OpenAI Wants AI Agents to Run Your Workday. Can It Convince Anyone Beyond Coders?

Key points

• OpenAI launched ChatGPT Work last month, available on its $20-per-month subscription plan.
• An OpenAI-backed study found that in June 2025, fewer than 1% of individual subscribers were using Codex, the agentic coding tool ChatGPT Work grew from.
• The desktop app has around 20 million users, versus more than one billion people who use ChatGPT through a web browser.
• OpenAI's lead desktop engineer has given the app full access to his inbox, Slack, Notion, and Figma to test it properly.
• Vertical AI competitors such as Harvey (for lawyers) and Clay (for sales teams) are already chasing professional users with a focused approach.

Andrew Ambrosino has handed a piece of software the keys to almost everything: his inbox, his Slack messages, apps he uses to design and plan. As lead engineer for OpenAI's desktop app, he granted the program access to all of it, not because he has no concerns, but because he thinks it is the only honest way to stress-test what he is building.

"If I'm asking it to write a document, is there a possibility that it's going to pull from a private DM on that subject and not know that it's not supposed to share some info? Yes," Ambrosino told TechCrunch. "I'll do it for the job."

That product is ChatGPT Work, released last month and priced at $20 a month. Built on Codex, originally a coding assistant, it aims to extend that kind of help to people whose jobs have nothing to do with writing software: accountants, investors, operations managers, finance teams.

What does ChatGPT Work actually do?

Think of it as a digital assistant that doesn't just answer questions but completes whole tasks on your behalf. It can scan a Slack conversation about an engineering problem and produce charts, or pull together emails and research into a polished investment memo. OpenAI says its own employees are using it to set up weekly metrics reports automatically.

The key concept is an AI agent, software that carries out multi-step tasks on its own without you clicking through each step. You describe what you want; the agent figures out which tools to open, what information to pull, and what to produce.

One OpenAI engineer described asking the tool to look at a Slack thread and "make some charts," then receiving back genuinely useful plots. Sam Altman, OpenAI's chief executive, is apparently using it to plan holidays.

Why aren't more people using it?

The numbers tell a stark story. Inside OpenAI, 98% of employees were using Codex by June 2025. Among paying individual subscribers outside the company, that figure dropped to less than 1%. Organisational subscribers sat at 17%, a figure the draft omitted.

Part of the gap is design. Early versions of Codex showed non-coders raw technical outputs, things like empty code diffs, that meant nothing to them. OpenAI has spent months reworking the experience so that ordinary workers can understand what the tool is doing and why.

Ambrosino compares the current interface to skeuomorphism, the old design trick of making a digital calculator look like a plastic one. Familiar shapes helped millions make the leap to touchscreens; extra buttons that power users consider unnecessary matter when most people have never tried anything like this before. As we reported on 20 August, Slack is pursuing a similar logic with AI agents embedded inside team channels, letting colleagues watch and redirect the software before it does anything irreversible.

"Without these products in front of the model, experts would know how to get the same results, but you wouldn't get to a billion people using the thing," Ambrosino said.

What does this mean for workers?

OpenAI is pitching ChatGPT Work at repetitive, data-heavy coordination jobs: turning a spreadsheet into a live planning tool, assembling weekly reports, summarising long Slack threads into action points.

For the average office worker drowning in messages and documents, that's a genuinely useful pitch. Granting an AI agent access to your inbox and work accounts is a real decision with real trade-offs, though. Ambrosino acknowledges the risk of private information surfacing in the wrong context. His implied advice is to start with lower-stakes tasks and watch carefully what the tool does before widening its access.

OpenAI's commercial logic is straightforward: agents that work longer burn through more tokens, which makes them more lucrative per user. Reaching nurses, teachers, and shop owners matters far more to OpenAI's long-term finances than perfecting tools for coders. Vertical competitors such as Harvey and Clay aren't waiting, and they'll swap in whichever underlying model performs best, owing no loyalty to OpenAI at all. That's the real competitive threat here, and a billion casual ChatGPT users who've never touched an agent is both the company's biggest asset and its most urgent problem.