Most people don’t think of Microsoft Teams as a place where they store personal documents. But most people have, at some point, shared something personal through a work communication tool — a medical certificate for HR, a copy of their passport for travel arrangements, a bank statement for expense reimbursement, a personal photo in a work-adjacent group chat.
Those files go somewhere. And since Microsoft 365 Copilot rolled out broadly, that somewhere now includes an AI system that can read, summarize, and reference them.
How Files in Teams Actually Work
When you share a file in Microsoft Teams, it doesn’t live in Teams itself. Teams is a communications interface. Files shared in Teams channels are stored in the connected SharePoint site for that team. Files shared in private one-on-one chats land in the sender’s OneDrive, with a sharing link extended to the recipient.
This is a technical detail most users never think about — it doesn’t affect how the sharing feels day-to-day. But it matters because the underlying storage is what Microsoft 365 Copilot has access to. Copilot isn’t just an AI assistant for typing help and meeting summaries. It’s an AI with access to your Microsoft 365 ecosystem: your email, your calendar, your SharePoint documents, your OneDrive — and by extension, every file you’ve ever shared or received in Teams.
What Microsoft 365 Copilot Can Do With Shared Files
Microsoft 365 Copilot is designed to answer questions across your entire organizational context. The pitch is productivity: ask “summarize the Q3 strategy documents” or “what did we decide in last week’s meeting about the budget?” and Copilot aggregates an answer from across everything it has access to.
This capability doesn’t have a clean boundary between “work documents” and “personal documents someone happened to share at work.” If you sent your passport scan to the HR channel three years ago, and that channel lives in a SharePoint site Copilot has access to, that document is part of the corpus Copilot can search and reference.
In practice, Copilot’s access is scoped to what a user would themselves be able to access in Microsoft 365 — it doesn’t escalate permissions. But that scoping is cold comfort if the user in question is an admin, an HR system account, or simply a colleague who was in the same channel.
The Gray Zone: Personal Documents in Work Systems
The category of documents people routinely send through work tools is not limited to spreadsheets and project plans.
Consider what HR-related sharing looks like: medical documentation to support a leave request, proof of address for payroll, passport or visa scans for travel, professional certifications, sometimes personal ID photos. These go through Teams because that’s the tool the employer has deployed, and because asking “where should I send my medical certificate?” typically gets answered with “just drop it in the HR channel.”
Consider what ad hoc file sharing looks like: someone screenshots a personal tax document to ask a colleague for advice, shares a photo of a family event in a work social channel, sends a personal contract draft to a trusted colleague for a second opinion.
None of these were sent to Microsoft. They were sent to a specific person or a specific team. But the infrastructure they traveled through, and the AI system that now reads that infrastructure, doesn’t draw the same distinction.
What Microsoft Says About Copilot and Your Data
Microsoft’s privacy documentation on Copilot for Microsoft 365 states that the service does not train foundation models on customer content by default. Tenant data — the emails, documents, and chats in your organization’s Microsoft 365 environment — is not used to improve Copilot’s underlying models. This is a meaningful commitment and a real distinction from consumer AI products that do train on user content.
But training-use and access-and-processing are two different questions. Even if your documents aren’t being used to train a model, they are being read and processed every time Copilot generates a response that draws on them. That processing involves sending content to Microsoft’s Azure OpenAI Service, where the AI computation actually runs.
Microsoft’s documentation acknowledges that Copilot prompts, responses, and the content it accesses to generate those responses are processed in the Microsoft 365 service boundary — which means they stay within Microsoft’s infrastructure rather than being sent to OpenAI’s consumer services. But they are still transmitted to compute infrastructure outside the original SharePoint or OneDrive container where they were stored, and they are processed by AI systems that your files were not originally shared with in mind.
The Permission Model Doesn’t Mean What Users Think It Does
A common response to concerns about AI reading your files is that access controls prevent the AI from seeing anything you haven’t authorized. This is technically accurate but practically misleading.
