You wouldn't hand a new hire the master key on day one. So why give an AI agent the keys to your whole inbox?
AI agents don't just answer questions anymore, they do the actual work. In ChatGPT Business, agents can read and send email from shared Outlook mailboxes and change shared calendar events [1]. Claude can connect to Gmail, Calendar and Drive to search and send emails and manage your calendar [2]. Microsoft's Copilot can take on multi-step work in Outlook and other Microsoft apps [3], and Google announced its own agent for Gmail, Drive and Docs on October 8 [4].
Plenty of owners are already using them. In an SBE Council survey of 516 small employers, 58% said they use AI agents regularly or occasionally, and another 14% are planning to [5].
I think agents are worth using. But the real risk usually isn't the AI fumbling a reply. It's handing it more keys than the job needs.
What too many keys looks like
On October 8, researchers at Zenity Labs showed how one prompt sent to one public-facing AI agent on Amazon's cloud let them grab its credentials and reach other agents in the same account, including private conversations and secrets [6] (a research demo, not a real breach). The lesson learned carries over to any agent: it can only leak what it can reach.
The big tools are built with this in mind. Microsoft's Copilot asks for approval before sending email or changing a calendar [3]. Claude asks before each action in Gmail, Calendar and Drive [2]. ChatGPT's workspace agents let you choose which tools and data an agent can use, and ask before sending email [7].
Start with one job
Starting small is safer. It's also how you learn. You need to wrap your head around how these things work before eating the entire AI elephant at once.
- Pick one low-risk job. Drafting replies to quote requests, or summarizing a folder of project docs.
- Connect only what that job needs. One app or folder, not your whole inbox and Drive.
- Keep approvals on. Let it draft, and read everything before it goes out. If your tool can't ask before it sends, pick a job where it never touches Send.
- Keep it away from money and promises at first. No payments, customer commitments or employee records.
Once it's done that one job well for some time, give it the next key. Then the next. Along the way, you verify it is working to your satisfaction. The same way you would any new hire.
Want help picking the first job to hand an agent? Grab 15 minutes and let's talk it through.
| Book a 15-minute call |
Sources
- OpenAI: ChatGPT Business release notes
- Anthropic: Use Google Workspace connectors
- Microsoft Support: Copilot for long-running, multi-step work
- Google Cloud: Welcome to Gemini at Work 2026: Introducing the Gemini agent
- SBE Council: New Small Business Check Up Survey
- Zenity Labs: How A Single Prompt Collapsed The Entire Cloud Security Model
- OpenAI: Introducing workspace agents in ChatGPT