OpenAI is trying to move AI agents beyond software development and into the everyday workflows of accountants, investors, doctors, operations teams, and other computer-based workers. Its latest effort, ChatGPT Work, connects a model to workplace tools and asks it to complete multistep tasks rather than simply answer questions.

The expansion is strategically important, but early evidence points to a difficult adoption problem. TechCrunch reports that OpenAI employees use its agentic coding tool at very high rates, while usage among external subscribers remains limited. The gap suggests that building capable agents is only one part of the challenge; making them understandable, controllable, and trustworthy for people outside engineering may determine whether the market grows beyond technical early adopters.

ChatGPT Work takes agents beyond coding

ChatGPT Work was released last month and is available on OpenAI’s $20-a-month subscription tier, according to TechCrunch. The product is described as a broader version of Codex, the company’s coding agent, designed for users who may not work with source code or command-line tools.

The core proposition is access to existing digital workspaces. OpenAI’s desktop application can connect with email, Slack, Notion, Figma, and other applications, allowing an agent to gather information and take action across several systems. Examples cited by TechCrunch include creating weekly metrics reports, turning spreadsheets into planning tools, assembling investment research, and producing charts from workplace conversations.

That approach puts ChatGPT Work in the same broad category as products such as Claude Code, Perplexity AI’s browsing agent, and other AI agents that operate across a user’s existing software rather than inside a single isolated chat. OpenAI is betting that a general-purpose agent can serve many professions if the surrounding product handles the complexity of tools, permissions, and task planning.

The company’s product leaders frame the change as a move from conversational assistance to autonomous execution. That is an OpenAI position, not an independently verified assessment of how reliably the system performs across different workplaces.

The adoption gap is the central test

The strongest evidence of the problem comes from usage data cited by TechCrunch from an OpenAI-backed study. In June, 98% of OpenAI employees reportedly used Codex, compared with 17% of organizational subscribers and less than 1% of individual subscribers.

Those figures are vendor-backed and should not be treated as an independent measurement of the overall market. They nevertheless illustrate the distance between internal familiarity and external adoption. OpenAI engineers work in environments where coding agents fit naturally into existing processes. Most other users must first understand what to delegate, how to provide context, and how to review the result.

OpenAI engineering leader Andrew Ambrosino told TechCrunch that the company initially saw non-engineering employees struggle with Codex because the product exposed technical concepts such as code changes and empty diffs. The team has since worked to make the experience more general-purpose, with the goal of reducing the amount of specialized knowledge required to use an agent.

This also explains OpenAI’s emphasis on visible controls and workflow buttons. A technically sophisticated user might be able to invoke the right capability through a prompt, but mainstream users may not know that the capability exists. Ambrosino argued that discoverability matters during the early stages of adoption, even if some controls eventually disappear as users become more familiar with agentic software.

More access creates more risk

Giving an agent access to workplace systems also changes the safety problem. The model is no longer only generating text for a user to inspect. It may read private messages, retrieve records, update documents, create calendar events, or act across multiple applications.

Ambrosino acknowledged a risk that an agent could draw information from a private direct message while preparing a document and fail to understand that the information should not be shared. He told TechCrunch that he accepts the risk in his work testing the product, while noting that he had not encountered that outcome.

That anecdote is not evidence that the problem has been solved. For enterprise buyers, the relevant questions include whether access can be limited by application, project, data type, or user role; whether actions require approval; and whether administrators can audit what the agent viewed and changed. The source material does not establish how ChatGPT Work handles each of those controls.

The product also faces the reliability problems of the software it must navigate. Many business systems were not designed for autonomous agents, and inconsistent interfaces, incomplete records, and ambiguous permissions can make a seemingly simple task difficult. An agent that successfully extracts a poorly formatted calendar once is useful; an enterprise system must perform predictably across thousands of varied cases.

Why the bet matters to AI builders and buyers

For OpenAI, broader agent adoption could increase the value and usage of its models because longer-running tasks consume more inference resources. More importantly, the company needs to reach professions beyond software engineering if it wants AI agents to support a much larger share of knowledge work.

Competition is already forming around narrower workflows. Harvey focuses on legal work, while Clay targets sales-related processes, according to TechCrunch. Those companies use a model-agnostic strategy, allowing them to select different underlying models when performance or cost changes. OpenAI, by contrast, is attempting to make one consumer-facing product useful across many functions.

That creates a trade-off for builders. A general agent can reduce the need to buy and integrate multiple specialized tools, but a vertical product may offer better permissions, terminology, review processes, and domain-specific reliability. Product teams evaluating ChatGPT Work will need to compare the convenience of a broad assistant with the controls and predictability of a focused system.

The adoption figures also suggest that user experience may matter as much as model capability. Engineers can tolerate configuration, technical feedback, and occasional debugging. Finance, communications, and operations users are more likely to need clear task boundaries, visible sources, reversible actions, and straightforward escalation when the agent is uncertain.

What to watch next

The first signal will be usage of ChatGPT Work relative to OpenAI’s broader ChatGPT audience. TechCrunch reports that the joint app is used by 20 million people, while OpenAI says more than a billion users prompt ChatGPT online. The comparison is not a like-for-like adoption metric, but it shows the scale of the company’s challenge.

Enterprise buyers should watch for published information about active usage, task completion rates, error rates, permission controls, audit logs, and the frequency with which users must intervene. OpenAI’s willingness to disclose those details would make it easier to distinguish genuine workflow adoption from curiosity-driven trials.

The market should also monitor whether users prefer general agents or specialized products such as Harvey and Clay. Another important signal will be whether OpenAI can make agents useful without requiring customers to reorganize their existing software stacks or accept broad access to sensitive data.

Creati.ai perspective

OpenAI’s news is less about adding another assistant than about testing whether autonomy can become a normal workplace interaction. ChatGPT Work addresses a real limitation of chat interfaces: users often need help coordinating information and actions across several systems, not merely drafting a response.

But mainstream adoption will depend on constrained autonomy, not autonomy alone. Builders and enterprises are likely to reward products that explain what they accessed, ask for approval at consequential moments, and fail safely when instructions or data are ambiguous. OpenAI has demonstrated the ambition to bring AI agents to non-engineers; the next question is whether its product design can earn the permission to act.