LLMs in Contract Review: Promise, Limitation, and the Bar-Association Question

Large language models can now review a standard commercial contract faster than a junior associate and flag the same risk categories — indemnification scope, limitation of liability, intellectual property assignment, termination triggers — with comparable recall rates on well-structured agreements. The capability is real and the improvement over the past two years has been dramatic. What LLMs cannot do is be held professionally accountable for what they miss. That asymmetry is the entire legal AI market.

The bar-association question — what obligations does an attorney have when using AI tools in client work? — is being answered jurisdiction by jurisdiction, ethics opinion by ethics opinion. The ABA's formal opinion 512 in 2023 addressed generative AI use in law practice and affirmed that attorneys have competence obligations that include understanding the technology they use, supervising AI outputs, and maintaining confidentiality over client data. These obligations do not prohibit AI use — they define what responsible AI use looks like.

This creates a specific product requirement that LLM-powered legal tools must satisfy: explainability and auditability at the output level. An attorney using AI contract review in client work cannot simply rely on a model's flagging of a risk clause. She needs to understand why the clause is flagged, what the relevant precedent or statutory framework is, and how confident the system is in the assessment. These requirements mean that legal AI products built on raw LLM outputs — without a retrieval architecture that connects outputs to specific legal authority — cannot be responsibly deployed in attorney workflows, regardless of their accuracy rates.

The companies building in this space that we find most interesting are those that treat LLM capabilities as a reasoning layer on top of a retrieval-augmented legal knowledge architecture, not as a standalone oracle. The reasoning is fast; the retrieval grounds it in verifiable legal authority; the output is an attorney-reviewable analysis with citations. This is not just a technical design choice — it is a professional responsibility compliance choice that determines whether law firms can actually use the product.