OpenAI's Agent Bet: 98% Adoption Inside, Under 1% Outside

An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex — against fewer than 1% of individual subscribers.

The middle number is the one to sit with: just 17% of organizational subscribers were using the agentic coding tool. Near-total adoption inside the company, negligible adoption outside it. TechCrunch, which reported the study figures, framed that split as both the challenge and the opportunity for OpenAI.

It lands as the company pushes agents well past software engineering.

What OpenAI is selling

ChatGPT Work, released last month, is available on OpenAI's lowest subscription tier at $20 a month. It is a modified version of the company's Codex coding tool, and the pitch is to give non-engineers the functionality software engineers already get from agents: a tool that does not just answer questions but completes multistep projects on its own.

The target list is explicitly not developers. The product is intended to let white-collar workers field AI agents — hooking LLMs into the digital workflows used by accountants, investors, doctors, and everyone else whose day is dominated by their computer.

Thibault Sottiaux, who leads OpenAI's core product work including Work, told TechCrunch that "ChatGPT can actually do entire, very complicated tasks for you all autonomously in a way that is delightful and safe."

Why the adoption gap is not surprising

OpenAI's own non-engineering staff had a rough start with the underlying tool. Andrew Ambrosino, lead engineer for OpenAI's desktop app, said teams like communications and finance began using Codex "at a time that it was actively hostile to them—asking them about code and showing them, 'oh, you have an empty diff for this thing,'" referring to a technical readout meant for software changes. The fix took months: "we started to make it more general purpose between February and now."

If it took that long to make the tool legible to colleagues sitting inside the company, the sub-1% number among individual subscribers reads less like disinterest and more like a product that has only recently stopped speaking exclusively to engineers.

The price of the capability is access

Getting full value means handing the model the keys. Ambrosino's own desktop app has access to and control over his inbox, his Slack account, his phone, and apps like Notion and Figma.

He is candid about what that costs. Asked whether an agent writing a document might pull from a private DM without knowing it should not share the information, he said: "Yes. I'll do it for the job. I will take the personal hit here and there if I have to. And I haven't had to."

That is a reasonable trade for someone testing the product professionally. It is a different calculation for a finance team with client data in the same Slack workspace — and it is the question any buyer should answer before rolling agent seats out beyond a pilot group.

The commercial math favors longer-running agents

Agents that work for longer stretches burn through more tokens, which makes them more lucrative for OpenAI on a per-user basis. That alignment is worth naming plainly: the vendor's revenue improves as agent runs get longer, while the buyer's cost does the same. Anyone evaluating agent tooling should be measuring completed tasks per dollar, not sessions started.

Reaching new professions matters for the whole industry, not just OpenAI. Coding has been lucrative territory for AI labs, but it remains a small subset of the professional work these tools need to enable to justify the investment in training and computation.

Vertical rivals are not waiting

While labs focused on software engineers, vertical-specific competitors have been chasing those same customers — Harvey in law, Clay in sales — with a model-agnostic approach, plugging in whichever AI works best at the time.

Industry analysts see this as a real threat to the labs. "If the labs cannot rapidly get ahold of the key complementary assets needed to scale AI in the market, value will accrue elsewhere," Christian Catalini wrote on a16z's "It's time to build" blog.

For buyers, the practical read is that a general agent from a lab and a vertical tool are not yet the same purchase. The vertical products carry domain workflows; the lab products carry the frontier model and the wider surface area.

Meanwhile, capital keeps repricing the layer underneath

Hugging Face has been approached to sell at a valuation of $13 billion or more, Business Insider reported over the weekend. No deal has been reached and it is not clear who the startup has been talking to, though it has reportedly been talking to banks to help evaluate bids.

The comparison points are steep. Hugging Face last raised in 2023 at a $4.5 billion post-money valuation, in a round led by Salesforce Ventures with Alphabet, GV, and IBM Ventures participating. Earlier this year it turned down a $500 million investment from Nvidia that would have valued it at $7 billion, saying it did not want a single dominant investor to sway decisions. CEO Clem Delangue has said the company is "close to profitability" and only "recently started to touch the money that [it] raised three years ago."

The talks come amid increased interest in core AI infrastructure providers, as evidenced by Stripe's $7 billion acquisition of OpenRouter.

The same heat is visible one layer up. General Intuition, which is building a foundation model that trains generalized AI agents to move through space and time, is in talks to raise at a $6 billion pre-money valuation from new investors including Valor Equity Partners, Point72 Ventures, and Seven Seven Six, according to sources familiar with the matter. That would come just weeks after the startup raised $320 million at a $2.3 billion valuation. It was spun out last October from CEO Pim de Witte's game clip-sharing platform Medal, using hundreds of millions of hours of gameplay and "action labels" — records of which buttons a player pressed and when — as its initial dataset.

What to watch

The gap between a $2.3 billion valuation and a reported $6 billion one in a matter of weeks, and between 98% internal Codex use and under 1% outside, are the two numbers that describe this market right now. Capital is repricing the agent stack faster than end users are adopting it.

Three things worth tracking over the next quarter: whether the 17% organizational figure moves once ChatGPT Work has been in the market longer; whether the Hugging Face talks produce an actual deal or turn out to be offer-fielding; and whether vertical tools or lab-built general agents win the non-engineering seats that both are now chasing.