Legal AI is one category name covering at least four different kinds of system. They are marketed in almost identical language, priced in comparable ranges, and demonstrated on similar material — which is why buying decisions so often turn on the quality of the demonstration rather than on the shape of the thing being demonstrated.

What follows is a way to read the architecture from the outside. It deliberately names no vendors: the distinctions are structural, they cut across the market, and a firm can place any candidate on this map in a single conversation.

1. The research platform with generative features

The core asset is a licensed corpus — case law, statutes, secondary materials — and the generative layer sits on top of it, answering questions against that corpus.

It is strong where the answer lives in the corpus and weak where the answer lives in your matter. The tell is what happens when you ask it about a document you supplied: whether that document becomes a first-class part of the answer or a temporary attachment to a question.

2. The task assistant

Built around a conversation or a prompt library, this class does drafting, summarisation and comparison well, and is usually the fastest thing in the category to deploy because it integrates with very little.

Its ceiling is structural rather than technical: it holds no state between exchanges. Nothing it produces is a step in a process that something else can pick up. The tell is what exists after the lawyer closes the window — in this class, nothing does.

3. The document-intelligence engine

Purpose-built for volume: extraction, classification, deviation-spotting across large sets. This class is genuinely differentiated, and for review and diligence it is often the strongest option available.

Its limit is scope rather than quality. It analyses corpora; it does not run matters. The tell is whether the output arrives as a report or as a change in the state of a file that someone now owns.

4. The workflow platform

The unit is the matter rather than the query. It models phases, owners, approval gates and a record; the model is invoked inside that structure rather than beside it.

It is the slowest to deploy, because it has to meet how a firm actually works. The tell is whether an approval is a state the matter is in, or a habit the team is asked to maintain.

The cross-cutting question

Deployment shape is a separate axis, and it is the one with professional-conduct consequences. Where inference runs determines whether a firm is making a disclosure at all — and the ethics guidance now puts that determination on the lawyer, not the vendor.

ABA Formal Opinion 512 requires a lawyer using generative AI to protect client confidentiality and to obtain informed client consent before disclosing confidential information to a tool.[1] That obligation is easy to satisfy when nothing leaves the firm, and demanding when the honest answer involves a third party, a retention window and a set of terms that can change.

The Southern District of New York sharpened the point in February 2026, holding that material generated through a consumer-tier AI assistant enjoyed neither privilege nor work-product protection — in part because the tool's terms permitted disclosure and training use, and in part because the work was not done at counsel's direction. The court expressly left open whether an enterprise product with training exclusion and contractual confidentiality would fare differently, while cautioning that contract alone does not establish privilege.[2]

Reading a platform in one meeting

Five questions place any candidate on both axes, and the hesitations are as informative as the answers.

  • What is the unit the system works on — a query, a document set, or a matter?
  • What exists after the interaction ends, and who owns the next step?
  • Where does inference run, and what leaves the perimeter when it does?
  • Is the matter boundary enforced beneath the application, or filtered inside it?
  • Can the firm produce the full record of an interaction without the vendor's help?

None of the four architectures is wrong. A firm that needs volume review and buys a task assistant has made an error; so has a firm that needs a matter system and buys a research platform. The category name will not tell you which you are looking at. These questions will.