August 26, 2026
TechCrunch's August 24 report on OpenAI's ChatGPT Work led with a striking internal statistic: an OpenAI-backed study found that 98% of OpenAI's own employees used Codex, the coding tool ChatGPT Work is built on, in June. Meanwhile, only 17% of organizational subscribers and under 1% of individual subscribers used it. OpenAI now has 20 million people on its joint desktop/mobile app and over a billion people prompting ChatGPT online -- so this isn't a story about people not knowing the tool exists or not having access to it. It's a story about a tool that gets used constantly by the people who built it, and barely touched by everyone else who paid for it.
The TechCrunch piece quotes Andrew Ambrosino, lead engineer on OpenAI's desktop app, framing the challenge plainly: without products in front of the model, experts would know how to get the same results, but you wouldn't reach a billion people. That's the right diagnosis of the problem, but it's worth pushing on why the 98% figure is so high in the first place. OpenAI's own employees aren't getting more value out of Codex because the interface is friendlier for them. They're getting more value because they already know exactly what task they're asking it to do, and exactly how their own codebase, tools, and workflows fit together. They don't need to figure out scope -- they already have it in their heads. Everyone else has to supply that scope themselves, in a chat box, task by task, with no guardrails telling them what a 'good' request even looks like inside their own company's Slack, Notion, or Figma setup.
ChatGPT Work is a genuinely capable product, and this isn't an argument that it or agents like it are poorly built. For a power user who already understands their own systems -- how their CRM, their invoicing, their support queue actually connect -- a general agent that can move across email, Slack, Notion, Figma, and a calendar is a real force multiplier. That's exactly the population inside OpenAI hitting 98% adoption. The problem is that 'general purpose' and 'usable by everyone who buys it' are different claims, and the TechCrunch data shows the second one failing at scale. A horizontal agent asks every new user to do the work OpenAI's own engineers already did for themselves: define the task, know the systems, know what a correct outcome looks like. Most people buying a $20/month seat don't want to do that work. They want the tool to already know it.
This is the core argument for purpose-built software over a single agent expected to handle everything: when a tool is scoped to one team's actual workflow, the user doesn't need to be an expert prompt engineer or a systems architect, because someone already built that logic into the tool. A CRM built for how a specific sales team actually tracks deals doesn't need the rep to explain their pipeline stages every time. A helpdesk built around a support team's actual ticket flow doesn't need an agent to guess which Slack channel a customer complaint should route to. That's the gap between the 98% and the under 1% in miniature: the OpenAI engineer already has that context; a purpose-built tool bakes the context in so nobody else has to have it memorized. It's why we built ViibeStack around letting teams generate internal tools and workflows that match how they actually work, rather than asking every employee to become fluent in directing a general model.
General AI agents are only going to get more capable over the next few years -- better reasoning, better tool use, better memory of context across sessions. None of that closes the gap TechCrunch documented, because the gap isn't a capability shortfall. Codex is plenty capable; 98% of the people who deeply understand their own systems prove that. What's missing for everyone else is scope: a bounded, pre-defined understanding of the specific workflow they're trying to automate. That's a product design problem, not a model problem, and it's why the adoption numbers for a horizontal 'agent for everything' will keep looking uneven compared to tools built for one team's specific job, whether that's marketing campaigns, project tracking, or HR workflows. We think the more useful question for teams evaluating AI tools right now isn't 'how capable is this agent,' but 'does this tool already know what I'm trying to do, or do I have to teach it every time' -- a question worth asking before deciding between a general AI app builder and a bolt-on agent.
Sources