OpenAI Wants Everyone, Not Just Programmers, to Have an AI Agent — Here's Why That's Hard

First, what's an "AI agent"?

Most people's experience with AI so far is chatting with something like ChatGPT: you type a question, it types back an answer. An "AI agent" is a step beyond that — instead of just answering, it actually goes and does things for you. It can open apps, browse the web, edit documents, and complete multi-step tasks mostly on its own, checking in with you along the way instead of waiting for you to spell out every single step.

The gap OpenAI is trying to close

OpenAI already has an AI agent built specifically for programmers, called Codex, and it's wildly popular inside OpenAI itself — reportedly used by 98% of OpenAI's own employees. But outside the company, adoption falls off a cliff: only about 17% of people at companies that pay for the business version use it, and less than 1% of individual paying subscribers do. In other words, the people building it love it, but almost nobody else has figured out how to make it part of their daily work yet.

To close that gap, OpenAI is launching ChatGPT Work, a $20-a-month product aimed squarely at people who aren't software engineers — accountants, doctors, marketers, and other office workers who've never touched a line of code but still do complex, multi-step tasks all day.

What it actually does

ChatGPT Work connects to the tools people already use — email, Slack, Google Calendar, Notion, and the design tool Figma — and tries to complete tasks across all of them without much hand-holding. That also means it can end up mixing information from different places in ways that might surprise you. One of OpenAI's own engineers admitted that if you ask it to write a document, there's a real chance it could accidentally pull in information from a private Slack DM.

What OpenAI learned from a rival

Interestingly, OpenAI says it originally built Codex's web version assuming its AI model was capable enough to just be told what to do and left alone. That didn't work well. It was actually a rival product — Anthropic's Claude Code — that showed a better way: have the AI check in conversationally with the user before barreling ahead, rather than disappearing to do a task and coming back with a surprise. OpenAI has since adopted that same back-and-forth approach.

The problems still unsolved

A few things are making this harder than building a coding assistant:

• Permissions are confusing. Users reportedly struggle to understand what the AI is and isn't allowed to access, and some features only work on the web version, not mobile.
• It can get expensive fast. One tester burned through 80 million "tokens" (the units AI models use to measure how much text they process) in just four days — racking up roughly $65 in computing costs against a $20 monthly subscription. OpenAI is effectively subsidizing heavy users by more than 3x right now.
• There's no simple pass/fail. Code either works or it doesn't, which makes it relatively easy to train an AI to get better at it. A business presentation or strategy document doesn't have that same clear right-or-wrong answer, which makes training much trickier.
• Specialized competitors exist. Companies like Harvey (built for lawyers) and Clay (built for sales teams) are going after the same office workers with tools built for one specific job, and they're not tied to any single AI model.

The bigger bet

OpenAI's argument is that none of the interface tweaks matter as much as simply having a better underlying AI model — and that as models keep improving and get more efficient with how many tokens they use, the rest of these problems (cost, complexity, trust) will get easier to solve over time.