August 28, 2026. On August 25, OpenAI introduced an Admin plugin for ChatGPT Work and Codex that lets a workspace administrator analyze usage, manage members and permissions, and approve spending requests through a single conversation. The number that matters for operators is in the same post: OpenAI says a ChatGPT Work agent in Slack now resolves about 45 percent of its own IT ticket volume, and that its support team cleared a backlog even as ticket volume roughly doubled. That is a concrete before-and-after for anyone weighing an AI support or operations agent for a client.
What OpenAI shipped
- On August 25, OpenAI introduced an Admin plugin for ChatGPT Work and Codex that lets an administrator review usage, manage members and permissions, adjust limits, and approve or deny spending requests in one conversation.
- The plugin is permission-aware. OpenAI says it works within each user's existing role, does not grant broader access, maps instructions to supported read or write actions, and surfaces higher-impact changes for review before they are applied.
- OpenAI reported that a related ChatGPT Work agent in Slack resolves about 45 percent of its internal IT ticket volume, and that the team cleared its backlog even as support volume roughly doubled.
The number that matters
Ticket deflection is the clearest ROI story in support automation, and OpenAI put a real figure on its own. A Slack agent that triages requests, checks approved policies, completes supported tasks and escalates exceptions resolved close to half the volume, which let the team move from reacting to issues to planning around live dashboards. For an agency that is a template you can reproduce on a client's stack, and it is the opposite of the ungoverned agent in OpenAI's Hugging Face incident report from the same week, where an agent found its own path to more access.
What it means for operators
Three choices make this work in a small business. Keep the agent permission-aware, so it inherits a real user's access rather than a broad service account. Route approvals to where the team already works, Slack or Teams, so a human clears the high-impact requests in seconds. And instrument it, turning ticket history into a dashboard so you can see what the agent is deflecting and what it is missing. Do those three and an AI operations agent stops being a demo and becomes something you can put in front of a paying client, whether on a custom build or inside GoHighLevel.
Frequently Asked Questions
It is a plugin for ChatGPT Work and Codex, introduced by OpenAI on August 25, 2026, that lets a workspace administrator review usage, manage members and permissions, adjust limits, and approve or deny spending requests through a single conversation instead of moving between analytics and settings screens.
No. OpenAI says it works within each user's existing role and permissions and does not grant broader access. It maps instructions to supported read or write actions, honors workspace approval requirements, and surfaces higher-impact changes for human review before they are applied.
OpenAI reported that a ChatGPT Work agent running in Slack resolves roughly 45 percent of its internal IT ticket volume, and that the team cleared its backlog even as support volume roughly doubled. It is a concrete example of ticket deflection from an AI support agent.
Connect an agent to the client's help desk, keep it permission-aware so it inherits a real user's access, scope it to policy-safe actions, route approvals to Slack or Teams for a human to clear exceptions, and turn ticket history into a dashboard to track what it deflects.