ChatGPT Work Just Shipped: The 30-Minute Test That Tells You If It Replaces a Task You Hate
OpenAI launched ChatGPT Work (on GPT-5.6) on July 9: an agent that connects to your email, Drive, Slack, CRM, and calendar and ships finished docs, sheets, and decks. Here's the operator so-what and one move to make this week.
What happened. On July 9, 2026, OpenAI launched ChatGPT Work, an agent built into ChatGPT and powered by its new GPT-5.6 model. Unlike a normal chat, it can connect to your Slack, Teams, Google Drive, SharePoint, email, calendar, CRM, and project tools, gather context across them, and produce finished deliverables: documents, spreadsheets, presentations, reports, even simple websites. It runs on web, mobile, and a new unified desktop app (Mac + Windows) that's free for all ChatGPT users. Rollout started with Pro/Enterprise/Edu, expanding to Plus and Business.
The so-what for operators. This is the shift from "AI that answers" to "AI that does the task and hands you the file." The real value: it can now reach into the tools where your work actually lives and come back with a draft that's 80% done. For a small business owner, the honest question isn't "is it impressive", it's "which recurring task can I stop doing myself."
One move this week: the 30-minute replacement test
- Pick ONE task you do every week that's really gathering + formatting (a weekly numbers recap, a client update doc, a proposal from a template, a meeting-prep brief).
- Connect only the ONE tool that task needs, nothing more. Least access wins.
- Give it a tight brief: the goal, who it's for, the format, and where the source data lives. Ask for a finished draft.
- Grade the output on one scale: how many minutes did it save vs. how many did I spend fixing it? If it nets positive, it's a keeper. If not, tighten the brief and rerun once.
The caution. An agent with access to your email, Drive, and CRM is also an agent that can be tricked by content it reads (see prompt-injection and memory-poisoning). Rules: connect the fewest tools possible, keep a human approval step before anything sends or publishes, and don't point it at inboxes full of untrusted inbound mail until you've watched how it behaves. Keep sensitive client and financial data inside your business workspace, and keep it out of personal accounts.
Bottom line: adopt it for exactly one task, prove the time savings, then expand. That's how an operator gets paid by a model change instead of distracted by one.