Legal AI
· 6 min read

Copilot for Legal, Claude for Legal: where do documents go?

By Amine Gaaliche · drafted with Claude, reviewed and edited by hand

It depends on settings most lawyers never see. Inside Microsoft Copilot, the terms that cover a document change with the AI model in use, and an administrator decides which models are on. Each connector adds another service with its own terms. A firm that removes identifiers before a document reaches the assistant can give clients one answer, whatever the settings.

The legal products now live in tools you already use

Each launch brings AI to where lawyers already work. Copilot for Legal works in Word, where it can redline a clause as a tracked change. Claude for Legal and Gemini Enterprise for Legal connect to document systems and research services. OpenAI's Astra for Law is so far open only to selected large firms.

For a firm, this looks like one less vendor to approve. Many firms already have a Microsoft contract and keep their documents in SharePoint. So when a client asks where its documents go, "we use Copilot" sounds like a complete answer.

One product, several answers

Microsoft's own documentation shows why it is not. Copilot can run on Anthropic's Claude models as well as on Microsoft's default models. For the standard Claude models, Anthropic works as Microsoft's subprocessor, under Microsoft's product terms and data protection addendum.

Two details change the answer. First, these models are outside the EU Data Boundary, Microsoft's commitment to process EU customers' data in the EU. That is why they are off by default in the EU, EFTA and UK, and on by default elsewhere. Second, some advanced Anthropic models come "with Data Retention". For those, Anthropic acts under its own terms, "not subject to your Microsoft Customer Agreement". It keeps most inputs and outputs for up to 30 days, and flagged content for up to two years.

Microsoft documents this clearly, and its defaults are cautious. The retention models are off everywhere until an administrator opts in, and users can see in the product when a model keeps data. The point is where the answer now lives. Two lawyers in the same firm, in the same Word window, can be covered by different terms. The difference sits in an admin setting the lawyer never opened.

Every connector is another service

Connectors are what make these products useful. When the assistant calls one, it sends that service a request, and the request can carry text from the matter. Each service handles it under its own terms.

Some connectors read from places your documents already are, such as iManage or NetDocuments. Others are separate research, drafting or e-discovery services. Copilot for Legal lists more than two dozen partners, and Claude for Legal 20 platforms. So a client's question used to have one vendor's answer. It can now have as many answers as the connectors in use on that matter.

One answer that does not depend on settings

You can make the question simpler by changing what goes in. A model reviewing a supply agreement needs the clauses, not the names of the buyer's directors. Remove names, addresses and account numbers before the document reaches any assistant. Then the answer to the client is the same whichever model or connector runs: those details stayed on our machines.

With promptShield, identifiers become codes such as [P0ABCD] on your own computer. It reads Word, Excel and PowerPoint files as well as PDFs, and can save the result in the original format. Within a folder, the same name keeps the same code, so the model can still follow who is who.

Decode puts the real values back on your machine. Detection and redaction run locally by default. Only the optional Deep Analysis feature can be set to send page text to a remote model. How the codes work is explained separately.

What this does not cover

Removing identifiers does not make a document non-confidential. ABA Model Rule 1.6 covers everything about the representation, and deal terms or strategy can point to a client without a name. So still check which models and connectors your tenant allows, keep terms that exclude training, and write the rule down.

Whether a specialist platform is worth its price on top is a separate question, covered in why pay for Harvey if it runs on Claude and GPT.

Frequently asked questions

What is Copilot for Legal?

Microsoft's legal offering, launched on 15 October 2026. It adds more than 20 legal skills to Copilot in Word, such as clause review and drafting from approved precedents, with plugins from legal tech partners including Harvey and Legora.

Does Microsoft Copilot send documents to Anthropic?

Only when a Claude model handles the request. For standard Claude models, Anthropic acts as Microsoft's subprocessor under Microsoft's terms, outside the EU Data Boundary. They are off by default in the EU, EFTA and UK.

Is Claude for Legal safe for client documents?

On a commercial account, Anthropic says it does not train on inputs by default. Each connector you enable is a separate service with its own terms, so check those too, and remove identifiers first.

Do legal AI connectors share data with third parties?

A connector sends requests to its own service, and a request can include matter text. Document-management connectors read where your files already are. Research and drafting connectors are separate providers.

Found this useful? Let us know.

Share