AI Gateway for Legal Firms: Keeping Privileged Data Inside Policy

A lawyer drafting a brief pastes a section of a privileged memo into a model to tighten the argument, and a paralegal summarizing discovery drops a batch of client documents into a chatbot to speed the review. Both are ordinary uses of AI in a firm, and both send confidential client material to a model endpoint with, at most, a log line showing a call to a model host. The firm carries a duty of confidentiality under ABA Model Rule 1.6 and its state equivalents, and it usually cannot say which client's data left, under whose login, or where it went.

I want to walk through how an AI gateway puts a firm's model traffic under matter-aware policy and produces the record a firm needs to show privileged data stayed inside its confidentiality obligations, because the duty does not pause when the tool is AI.

The confidentiality duty meets a tool that keeps no record

A firm's obligation is to protect client confidences and to supervise how they are handled. That obligation assumes the firm can see where client data goes. AI tools break the assumption quietly: the data leaves through an HTTPS call that looks like any other, the model provider may retain or train on inputs depending on the terms, and the firm's own systems record only that a call happened. When a client, a court, or the firm's own general counsel asks whether privileged material was exposed to an outside model, the honest answer is often that nobody kept the record that would settle it.

The problem compounds with matters that carry ethical walls. A firm screens certain lawyers from certain matters, and that screen has to hold across every system the firm runs. If AI access has no notion of matter or identity, a screened user can send a walled matter's data to a model and leave no trace, which the firm has no way to detect. My view is that AI use inside a firm should be held to the same confidentiality supervision as any other channel client data flows through, and today it usually is not.

Matter-aware policy at the request boundary

An AI gateway sits inline on model calls and reads them as actions with an identity and a data context. It resolves which lawyer or staff member is calling, classifies the material in the prompt, and applies policy that can account for matter and role, not just a blanket allow. A call carrying a walled matter's documents from a screened user is blocked. A call carrying privileged material to a provider whose terms the firm has not cleared for confidential data is redacted or denied. The firm sets the rules once, and the gateway applies them on every call rather than trusting each user to remember them.

That turns confidentiality supervision for AI into three behaviors. Identity binds every model call to a named person rather than a shared firm key. Classification recognizes privileged and client-confidential material in the prompt. Enforcement blocks or redacts calls that would send protected material outside policy. The same identity-aware AI gateway machinery that serves other regulated settings is what makes matter and role part of the decision here.

The record a firm can show

When the question is whether privileged data was handled correctly, a per-decision record answers it. A record naming, for each call, the person, the matter or data class, the model and destination, the policy applied, and the outcome lets a firm show that client data stayed inside its obligations, or identify precisely what did not and respond. That evidence supports the firm's confidentiality supervision the way access logs support it everywhere else, and it aligns with the broader discipline of what to log for AI compliance. Produced at the request boundary rather than by the application, the record documents AI access independently of the tool that made the call.

DeepInspect

This is the gap DeepInspect closes for a firm. DeepInspect runs as a stateless proxy between your lawyers, staff, applications, and any LLM. It binds a named identity to each model call, classifies privileged and client-confidential material in the prompt, applies per-role and matter-aware policy with detection and redaction of sensitive data, and writes a tamper-evident per-decision record of identity, data class, policy version, destination, and outcome.

Those records let a firm show client data stayed inside its confidentiality obligations and support the supervision the duty requires, produced independently of the applications lawyers use. DeepInspect does not give legal advice or set your ethics policy; it enforces the confidentiality rules you define on your AI traffic and records that they held. To bring AI use under your confidentiality obligations, book a vertical briefing for legal.