Not all AI PMs are the same: 5 roles you’ll really see

AI PM has become one of the most hyped words of this decade within the product management industry. Companies are racing to hire AI PMs, while on the other side, the PM community is constantly talking about what AI PMs actually are.

I’ve been in product management for a decade now and have been diving into this topic since last year. In that time, I’ve used multiple AI tools and built AI agents. A couple of weeks back, Pranav Pathak, product director at Booking.com, posted about how there is a lot of confusion about what an AI PM is.

He started a series where every Saturday, he’ll post about a different AI PM role. The posts are extremely informative, but I wanted to share my own opinion about AI PMs after building AI tools on my own and talking to AI PMs from the community.

If you’ve ever looked at two AI PM job descriptions and thought they had nothing in common, you were probably right. Because in reality, from my experience, there isn’t one AI PM job; there are five. And each requires different skills and different instincts, as well as has very different day-to-day work.

In this post, I want to share how these AI PMs differ from each other, what they build, what skills one needs to excel in these roles, and what their output and career path look like.

But before that, let’s understand what an AI PM is.

What is an AI PM?

Countless definitions of AI PMs are floating in the market, and to be honest, it also highly depends on your specific company and the product. To make matters easier, I’ve simplified it for you.

An AI PM is the one who uses or builds AI tools to solve a customer problem. This includes the type of problems you solve and the kind of systems you work with. It has nothing to do with knowing Python, training models, or having “AI” in your job title.

Here are three factors that distinguish AI PM from PM.

1. Working with probabilistic systems, not deterministic ones

Traditional PMs work with predictable inputs and outputs. While I’m not debating that a traditional PM doesn’t work with uncertainty, an AI PM’s work requires a greater degree of it.

AI PMs spend their time defining:

• What “good output” looks like
• How to handle hallucinations
• How to design fallbacks
• How to set confidence thresholds
• How to measure model quality using evals
• How to prevent failures

2. Collaborating closely with data science/machine learning teams and influence model direction

AI PMs partner daily with:

• Data scientists
• ML engineers
• Research teams
• Infra teams
• Data teams

And they contribute by:

• Defining model success criteria
• Prioritizing quality improvements
• Aligning training data with product needs
• Deciding when a model should be retrained or replaced
• Balancing accuracy, latency, and cost

A traditional PM might request a feature. But an AI PM will work on building models and optimizing them so that they solve a business problem.

3. Defining success with AI-specific product metrics

Traditional PMs measure their products based on acquisition, activation, retention, referral, and revenue. And in some cases, if it’s a B2B product, the metrics are generally related to efficiency.

---

---

But AI PM measures their product in different ways. Metrics include:

• Precision
• Recall
• Latency
• Inference cost
• Hallucination rate
• Accuracy
• Intervention rate for agents
• Model scores
• Retrieval quality

Overall, it depends on what problem you’re solving and what your solution includes.

5 types of AI PMs

In this section, I want to detail five types of AI PMs that companies are looking for in the current market. This industry is changing very fast, so by November 2026, this list might have expanded, but I believe that the core of AI PMs will still remain the same:

 

 

Agent

In my own work, I identify as an AI agent PM. I believe this is the space I want to excel in for the next two to three years. But before understanding AI Agents PM, we first need to understand what an AI agent is.

An AI agent automates manual tasks. But this automation is different from the traditional automation. Traditional automation follows rules; AI agents make decisions by connecting different tools.

For example, if traditional automation works like this: If A happens, do B; if B fails, do C.

But AI agents don’t follow fixed rules. They decide what steps to take based on the goal and the context. AI agents PM work on building, managing, and scaling such AI agents. Their use case is the strongest in domains that have manual operations, such as manufacturing, legal, insurance, customer support, HR, etc.

What outcomes do they own?

• Cost reduction — This is their main KPI. They primarily focus on hours saved through automation, support tickets deflected, and manual tasks eliminated
• Efficiency gains — Time-to-completion for automated workflows versus manual processes
• Scale — Number of tasks the agent handles that would otherwise require human work
• Quality — Error rates, escalation rates, and user satisfaction with automated outcomes

Real examples across companies

• Notion AI — Building AI that can automatically create databases, update properties, and reorganize pages based on user intent
• Intercom’s Fin — An AI agent that resolves support tickets end-to-end without human intervention
• Salesforce’s Agentforce — AI agents that nurture leads and instantly quote while also coaching sales teams

How to become an AI Agent PM

1. Process mining — Learn to identify high-volume, high-impact workflows worth automating. The goal here is to identify those operations that don’t require creative decision-making but are highly repetitive in nature. You’ll mostly find them in customer support, sales, or operations
2. Flow design — Understand how to model complex workflows by branching them into smaller steps and logic
3. Risk assessment — Get better at predicting where automation will fail and designing fallbacks
4. ROI modeling — Build business cases showing cost savings from automation
5. Build — Lastly, build AI agents on your own. That’s the fastest way to learn

Conversational

Conversational AI PMs build products where the primary interface is dialogue, specifically chat, voice, or messaging systems. A great example is ChatGPT — you ask a question and get an answer, and in the backend, models are helping you to find the right answer.

Unlike normal UX, conversation has no fixed structure, so trust and clarity matter more than design widgets.

What outcomes do they own?

