📰 Key Takeaways

OpenAI is testing a new mode where LLMs get deeply embedded in everyday digital life. Andrew Ambrosino, head of desktop apps, has given OpenAI’s desktop app access to and control over his inbox, Slack account, phone, and apps like Notion and Figma. He admits there’s risk here — for example, the AI might accidentally misuse sensitive info from private DMs while helping draft a document — but he thinks the occasional personal cost is worth it for work, and so far nothing has actually gone wrong.

Behind this test is ChatGPT Work, which OpenAI launched last month at the company’s lowest subscription tier — $20/month. The goal: let white-collar workers (accountants, investors, doctors, etc.) get the same kind of AI agent autonomy that software engineers already get from Codex — completing complex, multi-step tasks on their own instead of just answering questions. Thibault Sottiaux, head of core products, described this as letting ChatGPT complete an entire complex task “fully autonomously, delightfully, and safely,” calling it an embodiment of OpenAI’s mission to “bring everyone along.”

Commercially, this direction matters a lot for OpenAI: agents that run longer burn more tokens, which means higher revenue per user. And expanding beyond coding into other professions is key to the whole AI industry proving its massive training and compute investments are worth it. Right now, coding is still just a small slice of the professional work AI tools could theoretically empower. Meanwhile vertical-specific competitors — Harvey in legal, Clay in sales — have taken a “model-agnostic” approach, flexibly picking whichever model performs best at any given moment to win over these customers. That’s a real challenge for big labs like OpenAI.

💬 JudyAI Lab’s Take

The fact that OpenAI’s own head of desktop apps personally opened up his inbox, Slack, phone, Notion, and Figma to let an AI agent take over — testing this stuff on himself in public — is worth noting on its own as a signal for anyone watching the AI space.

This reflects a broader shift in the AI industry: from “answering questions” to “agents autonomously executing multi-step tasks.” ChatGPT Work is targeting non-engineering white-collar workers — accountants, investors, doctors — trying to replicate what Codex did for coding. But the business logic behind this move is pretty blunt: the longer an agent runs, the more tokens it burns, and that’s directly good for OpenAI’s revenue. It’s also a critical step for the whole industry to justify its massive spend on training and compute. Worth flagging: vertical players like Harvey and Clay have already gone “model-agnostic” — not locked into one model, but picking whatever performs best at the moment. That’s a real threat to the moat big labs are trying to build.

If you’re building AI products yourself, it’s worth asking: should your tools also stay flexible enough to swap models, the way vertical players do — instead of locking yourself into a single provider?

📅 Original Article Info

🔗 Further Reading