Feature Image

The "AI PM" is often described as a single role, but what the job actually involves depends on what is being built. Three product managers at the same company, on the same technology stack and in the same part of the organization, describe day-to-day work — and the definition of success — that differs sharply between them.

Henok leads product in Tech Research, where a win can be a team making the right technology choice, a methodology several teams use with more confidence, or a result that shows what is possible. Divita leads product for a recommendation foundation model, shaping what it learns and how its capabilities reach other teams. Humberto builds the experience users interact with directly, combining model capabilities, interaction design and user feedback into something valuable and trustworthy.

Shared vocabulary hides the differences. All three talk about "quality", "evals" and "shipping", but the meaning shifts with the problem space each owns. Impact looks different too: for research, a change in what the organization believes or funds is a major win; for a model PM, the payoff is a model improvement that lifts recommendations across multiple experiences; for an experience PM, it comes down to whether users love the product and return.

The research layer: product as technical leverage

Henok, Principal Product Manager, Tech Research, Personalization

The research product is technical leverage: applying scientific expertise where it can change what Spotify builds or how confidently it builds it. The role pairs today's product team needs with a view of what AI may make possible next.

AI is shifting which problems need research

Better models and tools let product teams do work that once required deep research involvement, freeing research to concentrate on shared methods, system design, and emerging capabilities where the right approach is still unclear. The questions that matter tend to span several teams — how to evaluate open-ended model behavior, for instance, or how to represent users and content in ways many products can use.

Prototypes are cheap; the distance between a prototype and a product that works reliably for hundreds of millions of users is not. A complex model can shine in a demo yet show no meaningful gain on the production tasks that matter. Evaluation, latency, cost, data and integration therefore become part of the research question, and they anchor investment decisions.

A promising architecture is not a roadmap

One project illustrates the pattern. An architecture that would unite capabilities now spread across separate systems began with a feasibility question. As evidence accumulated, the harder product question appeared: what should the organization do with it? A new architecture touches infrastructure, evaluation, interfaces and several teams' plans, and it would sit alongside systems that already work.

The product job is converting the larger idea into decisions the organization can reasonably make: start from a concrete problem, find a useful test, agree on what evidence justifies the next step. The path isn't all or nothing — the methods and components developed along the way can improve existing systems while the architecture question stays open. What began as exploration now informs both model and experience teams; finding a credible route through product needs, technical constraints and existing commitments is the harder work.

Long feedback loops in a fast-moving field

Research impact can take months to surface, often in another team's roadmap or metrics, and meanwhile the technology and the organization's priorities move. Influence becomes critical: inside the research team, providing product and strategic context without prescribing the answer; outside it, connecting work to decisions and bringing relevant teams in early, with a clear statement of what the evidence does and does not show. Start only when the research is complete and the decision window may have closed.

The pace of AI also demands deeper technical understanding than many prior PM roles. AI tools help in working through papers and prototypes, and keeping to one or two projects at a time allows judgment of the choices and obstacles involved. Success requires technical depth plus organizational judgment: identifying the questions that matter across teams, testing ideas before they become commitments, and linking evidence to decisions in time.

The model layer: a capability sold to other teams

Divita, Principal Product Manager, AI Foundation, Personalization

This platform is a language model purpose-built for recommendations and designed to reference real items on Spotify. Like any platform, it reaches users through other product teams — but its behavior is shaped through data, its outputs are probabilistic, and quality means something different to each team that integrates it. Those customers have their own roadmaps and metrics, and the same model must serve differing definitions of a good recommendation. A request from one team often points to something useful across several experiences, and spotting that shapes the model's direction.

Product fundamentals still apply, concretely: product-market fit means knowing which problems teams can solve with the model; a clear value proposition means explaining what it offers for their particular experience; business impact arrives through what they ship with it. Translating those needs into model behavior is where the work changes.

Shape the behavior, then prove it matters

Behavior is learned through training rather than specified in code. A spec alone will not make recommendations more relevant; the team must decide what the model needs to learn, which data can teach it, and how to know whether it has learned. Training takes time, so those choices set priorities for each cycle.

There are two real levers: what you teach the model and how you check what it learned — data and evals. Much of the job is deciding what the evals measure. What makes a recommendation good for a particular listener at a particular moment? Plausibility is only a starting point; the model must fit context, balance familiarity with discovery, and stay relevant over time. Evals turn those expectations into criteria that can be assessed consistently, linking what is measured in the model to what listeners experience. Offline evals reveal model behavior; online metrics show how it translates into the user experience. Work with partner teams to establish which improvements matter in their products sets priorities for the next training cycle.

