The decision to connect AI to your work email usually arrives as a small one. The tool is sitting right there, it can read a calendar and summarize an inbox, and the setup takes about four clicks.
Daniel Starr fields that request two or three times a week. He’s Managing Partner at Lightwire Solutions, a managed IT provider with offices in Decatur, Alabama and Hendersonville, Tennessee, and he’s the person small businesses call when something stops working. In a recent conversation on A Good Pour, he told me he never says no to it. He asks what the person is about to hand over.
That’s a different question than the one most owners are asking, and it’s worth sitting with before anything gets switched on.
Why does connecting AI to your work email need a decision first?
Because the connection hands the tool the whole account, not the one piece you had in mind. Daniel described an owner who wanted a chatbot on his calendar so he could ask it what was on his calendar. The connection that does that also reaches every message in his inbox.
Daniel isn’t against any of it, and I’d repeat his framing to anybody. He’s not there to judge what somebody wants to use a tool for, and he’s not going to gatekeep it. It’s your business. The stopping and thinking has to happen before the connecting, though, not after.
He compares calling your IT provider about AI to calling your CPA about a mortgage. Both sit inside the same industry. They’re different jobs. Your system administrator can connect a tool to your office. What that tool is allowed to see is a policy question, and it belongs to you.
What has to be true before you connect AI to anything?
Five things, and only one of them is technical.
Somebody decided what information is allowed in
Data classification sounds like an IT term. In practice it’s a short list of what your team can put into an AI tool and what they can’t. Daniel’s question about it lands hard: without that list, how would an employee know?
He puts it next to a filing cabinet. If somebody walked out your front door carrying your proprietary information, you’d call a lawyer that afternoon. The same information pasted into a personal account leaves just as surely, and nobody sees it go.
The account is the business version
The controls that fence in what an AI tool can reach live in the business and enterprise versions. A free personal login used for company work carries none of them. Daniel’s clients who do this well have the right version first and the policy written around it second.
You know what the tool can already reach
Copilot is built into Microsoft 365, which means it may already be sitting in your environment whether or not anyone decided to put it there. Configured loosely, it can surface SharePoint content the person asking has no permission to open. Daniel calls this the accidental part, and it’s the version he runs into most.
The team was trained
A tool switched on without training produces two groups. One won’t touch it. The other will paste anything into it. Daniel frames both in a single conversation: build the guardrails, then tell the first group they’re safe inside them and the second group they’re free inside them.
The work isn’t living in one place
A client of his built workflows inside an AI tool, hit a problem, reinstalled, and lost all of it. The work had been stored locally and nobody had asked where. Anything you build with AI is part of your operation now, so somebody has to be able to answer where it lives and what the business does on the day that tool is down.
Use this today
Open a blank document and run this with whatever AI tool your business already pays for:
I run a [type of business] with [number] people. We use [your tools, for example Microsoft 365, QuickBooks, a CRM]. Help me write a one-page list of what our team can and cannot put into an AI tool. Ask me questions until you are 95% sure you can complete this task, then give me the list in plain language a new hire would understand on the first read.
Ten minutes. What comes back isn’t a policy yet. It’s the list every policy starts from, and it’s the thing your team is currently guessing at.
A note on pace
None of this has to happen this week. Daniel’s own shop held a no-AI position for a while, on purpose, until they understood enough to put controls around it. Working out what you’re protecting before you connect anything isn’t falling behind. It’s the part that makes everything after it hold.
Questions people ask about this
Is it safe to connect ChatGPT or Claude to my work email?
It can be, on a business or enterprise account where an administrator has set the permissions and the company has written down what information is allowed in. On a free personal account used for company work, it isn’t. The risk sits in connecting the tool before anybody has decided what it may reach.
What is data classification, and does a small business need one?
It’s a written list of which information is fine to share with an outside tool and which isn’t. A small business needs it more than a large one, because there’s no second system catching the mistake. One page is enough to start.
Should I ask my IT provider about AI?
Ask them about the setup, the permissions, and the backups, because those are squarely theirs. What your team is allowed to do with AI is a policy question, and a good provider will tell you plainly that it isn’t their lane.
Why can Copilot see files my employee can’t open?
Microsoft 365 includes Copilot, and when the permissions aren’t configured it can pull from SharePoint content the user has no other access to. That’s a configuration problem rather than a flaw in the tool. Ask whoever manages your 365 tenant to check it.
The last part
The tools will keep getting easier to connect. The part that doesn’t get easier is deciding what you’re willing to hand them, and that decision has your name on it, not your IT provider’s.