How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma)

We spend a lot of time talking about AI as a consumer tool. We talk about ChatGPT, Claude, and Gemini as if they are these magic boxes we ask for favors. But if you want to see where the real value is being extracted, you have to look at how high-output product teams are integrating these models into their internal infrastructure. Figma is a prime example of this.

Ezra Mechaber, a growth PM at Figma, recently pulled back the curtain on how their product team actually functions. They aren't just using LLMs for better grammar. They have built what they call a PM Operating System (PM OS) and a specialized agent known as AutoXP. This isn't just about saving time; it's about changing the fundamental nature of the PM role from a task-doer to a systems-orchestrator.

The PM Operating System: More Than a Prompt Library

Most companies have a Slack channel where people share prompts. Figma took a different route. They built a structured repository of thirty-one distinct skills. These aren't generic "write me a user story" prompts. They are fine-tuned workflows that sixty different product managers use to maintain consistency across the organization.

For founders, this is the first lesson: standardization is the precursor to scale. If you have ten PMs writing PRDs in ten different ways, your AI implementation will be a mess. Figma’s PM OS ensures that the input and output formats are predictable. When the data is predictable, the AI becomes reliable. They’ve essentially turned the tribal knowledge of their best PMs into a digital asset that any new hire can plug into.

Moving from Assistants to Agents

The real shift happens when you move from a tool that helps you write to a tool that helps you execute. Figma’s AutoXP is the centerpiece of their growth strategy. It’s an agent designed to run experiments. In a traditional growth team, running a test involves a lot of manual coordination: defining the hypothesis, setting the parameters, monitoring the data, and deciding when to kill it or scale it.

AutoXP handles the bulk of this logic. It operates with what Mechaber calls a single human checkpoint. This is a critical distinction for anyone building AI-first workflows. Total automation is usually a disaster waiting to happen, but total manual control is a bottleneck. By placing a human at the high-leverage decision point and letting the agent handle the execution, they’ve managed to increase their experiment velocity without increasing their headcount.

What This Means for the Builder

If you are building in the crypto or AI space, you are likely feeling the pressure to do more with less. The Figma model gives us a blueprint for the lean organization of the future. The goal isn't to replace the PM; it's to remove the friction that prevents the PM from thinking strategically.

Building a "PM OS" forces you to define what a good product process looks like. You can't automate chaos. If you don't know your own process well enough to turn it into a set of thirty-one skills, you don't have a process problem—you have a management problem. AI just makes that problem more expensive if you don't fix it first.

The Skeptic's View on 'Growth Agents'

I’m naturally skeptical of anything that claims to automate growth. Growth is often about finding the non-obvious leverage points that models are inherently bad at finding because they rely on historical data. However, Figma isn't asking the agent to be creative. They are asking the agent to be rigorous.

The agent isn't necessarily coming up with the next billion-dollar feature. It's managing the plumbing of multivariate tests. It’s making sure the data integrity is there and the logs are correct. That is a perfect use case for LLMs. It’s the boring, repeatable work that humans eventually get lazy about.

The Takeaway for Founders

Stop looking for the "one prompt to rule them all." Instead, start mapping out every repeatable task your team performs. Figma’s success with AutoXP and their PM OS isn't due to a secret model or a special API key. It’s due to their willingness to document their internal logic and wrap an AI layer around it.

• Standardize first: You cannot automate a workflow that isn't defined. Build your own version of a skills repository.
• Identify the human checkpoint: Don't aim for 100% automation. Find the one spot where a human must say "yes" or "no" and automate everything leading up to and following that point.
• Build for the team, not the individual: A tool that helps one PM is a utility. A tool that sixty PMs use is an operating system.

The future of product management isn't about being a better writer or a better ticket-manager. It's about being a systems designer. If you aren't building the machine that builds the product, you're eventually going to be replaced by someone who is.

The goal is to move from a world of manual tasks to a world of managed agents. Figma is just showing us the first iteration of what that looks like in practice.

We need to stop thinking of AI as a search engine and start thinking of it as a middle-manager. It needs clear instructions, a defined scope, and a supervisor. Once you have those three things, your output potential changes entirely.

Read the original at Lenny's Newsletter →