There’s a small overlap of the outcomes with the traditional PM roles, but they still aren’t the same:

• Engagement — Session length, messages per conversation, return rate
• Retention — How many users come back to the conversational interface
• Satisfaction — CSAT scores, thumbs up/down rates, qualitative feedback
• Containment — Percentage of conversations that resolve user needs without escalation

Real examples from companies

• Anthropic/OpenAI — PMs building the core conversational experience (Claude, ChatGPT)
• Discord — Clyde AI chatbot PMs building conversation flows for community management
• Duolingo — Teams building conversational practice with AI tutors

How to become a conversational AI PM

This role requires a very good understanding of modern user interfaces. Some skills that you can focus on are:

1. Conversation design — Study how humans naturally converse. But the important part is to understand how to help users with the information they need via a simple design. While this might be a design-related skill, it still has a lot to do with how the backend works and how to show it on the frontend
2. Prompt engineering — This is a must-have since the product revolves around it. Get systematic about prompt structure, few-shot examples, and chain-of-thought
3. Evaluation frameworks — Build robust ways to measure conversation quality beyond simple metrics

Tool

AI tool PMs are a little more technical than the other ones simply because of the nature of the product. They build tools that data scientists, ML engineers, and researchers use to develop and deploy models.

Some of the prime examples are experiment tracking platforms, model registries, training orchestration systems, feature stores, annotation tools, and evaluation frameworks. Their target user groups could be internal teams of ML engineers or an external product that focuses on making the lives of ML and data scientists easier.

What outcomes do they own?

• Developer productivity — Time saved in model development, experiments run per week
• Platform adoption — Percentage of ML teams using your tools
• Infrastructure reliability — Uptime, training job success rates, deployment velocity
• Cost efficiency — Compute costs per experiment, optimization of training resources

Real examples from companies

• Databricks — PMs building MLflow, experiment tracking, and model serving features
• Netflix — Internal ML platform PMs building tools for their recommendation teams
• Uber — Teams building Michelangelo, their internal ML platform for thousands of models

How to become an AI tool PM

1. Deep ML knowledge — This surely requires knowledge of ML concepts, and hence an experience in software development is highly beneficial
2. Developer experience — Study what makes great developer tools (documentation, APIs, error messages) and understand what their biggest points are
3. Infrastructure thinking — Understand distributed systems, compute optimization, scaling challenges
4. Technical credibility — Build enough depth that ML engineers trust your product decisions

Research

These PMs focus a lot on commercializing AI research. Their job is to take new model breakthroughs coming out of research teams and turn them into reliable, usable, and scalable products. They influence research directions by showing which capabilities matter most for customers, and they ensure the business can actually ship the science.

What outcomes do they own?

• Model performance — Accuracy, latency, cost-per-token on production workloads
• Research velocity — Time from research breakthrough to production availability
• API adoption — Downstream product teams using your models, revenue from model APIs
• Quality bar — Defining what “production-ready” means and ensuring models meet it

Real-world examples from companies

• Google DeepMind — PMs bridging Gemini research and Google product integrations
• Meta — Teams productizing Llama models for internal and external use
• Anthropic — PMs taking Claude research iterations to production API releases
• Hugging Face — Teams building production infrastructure for open-source models

How to become an AI Research PM

---

---

• Research literacy — Get comfortable reading ML papers, understanding architectures, and training techniques
• Evaluation design — Learn to build eval frameworks that catch issues before production
• Technical translation — Get better at explaining research concepts to non-technical stakeholders
• Production systems — Learn about serving infrastructure, cost optimization, and reliability engineering

Data platform

Among all the other ones, data platform AI PMs are the most analytical. They build the data infrastructure and pipelines that make AI systems work.

Instead of working on user-facing features, they focus on the pipelines, platforms, and processes that collect, clean, store, and deliver high-quality data to machine learning teams. They partner closely with data engineers, ML engineers, and data scientists, acting as the bridge between what models need and how data should be structured, governed, and delivered.

What outcomes do they own?

• Data quality — Accuracy scores, completeness metrics, and freshness of training data
• Pipeline reliability — Uptime/SLA for data availability, processing success rates
• Cost efficiency — Storage and processing costs per GB, optimization of data pipelines
• Model enablement — Number of models trained on your data and time from data collection to model training

Real-world examples from companies

• Stripe — PMs building data platforms that power fraud detection models. In fact, this product is found in almost every payment company
• Airbnb — Teams managing massive datasets for pricing and search ranking models
• Spotify — Data platform PMs ensuring music metadata quality for the recommendation systems to use
• Scale AI — Product teams building data labeling and curation platforms

How to become an AI Data PM

• Data architecture — Learn distributed data systems, storage formats, and data modeling at scale
• Quality frameworks — Build systematic approaches to measuring and improving data quality
• MLOps knowledge — Understand how data flows through the ML lifecycle
• Cost optimization — Get better at balancing data quality with storage and processing costs

All of these PMs have an overlap, and there are cases where a PM is balancing two or more roles. But in a broad sense, this is how I see the classification of AI PMs.

Also, all of these roles appeared in the last few weeks globally, and so the only way to get into these roles is to pick the ones that suit your own vision and interest and build simple tools by your own self in addition to reading books.

Final thoughts

AI PM roles are changing at a very fast pace. No one knows what’ll happen by the end of next year. But we can always look at the trends and data, and see what makes sense.

And with the rise of Claude Code, Cursor, OpenAI, and other AI tools, it’s a given that product managers will have to include AI in their work as much as possible. And this has already given rise to AI PMs.

What do you think about AI PMs? And in which category do you identify yourself? Feel free to comment and share your feedback.

Featured image source: IconScout