ChatGPT Work: OpenAI's Agent That Delivers the Finished Output
Agentic AIOpenAI didn’t just ship GPT-5.6 on July 9. Alongside the model release, they launched ChatGPT Work — an autonomous agent that doesn’t generate answers about your workflow. It completes the workflow and delivers the output as a finished document.
What ChatGPT Work Actually Does
Built on GPT-5.6 and Codex infrastructure, ChatGPT Work accepts a single prompt and operates across applications and files for extended periods, returning finished deliverables: Excel models, Word documents, marketing campaign packages, adapted assets for different markets. The system handles multi-step projects with sustained context — not just session-level memory, but continuity across a complete project arc.
The integration layer is substantial. A “Unified Plugins Directory” spans twelve enterprise platforms: Google Drive, SharePoint, Slack, Microsoft Teams, Gmail, Outlook, Salesforce, Adobe, Zoom, LinkedIn, GitHub, Canva, and Dropbox. Users can invoke specific integrations with ”@” mentions, or allow the system to identify relevant data sources automatically based on task context. The desktop app adds local file access beyond cloud storage, extending the agent’s reach to whatever lives on the user’s machine.
The Enterprise Guardrail That Matters
Previous OpenAI agent products required users to approve every significant action — a friction model that reduces risk but also reduces utility. ChatGPT Work introduces an “Auto-Review” function: a secondary model that audits critical actions before execution rather than after.
During adversarial red-team testing, this layer “[blocked] 100 percent of attempts to extract protected data.” That specific claim is what enterprise procurement teams need to bring autonomous agents into environments touching sensitive systems. Whether the real-world attack surface matches the red-team conditions is a separate question, but the framing signals that OpenAI understands the enterprise barrier: it’s not capability, it’s the security sign-off.
Rollout follows a tiered structure — Pro, Enterprise, and Edu users first on web and mobile, followed by Plus and Business customers. Pricing runs on a consumption model shared with Codex and other workspace tools, scaling with task complexity rather than flat monthly billing.
The broader context: OpenAI is simultaneously retiring Atlas, its AI browser that launched eight months ago. The folding of Atlas capabilities into ChatGPT signals a consolidation strategy — fewer specialized agent surfaces, one primary product with expanding scope.
The So What
ChatGPT Work shifts the accountability question in enterprise AI adoption. When an agent produces the finished deliverable — not a draft, not a suggestion — the line between “tool that assists” and “agent that acts” blurs in ways procurement and legal teams are still working out. The twelve-integration directory and auto-review guardrails are specifically designed to lower that procurement barrier, but the real test is whether enterprise teams will accept AI agents with write access to their core systems at the scale of actual workflows. The teams that figure out the governance model for autonomous output — who signs off, what gets reviewed, what gets trusted — will run ahead. The teams still treating every AI output as a draft to rewrite are leaving the productivity gap on the table.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →