Recently I came across a public tweet summarizing an interview with the OpenAI Codex team, and it was dense with insight. It mentioned: the team has only 40 people — 1 PM, 2 designers, with the rest mostly engineers plus a few researchers — and the sole PM’s daily work mainly involves using Codex to process user feedback and issues and to do priority sorting.

What’s really worth noting here isn’t just that “the team is small.” It reveals a more important shift: AI is driving the cost of information processing, coordination, and execution inside organizations down to very low levels. Teams’ focus is starting to move from “human-driven processes” to “model-driven execution.”

1. The PM role is being restructured

In the past, a large share of a PM’s time went into collecting feedback, organizing issues, tracking status, pushing cross-team collaboration, and writing documents. Now, the most time-consuming and repetitive parts of that work are increasingly well-suited to AI.

The PM’s routine on the Codex team illustrates this exactly: AI can already help turn fragmented feedback into structured information, help with initial issue categorization, and help generate the basis for prioritization judgments. The PM’s value is therefore concentrating into a few high-value activities:

• Defining the problem
• Judging trade-offs
• Setting direction
• Making final decisions
• Aligning user, technical, and business goals

In other words, the PM’s center of gravity is shifting from “process hub” to “decision hub.”

2. High output from small teams will become more common

In a team of 40, engineering and research make up the overwhelming majority, which signals a clear trend: code generation and automation tools are amplifying the output of individual engineers.

This produces two outcomes:

1. Small teams can now accomplish things that previously required a much larger organization.
2. Team competitiveness increasingly depends on depth of AI usage, not just headcount.

In other words, for many teams in the future, the gap will show up not first in “how many people you have” but in “whether your workflows have been rewritten by AI.”

3. Feedback loops will become increasingly automated

The chain from user feedback to issue to priority sorting used to be long and heavily reliant on manual information handling. Now AI can do a lot:

• Automatically merge similar feedback
• Extract high-frequency issues
• Pull themes from community posts, tickets, and chat logs
• Do initial screening by frequency, impact, and severity
• Provide prioritization suggestions

This means product iteration will behave more like a real-time feedback system than a manual system driven by meeting cadence. The speed of information intake, synthesis, judgment, and execution will noticeably increase.

4. Design and research won’t disappear, but will become fewer and sharper

The team still keeps designers and researchers, which matters. It shows that experience, aesthetics, and user understanding haven’t lost value — only the way the work gets done has changed.

In the future, what will be more valuable is not purely low-repetition execution but:

• Defining standards
• Developing taste
• Validating hypotheses
• Judging boundaries

Designers become more like gatekeepers of high-level experience; researchers become more like problem-definers and evidence validators. AI can help them generate drafts faster, but it can’t replace their judgment about “what good experience looks like.”

Takeaways for us

1. Fix the workflow first, then add people

Many teams, when they hit efficiency problems, instinctively hire. Now the first questions should be:

• Which feedback can be automatically structured?
• Which problems can have AI produce a first draft?
• Which decisions can be converged first through data and models?
• Which repetitive coordination actions can be systematized?

If the workflow isn’t restructured, simply adding people amplifies the old problems.

2. Treat information flow as a product

Feedback, issues, requirements, and experiment results are all information flows. What AI is best at is exactly this kind of high-frequency, fragmented, repetitive information.

So what a team really needs to build is not just features, but a system for “how information flows, aggregates, gets judged, and lands.”

3. Organizations need more judgment, not more middle layers

When execution is accelerated by tools, what really creates the gap is:

• Direction judgment
• Trade-off ability
• Understanding of users
• Grasp of technical boundaries

Organizations will become flatter, but the bar for judgment for each person will be higher.

Implications for individuals

• PMs should learn to use AI for upfront organization — turning chaos into structure and information into decisions.
• Engineers need stronger product sense, because AI raises engineering output and makes it easier for engineers to directly participate in defining problems and validating solutions.
• Designers should strengthen aesthetic, narrative, and experience judgment, because mechanical image production will become increasingly less scarce.

Conclusion

What this tweet really reflects isn’t just one small team’s efficiency — it’s a bigger technological trend: software is eating organizational processes.

In the past, software helped people complete tasks; now AI is starting to help teams complete “organizational actions” — reading feedback, dividing tasks, writing summaries, setting priorities, generating solutions.

Future competition isn’t just about how many people you have, but how many organizational actions you can hand to AI, and how much human time you can free up for genuinely high-value judgment.