OpenAI’s ChatGPT Work Makes Hours-Long AI Agents an Enterprise Oversight Problem

OpenAI began rolling out ChatGPT Work on July 9, an agent built into ChatGPT that gathers context from a user’s connected apps and files, breaks a stated goal into steps, and keeps working on its own for hours to produce finished spreadsheets, slides, documents, reports, and web apps. Reuters reported that the feature rolled out first to Pro, Enterprise, and Edu plans, with Plus and Business accounts following within days, an order that puts paying business customers ahead of the general ChatGPT audience. The launch lands five months after a similar move by Anthropic briefly wiped billions off software company valuations, which makes ChatGPT Work less a standalone feature update than the newest entry in a race among OpenAI, Anthropic, and Microsoft to ship AI agents built to act unsupervised for hours at a stretch. That race is turning into an operational question for anyone who has to decide what an agent is allowed to touch while nobody is watching.

What ChatGPT Work Actually Does

PYMNTS reported that ChatGPT Work can take action across a user’s apps and files, spend hours on a project if needed, and produce sheets, slides, docs, web apps, and other finished materials, quoting OpenAI’s own announcement: “The best way to learn how to use ChatGPT Work is to give it a task you already know well: analyze a month-end budget variance, turn source materials into a marketing campaign brief, or prepare for a sales meeting.” OpenAI’s description of the oversight model is a checkpoint, not a step-by-step approval: “You can follow its progress, answer questions, change direction and approve important actions.” According to PYMNTS, the agent can take on an entire workflow from a single request and keep working while the user is away from their computer or phone.

ChatGPT Work is powered by GPT-5.6, a new frontier model OpenAI shipped the same day. PYMNTS reported that OpenAI CEO Sam Altman said one of the new models, GPT-5.6 Sol, is “54% more token efficient on agentic coding jobs” and is “as good or better” than competing models, telling CNBC the same day: “Every enterprise now is thinking about spend and the value they’re getting in exchange for AI, and this is what we really want to do.” On desktop, ChatGPT Work is available on every plan rather than gated to paid tiers, and Reuters reported that OpenAI paired the launch with a new ChatGPT desktop application and a hosted-websites feature that lets users build and share websites directly through the agent. The company’s earlier agentic offerings, Operator and deep research, were folded into ChatGPT Agent for individual users, alongside a separate Workspace Agents product for enterprise workflow automation; ChatGPT Work extends that lineup rather than replacing it.

The Agent Race Already Spooked the Market Once

OpenAI is not the first vendor to bet on agents that work for hours unattended, and the last time a competitor made that bet, investors reacted badly. In February, Fortune reported that Anthropic’s release of industry-specific plug-ins for its Claude Cowork tool triggered a broad selloff across enterprise software stocks, as investors worried that AI agents would make traditional software-as-a-service companies obsolete. Financial data providers took the brunt of it: FactSet Research Systems fell 10 percent, and S&P Global, Moody’s, and Nasdaq all declined sharply the same week, with the anxiety spanning sectors including legal services, financial data, and real estate software.

The selloff was not the end of the story. Gartner analysts pushed back in a research note quoted by Fortune, writing that “predictions of the death of SaaS and enterprise applications are premature” and describing Cowork and its plug-ins as “potential disrupters for task-level knowledge work” that are “not a replacement for SaaS applications managing critical business operations.” Reuters reported that Microsoft, an OpenAI investor, answered Anthropic with its own Copilot Cowork agent shortly after. ChatGPT Work is now the third major vendor to ship an hours-long autonomous agent within the same stretch of months, which looks less like the market settling the SaaS-replacement question and more like the three biggest AI vendors deciding to ship the same bet regardless.

Hours of Unattended Access Changes What Oversight Has to Mean

For a chat product, “follow its progress, answer questions, change direction and approve important actions” is a reasonable design. It assumes a person is available to notice when something needs approving. An agent that runs for hours across a user’s connected apps and files while that person is away from their computer, which is explicitly how OpenAI describes ChatGPT Work operating, shifts the real control point from continuous human review to whatever the agent decides counts as an “important action” worth pausing for. Neither OpenAI’s announcement nor the wire reporting reviewed for this piece specifies what triggers that pause, what gets logged while the agent runs unsupervised, or whether an in-progress task can be rolled back once it has already touched a connected file or app.

That gap matters more because of who gets access first. Pro, Enterprise, and Edu accounts, the tiers most likely to already have administrative controls, data loss prevention tooling, and IT oversight in place, are exactly where OpenAI chose to start the rollout. That is a defensible product decision: business customers are the ones asking for multi-hour automation. But it also means the first wave of real-world ChatGPT Work usage will happen inside organizations with the most connected systems for an unattended agent to reach, including shared drives, CRMs, ticketing systems, and internal wikis. The Anthropic selloff showed that markets already treat multi-hour autonomous agents as a structural threat to how software gets bought. IT and security teams evaluating ChatGPT Work face a narrower but more concrete version of the same question: not whether agentic AI will change how work gets done, but which of their own systems they are willing to let an agent reach without a person watching in real time.

What to Watch Next

Three vendors, OpenAI, Anthropic, and Microsoft, are now shipping agents built on the same premise: hand over a goal, walk away, and come back to finished work. Gartner’s framing, that this exposes how much day-to-day knowledge work remains manual and ripe for automation, is probably right as far as it goes. It leaves out the part enterprise buyers still have to solve for themselves: multi-hour agent autonomy does not remove the need for oversight, it just moves the burden from watching each step to designing the boundaries an agent cannot cross before anyone checks in. Whichever vendor makes that boundary-setting easiest, not just whichever ships the most capable agent, is positioned to win the next phase of this race.

Sources

• OpenAI: ChatGPT is now a partner for your most ambitious work
• PYMNTS: OpenAI Launches ChatGPT Agent That Executes Complex Workflows
• Reuters via BNN Bloomberg: OpenAI launches ChatGPT Work, deepening race for workplace AI tools
• Fortune: Anthropic’s Claude triggered a trillion-dollar selloff. A new upgrade could make things worse

Featured image: a paperless work desk with an iPad, iMac, and MacBook each showing a different app, by FirmBee (Pixabay), public domain (CC0) via Wikimedia Commons; cropped, resized, and converted to WebP.