Demonstrating what a shared model enables

A shared model should let one improvement benefit many experiences, but each team already runs a system optimized for its own needs. The first task is showing the shared model meets that experience's requirements for quality, latency and cost, with criteria defined and assessment agreed alongside the partner team. Meeting them is only the baseline. The longer-term question is the gain from building on a common model: can something learned in one experience improve recommendations in another, and can a team build on an existing capability to try something new? Part of the job is converting that promise into specific testable questions.

No single evaluation answers them. Tracking how the model develops across training cycles and how those changes translate across experiences distinguishes a promising result from a pattern worth planning around. A foundation model takes time and resources, and benefits compound as more teams adopt it; being clear about where benefits are showing up — and what remains to be established — guides prioritization. The case for the model is made over time through evidence teams can use in their own product decisions.

Good is not enough if nobody ships with it

Success means teams ship with the model. A good model is necessary but its value depends on another team being able to solve a problem that matters to them. Much of the enabling work looks nothing like model work: finding a team whose current system has hit a ceiling on a metric they care about, then making the model a compelling and practical path forward. That requires agreeing on what the capability must deliver, defining the integration contract, and owning the testing of the riskiest assumptions before the partner commits.

The team's best result this year came from a surface team that wanted an outcome the model was positioned to deliver; the collaboration worked because success measures were agreed before the architecture was. The role owns the model, and equally owns turning its capabilities into something another team can confidently ship and users can benefit from.

The experience layer: when the model is one ingredient among many

Humberto, Principal Product Manager for Sessions, Personalization, builds the surface users actually touch — where a vague mood has to turn into a set of songs, and the system must do something useful even when the request is ambiguous. His customers are Spotify's users directly. That is the simplest difference from the research and model roles, and it changes almost everything about the job. The other layers are his base technology; his task is turning that technology into something delightful. Because the experience team sits closest to the user, it has the most direct view of which model gaps and open research questions matter most right now — which makes the dependency on the other layers less a bottleneck than a shared prioritization signal across the stack.

Designing for open-ended input

Traditional products constrain the user's action: a button tap, a menu selection, a search keyword. The response is predictable, and so is the test — you know what the user did, what the system should do, and how to verify it. An AI experience inverts that. Input is open-ended: vague, underspecified or complex requests that map to no single entity in the catalog, with agents interpreting intent and constructing a response.

This reshapes constraints that once looked fixed. Latency that would be unacceptable in a keyword search turns out to be perfectly tolerable in a conversational interaction, since users understand they are asking for something harder. But questions such as how the system should respond when it is not confident, or how to recover when something goes wrong, have no stable answers — and they belong to the product, not the model.

Evaluation as a continuous product

A/B testing and offline evaluation still matter, but they are not sufficient alone: when the input space is open-ended, you cannot predefine every scenario worth testing. So the team builds new instruments — datasets drawn from real usage patterns, feedback loops that reveal where the system misunderstood a request, and LLM-based judges alongside custom evaluators tuned to the interactions the product supports, used to measure whether a change is actually an improvement or merely different.

Evaluation is therefore not a phase before launch but a continuous product in itself. A general-purpose eval spanning the foundation model's many use cases is a starting point, not the finish line; the experience team must close the gap between "the model is generally good" and "this specific interaction was good for this specific user."

Conversation as a feedback channel

In a conversational product, users do not just interact with the system — they talk to it, and when something goes wrong they say so, in their own words, in real time. That directness did not exist on previous surfaces. Frustration with a single interaction is simultaneously feedback on the experience, the model and potentially the training data. A user saying "that is not what I meant" is more than a local product issue: the job is to build mechanisms that detect that signal at scale, interpret it, and pass it back through the collaboration chain so every layer improves.

For someone considering the role, this is the version worth describing. You have the most direct line to the user and the most tangible ownership of the outcome — but you build on systems you influence rather than control, for users whose expectations are being reshaped in real time by every other AI product they encounter. The craft lies in making something trustworthy out of ingredients that are probabilistic by nature.

No single "AI PM" skillset

Across the stack, the longer the teams work together, the clearer it becomes that there is no one simple new "AI PM" skillset. Product management remains the same craft it has always been: defining the problem clearly, knowing the customer deeply, measuring what matters, and building conviction through evidence rather than enthusiasm. What the AI stack adds is not a different job but a layer of specialization on top of those fundamentals — and those layers come in more flavors than people realize, varying from product to product. The role is varied and complex, and only gets more exciting with the rate of change.