How OpenAI’s PM grades product progress using strict AI rubrics

This article summarizes an episode of Aakash Gupta’s video series featuring Abhi Muchhal, international growth PM at OpenAI.

Photo credit: Shutterstock

Product managers often waste time writing bloated requirements documents. Abhi Muchhal, international growth PM at OpenAI, argues that PMs should stop writing endless pages and instead build working prototypes first.

This approach speeds up product development while exposing the limits of automated workflows. Crucially, it also changes how the industry measures global AI adoption.

Software development suffers from scattered information

Modern product work happens across a chaotic web of analytics dashboards, team messaging channels, code repositories, and calendar invites. Operating efficiently in this environment requires a new way to manage tasks.

Muchhal relies on an automated setup that connects isolated data sources, limits permissions, and produces tangible products that teams can instantly inspect and ship.

The evolution of these tools changes how humans interact with software, moving users from simply asking questions to executing complex, multi-step workflows.

“With Codex, [ChatGPT] has truly become an agent. I can give it parts of my job to do end to end,” Muchhal notes.

To design this system successfully, leaders must follow these operational rules:

• Connect the real sources: An assistant only becomes useful when it can reach analytics dashboards, team channels, cloud storage, knowledge bases, repositories, calendars, and local applications.
• Define inputs and outputs: Ensure growth dashboards name exact markets, performance metrics, and target audiences to prevent vague AI hallucinations.
• Check work automatically: Use browser testing tools that allow automated systems to see the interface through screenshots, removing the need for a human to inspect every minor code change.
• Automate stable work: Schedule routine tasks, like message sorting, daily dashboard refreshes, and weekly update drafts, to run automatically before anyone even opens a laptop.
• Turn repeated decisions into skills: Create reusable workflows, such as a standardized growth experiment process that writes hypotheses and highlights risks before a decision meeting occurs.
• Set permissions by risk: Allow automated reading to run freely in the background, but mandate strict human review for high-risk actions like messaging customers.
• Build a prototype before writing a document: Ensure the working product serves as the primary output, leaving the traditional requirements document as a mere reference guide.

Instructions control the quality of the work
“Context is king,” Muchhal points out. He emphasizes that context goes beyond pointing to connectors or Slack channels. It means telling Codex specifically what kinds of posts matter and should be indexed.

Silicon Valley assumptions hurt global plans

Fixing internal workflows is easy compared to the challenge of winning users in emerging global markets.

The standard AI playbook was built around Western office workers: university students and software engineers at desktop computers. That assumption breaks down fast beyond Silicon Valley.

“In Germany, about 60% of adults with internet access are knowledge workers,” Muchhal explains. “In the US, that share drops to 40% to 50%. In Brazil, it’s only 10% to 20%. In India, even lower.”

Images remove the need to read and write

Faster development requires better skills

Measuring results replaces blind excitement