OpenAI Wants Every Office Worker to Have an AI Agent. Will They Trust It?
OpenAI has spent the last few years selling ChatGPT as something you talk to. Its newest push, ChatGPT Work, is trying to sell something different: an AI agent you hand things to. The pitch is simple — give it a multistep project, walk away, and come back to it done. The harder question is whether people who aren't engineers are actually ready to do that.
From "coding tool" to "everyone's tool"
ChatGPT Work launched last month on the standard $20/month subscription tier, and its job is to take Codex — OpenAI's coding agent — and stretch it well past software engineering. The target now includes accountants, investors, doctors, and other white-collar workers who've never touched a terminal.
That's a big shift. Codex was built for people who already think in tasks, files, and diffs. Most office workers don't. So OpenAI has been reworking the product to hide the parts that used to confuse non-engineers — like raw technical readouts — and lean into general-purpose project completion instead.
The adoption gap tells the real story
The numbers OpenAI shared are blunt about where things actually stand. Internally, 98% of OpenAI's own employees used Codex in June. Outside the company, only 17% of organizational subscribers touched it, and less than 1% of individual subscribers did.
That's not a rounding error — that's a product that works great for the people who built it and barely registers for everyone else. ChatGPT Work is OpenAI's attempt to close that gap by making the agent legible to people who just want a task done, not a codebase reviewed.
What it can actually do
Early testing turned up some genuinely useful examples: auto-populating a Google Calendar straight from formatted emails, building a financial dashboard that updates itself, assembling a queryable database of space launches, and generating weekly research digests pulled from academic sources.
That's real multistep, autonomous work — not autocomplete, not a chatbot answering questions one at a time. It's an agent making a plan and executing it.
The catch: cost and trust
Two things stood out as rough edges. First, cost: four days of casual use burned through more than 80 million tokens, running roughly $65 — over three times the monthly subscription price. OpenAI says efficiency and pricing will improve, but right now the meter runs faster than people expect.
Second, and bigger: trust. To do any of this, the agent needs real access — your inbox, your Slack, sometimes your banking data. Permission settings reportedly felt circular and hard to reason about, which is exactly the kind of friction that makes people hesitate to grant an AI that level of reach. Competing tools like Claude Code, Harvey (for law), and Clay (for sales) have leaned toward tighter conversational feedback loops instead of full autonomy — letting the user stay in the loop rather than stepping away entirely.
That's the real tension in this whole category: the more useful an agent is, the more access it needs, and the more that access has to be earned rather than assumed.
The bigger picture
OpenAI closing the gap between "engineers use agents" and "everyone uses agents" is a sign of where the whole industry is heading — multistep, autonomous, hands-off AI work becoming the default, not the exception. The products that win that shift won't just be the most capable ones. They'll be the ones people actually trust enough to hand the keys to.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.