is the same public fetch behind our count of AI job ads, re-run on 29 September 2026: 3,610 postings from Hacker News "Who is hiring?" (July to September 2026), Remotive and Arbeitnow. 109 had product manager or a close variant in the title: 95 from Arbeitnow and 14 from Hacker News, so it is a mostly European, startup-leaning picture in which Australia barely appears. Among those 109:
| What the ad mentions | Postings (of 109) | Share |
|---|---|---|
| AI, LLMs, ML, agents or a model vendor | 77 | 71% |
| Ownership or autonomy | 68 | 62% |
| Stakeholders or cross-functional work | 56 | 51% |
| Customers or product sense | 48 | 44% |
| Communication | 47 | 43% |
| Production or shipped systems | 33 | 30% |
| Agents or agentic systems | 29 | 27% |
| A portfolio, GitHub or side projects | 24 | 22% |
| Experimentation or A/B testing | 18 | 17% |
| A degree | 18 | 17% |
| Anthropic or Claude | 15 | 14% |
| Evals | 12 | 11% |
The counts file is public, with the method in it. Three readings:
- Ownership comes first. More than three in five ads ask for ownership or autonomy, and half for working across teams. The PM job is still the PM job.
- Agents have arrived in product roles. 27% of the ads mention agents, often because the product itself is an agent or is being rebuilt around one. Six mention the Model Context Protocol.
- Evals are under-asked. Only 11% of these ads say "evals", far fewer than mention AI. Our reading is that most employers have not yet written the skill into the job description, not that they do not need it. A PM who can read and write an eval stands out precisely because the ads rarely ask.
On experience: 40 of the 109 stated a number of years, with a median of five. 44 said remote, 36 hybrid and 23 on-site. 14 printed pay, across mixed currencies and seniorities, so we do not summarise it.
Do AI product managers need to code?
No, but they need to understand what the engineers are building well enough to make calls in the room. The ads bear this out: Python appears in 4 of the 109 and SQL in 6, so coding is rarely a requirement. What is asked for is judgement about systems: 30% mention production or shipped systems, and 23% system design or architecture.
The practical line is this. You should be able to read an eval result and say whether the feature is ready; estimate roughly what a change of model does to cost and speed; and build a rough prototype yourself with AI app builders, so you can test an idea with users before asking engineers for a sprint.
What does a portfolio look like for an AI product manager?
22% of the ads mention a portfolio, GitHub or side projects, which is more than you might expect for a non-engineering role. For an AI PM, a strong portfolio is a small set of real artefacts:
- A spec with eval criteria an engineer has reviewed and agreed can be measured.
- A prototype tested with a handful of users, and a note saying what the findings changed.
- A shipped feature, even a small one, with its eval running.
- An experiment and a decision memo where the decision follows from the data.
- A strategy or roadmap that says what you bet on and why, including build, buy or fine-tune.
How do you move from product manager to AI product manager?
If you are already a PM, you are closer than the title change suggests. The gap is usually AI literacy (what models can and cannot do, how evals work, where cost and latency come from) plus one shipped AI feature you can talk about in detail. Designers and engineers moving into product bring half of it already: designers know user research, engineers know the systems.
The route most people take is to volunteer for the AI feature on their current team, learn evals on it, and then describe that work in interviews. If your team has no AI feature, build one yourself as a prototype and test it with real users.
Where should you start this week?
Pick one AI feature you use every week and write the spec its PM should have written. Choose twenty realistic inputs, write down what a good output looks like for each, run them, and count how many pass. Then write one paragraph on what you would change and one on what it would cost. The AI product manager role page describes the day-to-day, and the free AI product management skill check takes a few minutes.
Square 1's AI Product Manager Bootcamp is twelve weeks, live on Zoom with one instructor, about 15 hours a week, in six blocks (AI literacy for PMs, discovery, shipping, measuring, strategy, and a partner's brief with a hiring sprint), each ending in a project and a gate you must pass: a spec an engineer reviews, a prototype tested with five users, a feature shipped with a partner engineer and its eval running, an experiment with a decision memo, and a strategy defended in a recorded viva. No code is required. Two on-demand courses, recorded by an instructor and graded by Nova, Square 1's AI tutor, cover pieces of it: Vibe-Coding a Product for building a working prototype without being an engineer, and Evaluating AI Systems for the eval side, which assumes some Python. All three are taking a waitlist today.