Microsoft 365 Copilot respects permissions — it doesn’t show you documents you don’t have access to. But most users in an organization have access to a substantial amount of organizational content, including files shared in public channels, department-wide SharePoints, and any document explicitly shared with them.
If you’ve ever received a file, you almost certainly have access to it in Microsoft 365. That means Copilot can reference it in responses to you. If you’ve ever shared a file with someone, they have access to it — and their Copilot can reference it in responses to them.
The permission model was designed around human access patterns, where most people don’t systematically review everything they have read access to. An AI assistant working across your entire permission scope does exactly that — and does it in response to prompts from every user in the organization whose access overlaps with yours.
Admin Visibility: What Copilot Interaction Logs Contain
There’s a separate question about what administrators of a Microsoft 365 tenant can see about Copilot usage.
Microsoft 365 provides admins with audit logs that include Copilot interaction data — the prompts users submitted, and metadata about what content was accessed in generating responses. This is a legitimate enterprise feature: organizations have compliance, legal, and security reasons to audit how AI is being used on their platforms.
But it means that the files you share through Teams, and the AI interactions that reference them, are logged in a way that is accessible to your organization’s administrators, your IT and security teams, and potentially, depending on your employer’s policy, to HR and legal.
This is true of a lot of enterprise software. The distinction with AI interaction logs is that they capture not just “this file was accessed” but summaries of content and the context in which it was retrieved — which can reveal the substance of what a document contains, even to administrators reviewing logs rather than the files themselves.
Practical Steps to Limit Exposure
The goal isn’t to refuse to use work tools — it’s to be deliberate about what you put in them.
Keep personal documents out of work systems. If HR asks for a document and provides a Teams link, consider whether a secure email to HR directly accomplishes the same thing, or whether the work system is genuinely the only option. Often it’s the path of least resistance rather than a requirement.
Understand what’s shared vs. what’s sent. A file dropped in a public Teams channel is accessible to everyone in that team and potentially to organizational AI systems. A file sent directly to one person in a private chat has a more limited footprint — but it still lives in Microsoft 365 infrastructure.
Review what’s in your OneDrive and SharePoint. If your organization has Microsoft 365 Copilot deployed, look at what documents live in systems Copilot has access to. Files you uploaded years ago as part of an onboarding process, or for a project that’s long over, may still be there.
Have a separate home for genuinely personal documents. Documents you need to keep but don’t want in work infrastructure — medical records, financial documents, personal IDs, family photos — belong in a system that exists for you, not for your employer. That means personal storage, not work-issued tools.
The Broader Pattern This Reflects
Microsoft’s situation is not unique. It’s the version of a pattern that appears across enterprise software in 2026: AI capability has been layered onto existing platforms, with access scoped to the existing data that was already there.
The underlying data — files, emails, messages — predates the AI. It was shared under one set of expectations (this goes to this person, in this context). The AI operates under a different set of expectations (this tool can access anything you’ve shared in this ecosystem).
Users generally haven’t reconciled these two sets of expectations. The files are already there. The AI is new. And the people who shared the files often did so before they knew an AI with broad access was coming.
This isn’t a reason to stop using enterprise software. It’s a reason to stop using enterprise software as your personal file system — or at minimum, to be clearer about the distinction between documents you’re sharing for work purposes and documents you’d prefer to keep genuinely private.
Keeping Personal Documents Separate
A private storage service that exists outside your work infrastructure is the structural solution to this kind of exposure. Not because enterprise tools are malicious, but because they’re designed to be accessible — to the organization, to the tools the organization deploys, to the AI capabilities that get layered on.
A service like daftei is built for the opposite purpose: a personal, private place for your documents, photos, and files that doesn’t share your content with employers, doesn’t run AI on your stored files to build features, and doesn’t use your content to train models. Files are encrypted with AES-256 at rest and TLS 1.3 in transit. No ads. No AI training on your content.
Work tools are for work. Personal documents belong somewhere else — and that distinction matters more now that enterprise AI has turned document storage into document access.