The hard part is not the model.
It is who may see what.

There is a specific one. On 29 July 2024 the ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512, its first formal guidance on generative AI, addressing competence, confidentiality, communication with clients, candor to the tribunal, supervision and fees. It does not prohibit these tools; it sets out the conditions for using them properly, including that a lawyer must think carefully before putting client information into a tool and that informed consent can be part of the analysis. State bars have issued their own guidance too, so the applicable rules depend on your jurisdiction. The practical consequence for a build is that architecture does most of the ethical work: a system inside your own environment, inheriting your permissions, with an audit trail, answers most of these questions structurally rather than by policy.

By not letting it write authorities at all. Citations come from retrieval over sources the system can point at — your documents, your subscribed materials, published sources — rather than from the model’s memory, and anything it cannot support is marked as an open question instead of being written over smoothly. This is not a theoretical concern: courts have sanctioned lawyers for filing briefs containing authorities that a tool fabricated, and that specific failure is the reason the design refuses to produce a confident-sounding paragraph it cannot substantiate. A lawyer still verifies. The system’s job is to make verification quick rather than to make it feel unnecessary.

They have to be enforced by the same system that enforces them today, which in most firms means the document management platform. This is the single most common way legal AI projects go wrong: someone builds a retrieval index over everything the firm holds, and it quietly becomes a route around permissions the firm spent years configuring. We build retrieval that inherits matter-level permissions at query time, so a lawyer who is not on a matter gets nothing — not a summary, not a redacted answer, not even confirmation that responsive documents exist. Attempts are logged, because a refusal is as much a part of the record as an answer.

Almost certainly not by virtue of being a law firm, and it is worth saying plainly because these labels get attached to legal pages indiscriminately. The regime that governs your data is the Rules of Professional Conduct, your clients’ outside counsel guidelines, and whatever contractual security obligations you have signed up to — which are frequently stricter and more specific than the certifications vendors advertise. Where a firm genuinely handles protected health information for a client matter, HIPAA may apply to that engagement, and that is a real conversation with a real answer rather than a badge on a web page. We would rather scope the actual requirement than decorate the page.

Not for client material, and that is the whole reason this page exists. Consumer and general-purpose tools are a poor fit for confidential matter content because of where the data goes and what may be retained. What works is an enterprise model instance running inside your own Azure or AWS environment, where inputs are not used for training, logging is under your control, and the data never leaves the boundary your firm is responsible for. The models themselves are largely the same; the difference is entirely in where they run and what surrounds them.

Regulatory and filing monitoring, almost always. It produces something genuinely useful within weeks, it engages no client confidentiality because the sources are public, and it lets your risk people watch the system behave before anything touches a matter. From there the natural second step is retrieval over the firm’s own documents, which is more valuable and requires the permission work to be done properly. Starting with drafting is the common instinct and usually the wrong order, because it is the highest-scrutiny use with the least trust established.

It depends on which environment you run, which systems have to be connected and how much of the permission model is already clean, so there is no list price. The first step is a short assessment: where confidential material actually lives today including the copies nobody talks about, what your security team will and will not approve, and which use case gives you something useful fastest. That produces a scope, an acceptance definition and a fixed price in writing before any build starts.