Zero Trust for AI: Keep Copilot and Claude Inside Your Matter Walls

Zero Trust for AI means every person and every AI assistant gets only the access it needs, checked every time, with protection attached to the data itself. For a law firm, that keeps Microsoft 365 Copilot and Claude from surfacing a client file to someone who has no business seeing it.

Cocha Technology Inc. is a Houston-based AI governance and security advisory firm serving clients in the United States, Canada and the European Union. We help law firms and other regulated organizations deploy Claude and Microsoft 365 Copilot without exposing client data. This page explains how we think about it.

People used to call this the fortress mindset. The picture still holds if you move the walls. They belong around each matter and each sensitive file, rather than around the office.

Why AI changes the security question

Copilot answers from content the signed-in user can already open. It doesn’t grant new access. What it does is make old access easy to find.

That’s the problem. A folder shared with “everyone” in 2019 was mostly harmless because nobody went looking. Now anyone can type a question and get a summary of it. The same holds for any AI assistant you connect to firm data. It works with the access you give it, so sloppy access becomes fast, searchable exposure.

So the work starts before the license. Fix who can see what, then turn the assistant on.

Identity and least privilege come first

Most firms already have a firewall and antivirus. Good. Neither one stops a person signed in with valid credentials from opening a file they shouldn’t, and neither stops an AI assistant acting on that person’s behalf.

Identity is where control lives now:

  • Strong sign-in. Multi-factor authentication on every account, with phishing-resistant methods where you can use them.
  • Conditional access. Sign-in rules that check the device and the sign-in risk before granting access to client data.
  • Clean offboarding. Former staff and contract attorneys lose access the day they leave, not months later.
  • Least privilege at the matter level. People see the matters they work on. Everyone else doesn’t.

Ethical walls are least privilege with professional stakes. If a screened lawyer can’t open a matter in your document system, Copilot can’t summarize it for them either. If the wall exists only as a memo, it isn’t a wall. ABA Formal Opinion 512 says a lawyer’s confidentiality duties apply to what goes into generative AI tools, including disclosure to others inside the same firm.

A passed audit and a working control are different things

An audit tells you where you stood on the day of the audit. Security is what holds the rest of the year.

AI makes the gap wider. An AI acceptable use policy that says “don’t paste client data into public chatbots” is necessary. It won’t stop anyone on its own. Controls do that:

  • Sensitivity labels that mark and, where needed, encrypt privileged and client-confidential content.
  • Microsoft Purview DLP rules that block or warn when labeled content heads somewhere it shouldn’t, and can keep labeled files out of Copilot’s answers.
  • Monitoring for drift, because permissions and sharing settings slide back over time.

Compliance is what you get when the program works. It makes a poor target on its own.

Zero Trust for law firm AI, step by step

You don’t need an enterprise budget for this. You need an order of operations.

  1. Verify identity. MFA everywhere, conditional access, and devices that meet your encryption and update requirements before they reach client data.
  2. Clean up permissions and oversharing. Find broad sharing links, “everyone” groups, and inherited access on sites and folders that hold matter files. Restrict them. Rebuild ethical walls in the permission model itself.
  3. Protect the data. Apply sensitivity labels and Purview DLP so the protection travels with the file, wherever an AI tool picks it up.
  4. Find shadow AI. Discover which unapproved AI tools people are already using, then steer them to the sanctioned ones. See our guide to shadow AI.
  5. Write the rules down. An AI acceptable use policy people can follow. Our AI acceptable use policy template is a starting point.
  6. Watch and respond. Audit logging on AI activity, and a named owner with a plan for what happens when an assistant surfaces something it shouldn’t.

Our Copilot readiness checklist walks through steps 2 and 3 in more detail.

How Cocha works

Gabriella San Miguel, Co-Founder and Principal of Cocha Technology Inc.

We lead with AI governance and security advisory. Cocha implements AI TRiSM controls using Gartner’s AI TRiSM framework. In practice that means permission and oversharing cleanup, sensitivity labels and Purview DLP, shadow AI discovery, and AI acceptable use policy.

Cocha also offers Managed IT Services: helpdesk, monitoring and day-to-day support, available 24×7, with industry-standard response times. We support businesses remotely across Texas and onsite in Houston. It pairs with the AI work. A firm that has fixed its permissions still needs day-to-day support behind it.

Cocha is led by co-founders Steven R. Combs and Gabriella San Miguel, both Co-Founder and Principal. For more on law firm work, see AI governance for law firms.

Steven R. Combs

Where to start

Start with the free AI Readiness Snapshot. It is a 60-minute working session followed by written findings and a 30-minute review. It is read-only, makes no changes and installs no software. The findings point to your priorities across permissions, data protection, monitoring, shadow AI and Secure Score.

Frequently asked questions

Zero Trust means no user or AI assistant is trusted by default. Each request is checked, and access stays limited to what the work requires. For Copilot and Claude, that means cleaning up permissions and labeling sensitive content before the tools can reach firm data.

No. Copilot works from content the signed-in user already has permission to open. The risk is that it makes existing overshared content easy to find and summarize. That’s why permission and oversharing cleanup comes before rollout.

An ethical wall only holds against AI if it’s enforced in permissions. If a screened lawyer can’t open a matter’s files, an assistant acting for that lawyer can’t use them either. A wall that exists only as a memo gives AI nothing to respect.

No. An audit shows where you stood on one day. Security is the set of controls that keeps working afterward. For AI, a written acceptable use policy matters, and labels and DLP rules are what stop client data from leaking.

A managed service provider (MSP) keeps IT running day to day, from laptops to the helpdesk. A managed security service provider (MSSP) focuses on security monitoring and incident response. Cocha leads with AI governance and security advisory and offers Managed IT Services for day-to-day support. We don’t run a security operations center.

Start with the free AI Readiness Snapshot. It is a 60-minute working session followed by written findings and a 30-minute review, and it is read-only with no changes and no software. The findings point to your priorities across permissions, data protection, monitoring, shadow AI and Secure Score.

4321 Kingwood Dr, Suite 222

Kingwood TX 77339

281-607-0616

info@cochatechnology.com