OpenAI is building AI agents for everything. Will everyone use them?
TechCrunch Tim Fernholz ● Covered by 2 sources
OpenAI is letting ChatGPT run parts of people’s work life, from inboxes to calendars. The bet is that agents only feel useful when they get real access — and that scares people for a reason.
Based on reporting by TechCrunch, Tim Fernholz — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI’s newest push is simple to describe and uncomfortable to use: give the model more of your digital life, or don’t expect it to be very useful. Andrew Ambrosino, who leads engineering for the company’s desktop app, is already letting it into his inbox, Slack, phone, Notion, Figma, and more. He’s fine with the risk. For OpenAI, that kind of trust is the whole point.
The company’s big wager is ChatGPT Work, which launched last month and sits on OpenAI’s lowest paid tier at $20 a month. It is aimed at white-collar workers whose days are split across email, spreadsheets, docs, and internal tools. The pitch is not that the model answers questions better. It’s that it can take on multi-step jobs by itself, the way coding agents already do for software engineers.
That matters commercially as much as it matters technically. Longer-running agent tasks use more tokens, which is better business for OpenAI on a per-user basis. And the company can’t live on programmers alone. Coding tools have been the easiest place to show off agentic AI, but they’re a small slice of the professional work AI labs need to cover if the huge spending on training and compute is going to make sense.
The gap between inside and outside OpenAI is stark. A study backed by the company found that in June, 98% of OpenAI employees were using Codex, while only 17% of organizational subscribers and less than 1% of individual subscribers were using it. That’s the whole problem in one chart: people at OpenAI are willing to hand the machine the keys, but most customers are still standing in the doorway.
So the product has to become less like a coding tool and more like a useful, if slightly nosy, assistant. Ambrosino said the team spent months making Codex more general purpose after non-engineers found the old version actively hostile. OpenAI also knows that buttons still matter, even if some employees think users should just ask the model directly. Early on, discoverability beats elegance. Or, put another way, if the interface feels like a magic box, at least people will open the box.
Opinion: OpenAI is right that agents need real permissions, but the company is also asking normal people to trust a system that still makes basic setup feel like a maze. That’s not a small UX problem; it’s the product. The wider industry keeps pretending access will feel natural once the model gets smarter, when the bigger truth is that most users want automation with guardrails, not a digital intern with the master key.
My take — AI-written commentary, not fact-checked reporting
OpenAI is right that agents need real permissions, but the company is also asking normal people to trust a system that still makes basic setup feel like a maze. That’s not a small UX problem; it’s the product. The wider industry keeps pretending access will feel natural once the model gets smarter, when the bigger truth is that most users want automation with guardrails, not a digital intern with the master key.
Read more about this at: TechCrunch