8 min read

Is Microsoft Copilot safe for law firms?

The AI isn't the risk. Your SharePoint permissions are. Here's what actually determines whether Copilot is safe for client work, and why the free version is a different product entirely.

Minimal illustration of a shield and Microsoft Copilot icon on a deep navy background with a red redline mark.
By The Redline Editors

Most articles answering this question stop at one fact: Microsoft 365 Copilot doesn't train on your data. True, and almost beside the point. The thing that actually gets a firm into trouble with Copilot isn't the model. It's the SharePoint site somebody shared with "Everyone except external users" back in 2022 and never revisited. Copilot doesn't create that problem. It just makes it instantly findable by anyone who asks the right question.

So the honest answer is: safe, if you configure it properly first. Unsafe, if you switch it on and hope. Here's what "properly" actually means for a firm with five lawyers and no dedicated IT security function.

Free Copilot and Microsoft 365 Copilot are not the same product

Copilot Free, the chat box available to anyone with a Microsoft account, cannot see your firm's email, documents, or Teams messages. There's no client data for it to leak because it has no access to any. It's a general-purpose assistant with a browser window, not a firm tool, and treating it as one is where confusion starts.

Microsoft 365 Copilot is the licensed add-on that actually reads your work data: files in SharePoint and OneDrive, emails in Exchange, chats in Teams, grounded on your tenant's Microsoft Graph. That's the entire value proposition, and it's also why the stakes are different. According to Microsoft's enterprise data protection documentation, prompts and responses inside Microsoft 365 Copilot carry the same contractual commitments your firm already has for Exchange and SharePoint, and none of it trains the underlying foundation models. That's a real, auditable commitment. It just doesn't cover the oversharing problem at all, because oversharing isn't a Copilot setting. It's a permissions setting that predates Copilot entirely.

The actual risk: Copilot reads what you can already read

Before Copilot, finding an overshared file required knowing it existed: the right site, the right folder, the right filename. Practically nobody stumbled onto it by accident. After Copilot, an associate can type "summarize the settlement terms we offered in the Hendricks matter" and get an answer in twelve seconds, whether or not that associate was supposed to have access to that file. Copilot didn't leak anything. It surfaced something that was already exposed, just previously too hard to find.

This is now a well-documented pattern across Microsoft 365 tenants generally, not a legal-specific quirk, and Microsoft's own deployment guidance recommends running a SharePoint oversharing assessment before Copilot goes live, not after. For a law firm, that assessment needs to specifically catch: matter folders shared firm-wide instead of to the assigned team, "Everyone" or "Everyone except external users" groups still attached to client sites, anonymous or organization-wide links generated years ago for a one-time share, and old client sites that should have been archived and locked down when the matter closed.

None of this is exotic security work. It's a few hours in the SharePoint admin center, or one engagement with an MSP if your firm doesn't have the bandwidth. It has to happen before you flip the Copilot switch, not as a cleanup project after someone notices a problem.

Where the EU Data Boundary matters, and where it doesn't cover you

If your firm handles EU personal data, there's a wrinkle worth knowing about. Microsoft 365 Copilot is generally an EU Data Boundary service, meaning EU customer traffic stays within the EU boundary for processing. But Anthropic models deployed inside Microsoft's Copilot stack are currently excluded from that EU Data Boundary commitment. If your firm's Copilot configuration routes queries through an Anthropic model rather than OpenAI's, that specific traffic may not get the in-region processing guarantee you'd assume applies blanket-wide. Ask your Microsoft rep or your IT provider which model your tenant is actually using before you assume EU data residency is automatic.

Privilege: still untested, still your responsibility to protect

No published case has yet ruled on whether using Copilot on privileged material waives privilege. The governing framework is still ABA Formal Opinion 477R, which says cloud tools including AI don't waive privilege as long as the firm takes reasonable precautions to protect confidentiality. A properly configured Microsoft 365 Copilot deployment, with tight permissions, sensitivity labels enforced through Purview, and DLP policies actually turned on, is a defensible version of "reasonable precautions." A default deployment with SharePoint permissions nobody's audited since onboarding is a much harder position to defend if a client ever asks.

We covered the mechanics of what counts as reasonable precautions in our confidentiality checklist, and the disclosure question that often comes up alongside privilege in our piece on when lawyers must tell clients they're using AI. Both apply directly to a Copilot rollout.

What to actually do before turning Copilot on

  • Run a SharePoint and OneDrive oversharing audit first. Fix the "Everyone except external users" groups and dead anonymous links before anyone gets a Copilot license.
  • Confirm you're licensed for Microsoft 365 Copilot with commercial data protection, not relying on the free consumer version for anything touching client files.
  • Ask which model processes your firm's queries and whether EU Data Boundary coverage actually applies to your traffic, if you handle EU client data.
  • Apply sensitivity labels with real enforcement, not just a "Confidential" tag with no policy behind it, to matter files that should never surface in a Copilot answer.
  • Archive and lock down closed-matter sites. A live Copilot license querying five years of unmaintained client folders is where this goes wrong fastest.

None of this is a reason to avoid Copilot. For a small firm already living inside Microsoft 365, it's one of the lower-friction ways to get real AI assistance into daily drafting and research, and the contractual data protection terms are genuinely solid. The mistake is treating "Microsoft says it's enterprise-grade" as the whole answer. The AI is the easy part. Your permissions structure is the part that actually decides whether this is safe.

Get a full walkthrough of what reasonable precautions look like in practice, plus a running list of vetted tools, in The Lawyer's AI Starter Kit, free. For how Copilot stacks up against other options for a small practice, see the best AI tools for lawyers.

FAQ

Does Microsoft train its AI models on my firm's documents?

Not if you're on Microsoft 365 Copilot under a commercial license. Microsoft's enterprise data protection terms state that prompts, responses, and data accessed through Microsoft Graph are not used to train the underlying foundation models, and this is a contractual commitment, not a settings toggle. Copilot Free, used outside a licensed work account, carries no such guarantee for the same class of protection.

What's the actual difference between free Copilot and Microsoft 365 Copilot?

Copilot Free is a general chat tool. It can't see your firm's SharePoint, Outlook, or Teams data at all, so there's nothing for it to leak from those systems, but there's also no enterprise data protection, no Purview audit logging, and no conditional access tied to it. Microsoft 365 Copilot is the paid add-on license that reads your actual work data and answers grounded in your files. That capability is exactly why it needs governance the free version doesn't.

Can Copilot see documents I've marked as confidential?

Copilot inherits whatever permissions the logged-in user already has. If a document was shared with 'Everyone except external users' two years ago and forgotten, Copilot will read it and can surface it in an answer to anyone with that same broad access, confidentiality label or not, unless the label carries an encryption or access restriction that Purview enforces. A label alone that just marks something 'Confidential' without an enforcement policy behind it does nothing to stop Copilot from summarizing it.

All 9 guides on confidentiality →

ShareX / TwitterLinkedIn
Was this useful?

Disclaimer · Educational content about software and productivity, not legal advice. AI tools and regulatory guidance change often, so check any tool against your own firm's obligations and your regulator's current guidance (for example the SRA in England & Wales, or your state bar and the ABA in the US) before using it with client data.

Free starter kit

Want the safe-tools shortlist as a PDF?

A client-confidentiality checklist for the SRA (UK) and ABA Rule 1.6 (US), with a shortlist of 10 AI tools and 12 ready-to-use prompts. Free, no spam.

Get the free Starter Kit →

Go deeper: The Lawyer's AI Toolkit (£29) →