Since joining Lumia, I've had the chance to speak with a lot of law firms from mid-size regional practices to large global partnerships. Almost every conversation starts the same way: an InfoSec leader tells us the firm is early in its AI journey, but that the adoption of AI across the firm is far outpacing their ability to govern its use. On top of that, there seems to be a new Legal AI tool hitting the market every week, and as difficult as the governance challenge looks today, it is only getting worse. Combine all that with the stringent requirements of protecting Attorney-Client Privilege in the age of AI, and it's easy to see why law firms are having an especially difficult time addressing this challenge.
The Legal Context Is Unique
Most enterprise AI governance conversations center on data security and regulatory compliance. These matter for law firms too, but law firms carry an additional layer of obligation that doesn't exist in most industries: attorney-client privilege. Privilege isn't just a policy preference; it's a legal doctrine. If a privileged communication is exposed, even inadvertently, the consequences can extend to the client relationship, to litigation outcomes, and to bar compliance. That changes the risk calculus entirely. Layered on top of that, roughly 78% of attorneys are now using AI in some form. Many of those tools are unsanctioned, not because attorneys are trying to create risk, but because they operate in a deadline-driven environment where the power of AI to accelerate workflows is simply too transformative to overlook. The result is a shadow AI problem that's larger and harder to contain in legal services than almost anywhere else.
On top of that, there are client mandates. Certain clients - typically those in the heavily regulated industry - are explicitly restricting whether and how their outside counsel can use AI on their matters. Firms need to honor those restrictions, but enforcing them manually across a large attorney population becomes incredibly difficult to manage at scale.
What a Foundational AI Policy Should Actually Cover
Based on Lumia’s experience advising leading law firms on their AI governance strategy, we’ve identified the following six pillars of an AI usage policy designed for this environment:
1. Access control. Not all AI tools are created equal. Some, like DeepSeek, carry significant data security and geopolitical risk. Others may be legitimate consumer tools that nonetheless lack the enterprise data handling agreements that protect client information. Others may be licensed tools for a specific purpose, but untrained to support legal research. Firms need the ability to block high-risk tools outright, enforce attorneys using only firm-sanctioned enterprise accounts (not personal accounts), and directing case work to approved legal AI tools.
2. Data protection at the prompt level. This is where most firms underestimate the risk. Prompts often can contain privileged communications, client matter numbers, or HIPAA-protected ePHI. The firm needs a way to inspect and redact sensitive content before it leaves the endpoint, and ensure that policy enforcement layer exists across every application and account. Even when attorneys are using enterprise licensed tools, firms should maintain clear visibility into where this information is exposed, and for highly protected data like ePHI, it should be redacted before reaching any external AI endpoint.
3. File Inspection. Given the nature of legal work, attorney AI usage inevitably involves sharing files and documents for AI-enabled legal review. An AI usage policy must involve inspection of these files, audit trails of where they’ve been shared, and an ability to block exposure of a document containing protected information before it leaves the firm’s environment. For firms already utilizing a Document Classification system like Purview, their AI policy needs to cover how attorneys should handle those classifiers in an AI context.
4. Privilege protection. This one is critical for law firms to consider and a primary focus of ABA Opinion 512 regarding use of AI in legal services. Law firms using AI must protect client confidentiality, which means any AI usage policy implemented by a law firm has to distinguish between privileged and non-privileged information; something traditional DLP was never designed to achieve. The trick here is topic classifiers trained to recognize privileged content within prompts, then enforcing specific controls (redact, block, warn) on interactions containing privileged information.
5. Hallucination risk mitigation. Attorneys citing hallucinated case law is no longer a hypothetical. It is happening on a regular basis with real consequences to the attorney and firm they represent. According to a Thomson Reuters Westlaw study examining legal cases from June 30 through August 1, 2025 researchers uncovered 22 distinct instances where adversaries or the judiciary identified fabricated citations within legal submissions, frequently culminating in disciplinary motions or formal judicial sanctions. The AI usage policy must address this risk, and actively mitigate it by ensuring attorneys direct legal research to approved legal AI tools; and provide inline warnings of the need to manually verify any citations for accuracy.
6. Agentic activity governance. AI usage is quickly moving beyond attorneys asking questions in a chat interface. Increasingly, AI tools can take actions on behalf of users; retrieving documents, accessing matter data, and interacting with internal systems. The policy must define which agentic actions are permitted, which require human approval, which systems agents may access, and which actions should be blocked outright. Firms need visibility into downstream activities triggered by prompts, including what data the agent accessed, what tools it used, what actions it attempted, and whether those actions were appropriate for the matter, client, and user context.
7. Client-mandate enforcement. When a client places restrictions on the use of AI for their matters, it becomes another AI policy that must be enforced. A governance platform that knows which clients have restrictions and can trigger warnings or block when those clients are mentioned in AI interactions is the only scalable way to honor those commitments.
Governance Shouldn't Slow Attorneys Down
Law firms of all sizes consistently express concern that AI governance will create friction that undermines the productivity attorneys are trying to capture. It's a fair concern, and one we take seriously at Lumia, which was founded on the idea of enabling rapid AI adoption, not slowing it down.
The goal isn't to restrict AI use, it's to make AI use safe and auditable. Lumia deploys at the network level, which means there's nothing for attorneys to install and no change to their workflow. Policies are configured in natural language, so the firm's general counsel or IT team can define what "privileged content" means in their context without writing rules in RegEx, and enforcement is precise; creating as little friction as possible in attorney workflows.
Lumia gives law firms a clear picture of which AI tools attorneys are using, what's being sent, and whether the firm's policies are being followed across every protocol, every device, and every application.
Starting Point
For firms that are early in the AI adoption process, the good news is that a foundational policy doesn't need to be comprehensive on day one. The important thing is to start with the controls that carry the highest risk: blocking unauthorized tools, enforcing enterprise account usage, and protecting against obvious privilege exposure. From there, the policy can evolve as the firm learns more about how its attorneys are actually using AI.
Every firm we talk to is somewhere on this journey, and the most successful ones are those taking action early to create a governance foundation that protects the firm while enabling a rapid AI rollout.
