If your employees use Microsoft Copilot, ChatGPT, or Claude, those tools may already be connected to your email, Teams, SharePoint, and OneDrive. That changes the security question. Early worries were mostly about what people typed into a chatbot and whether they could trust the answer. Once an AI tool can read across your Microsoft 365 environment, it’s working with your business data directly, and it needs the same protection as any other system that does.
In a recent session, Xantrion talked this through with IT leaders from midsize organizations. Xantrion’s advice was that if you do nothing else, get three basics right. Clean up permissions, move people onto business subscriptions, and put every AI tool behind single sign-on. Then expand access in steps as you get comfortable.
The discussion produced seven takeaways for IT and business leaders.
1. AI Tools See Everything Your Users Can See
When you connect Copilot, ChatGPT, or Claude to Microsoft 365, the tool works as the signed-in user and can reach their files, mailbox, and Teams conversations. If an employee has permission to open something, the AI can find it. A confidential folder buried 50 clicks deep used to be protected mostly because nobody bothered to look. AI doesn’t get tired of clicking. If the permission exists, these tools will find their way in.
One participant rolling out Copilot gave a concrete example. Without guardrails, someone in logistics could pull up confidential finance documents just by asking the right question. Copilot needs particular attention because it lives inside your Microsoft tenant and is getting harder to switch off completely. So even if you haven’t formally rolled out AI, your data needs to be ready for it.
2. Basic #1: Fix Permissions Before You Connect Your Data
This is where to start. Before you turn on any AI connectors, check the access controls on your most sensitive areas, such as finance, HR, payroll, executive, and client-confidential folders. A useful test is whether someone could stumble into a folder through a shared link or an inherited permission. If they could, the permissions need work, and an AI assistant will eventually find the gap.
This applies whether your files live in SharePoint, OneDrive, or a traditional file server. Every other AI control you put in place depends on getting it right.
3. Basic #2: Replace Personal Accounts With Business-Grade Subscriptions
People use the tools that help them get work done, whether or not IT has approved them. Before OneDrive and SharePoint were standard, employees put company files in personal Google Drive accounts so they could work from home or share with clients. Giving them a sanctioned option mostly solved that. AI works the same way. One participant found staff pasting internal documents into free ChatGPT accounts to get summaries, and responded with user training and by blocking unapproved tools.
Business and enterprise plans for Copilot, ChatGPT, and Claude include commitments not to train on your data or expose it to other customers. They’re usually priced close to personal plans, and setup is often simple. Counting on employees to strip out Social Security numbers before pasting into a personal account won’t hold up for long. A company-managed account puts you in control of where that data goes.
4. Basic #3: Put AI Tools Behind Single Sign-On
Any AI platform other than Copilot should connect to your identity provider through single sign-on, for the same reasons you’d do it with any SaaS app. People reuse passwords, and people get phished. With SSO, the protections you’ve already set up for Microsoft accounts, including multifactor authentication, conditional access, and trusted device requirements, also cover your AI workspace. Without it, one stolen password could expose the payroll analysis or client files someone uploaded to Claude or ChatGPT.
Device controls help too. After finding employees opening confidential finance documents on personal phones while traveling, one participant is setting up conditional access policies so SharePoint and OneDrive content only opens on company-assigned devices. For more on these controls, see SSO Is Security Infrastructure: 7 Takeaways for IT and Business Leaders and Device Trust Builds the Foundation for Zero Trust Security.
5. Limit Connectors And Agent Features To What You’re Comfortable With
An AI tool that reads your data and makes suggestions carries a different level of risk from one that acts for you. Current tools blur that line, but admins have more control than many people realize. ChatGPT and Claude connect to your tenant through the Microsoft Graph API, so you can restrict them to read-only access. You can turn off individual capabilities, like creating calendar events or sending email. Connectors can run under a service account instead of the individual user. You can also disable agent features at the tenant level, such as Claude Cowork’s ability to control a browser and click through applications.
Xantrion uses this approach internally. One of its advisors has a ChatGPT agent that reviews his email, Teams chats, meeting notes, and recently changed SharePoint files for a client each morning, then sends him a briefing and updates a project board. The agent can read his Outlook calendar but can’t write to it. Participants also pointed out that the Microsoft 365 admin center now lists a growing number of AI agents that can connect to a tenant, so it’s worth reviewing that list regularly. Each restriction takes away some usefulness along with some risk. Make that trade-off on purpose rather than by default.
6. Roll Out In Phases, The Way You’d Open Firewall Ports
Start narrow and widen access over time. A practical first step is a business subscription with manual, one-off file uploads, which gives people useful results at low risk. Once permissions and data governance are in place, turn on connectors in read-only mode and train staff on the risks. When a specific use case needs more, open that one permission.
It works a lot like managing a firewall. Every open port carries some risk, but some have to be open for anything to work. The discipline is opening only what you need, for a clear reason, and reviewing it over time. One participant described the goal as swimming with the AI current while keeping a float, so you have somewhere to stop, take stock, and keep going.
7. Manage The Risk That’s Left Instead Of Waiting For Zero Risk
Not everyone in the room was convinced. Several participants said they take a zero-trust stance and don’t fully trust any AI vendor’s promises about training on customer data. Others argued that the major providers have strong reasons to keep those promises. Their business depends on trust, researchers and nation-state actors are watching them closely, and large enterprises have examined their controls in more depth than most midsize organizations could. Xantrion’s view is that a well-configured AI platform carries less risk of data leakage than many SaaS products, or an on-premises file server behind an outdated firewall.
Another participant raised a different worry: AI doing something nobody asked it to do. You can’t remove that risk completely, just as locks and alarms can’t guarantee nobody will break into your house. What you can do is set permissions and scope carefully, log and monitor activity, and treat AI output as a draft that someone reviews before it informs a decision. Opting out doesn’t make you safe either. Your competitors will get the productivity gains, and your executives and employees will expect these tools to be available.
Where To Start
If you’re deciding where to begin, start with the three basics: fix permissions, move everyone onto a business subscription, and put each AI platform behind single sign-on. After that, give connectors only the access each use case needs, and expand in steps.
If you’re evaluating Copilot, ChatGPT, or Claude, or want to check that your current setup is secure, start a conversation with Xantrion’s strategic IT advisors. They can review your permissions and identity controls with you and help plan a phased rollout.
