Law firms are handing AI the whole workflow, not just the draft

The brief benchmark

Harvey introduced LAB: USCL to test frontier models' ability to research and draft litigation briefs end-to-end, drawing its pipeline from RECAP, a community-constructed archive of PACER filings, with a final set of 100 cases and models evaluated pairwise in a round-robin tournament using Grading Reward Models. A small group of frontier models can now research and write complete, first-draft briefs in a single shot without attorney guidance. That is a real capability milestone: the brief is no longer a collection of assisted paragraphs but a complete draft produced in one pass. m:…2687

The same benchmark sets the limit just as clearly. There is a 700 Elo point gap between the highest- and lowest-rated models, with Opus 5.5 the highest-rated, and even the strongest briefs produced by models still fall short of the filing bar and show a long tail of failures. The gap between first draft and filing-ready is where lawyers still earn their fees, and Harvey is partnering with Grubel AI to make research in this area more accessible, which suggests the benchmark's authors see the remaining distance as a research problem, not a deployment detail. m:…2687

Fixing the research underneath

Legora's answer to the same quality problem starts one layer down, in the research itself. The company says AI models often fail at legal research by producing fabricated citations, citing outdated law, or missing exceptions to rules, and it has introduced Ontological Legal Research, starting with the United States. Legora collects US slip opinions directly from court websites within 24 to 48 hours of release, and the company says it achieves 99.997% accuracy in its legal data by having three separate people check every word on every page. Its ontological data structure makes the Legora Agent 40% faster, according to the company, and it plans to release an AI-native citator later this year. m:…2494

The context Legora gives for the work is geographic: in Germany, only about 3% of case law is digital, and in the UK, legislation from official sources is often two to four weeks behind. A research tool is only as good as the corpus it stands on, and the corpus itself is uneven. That unevenness is part of why fabricated or outdated citations keep appearing: the model fills gaps the data leaves open. m:…2494

Diligence and intake go end to end

Ropes & Gray shows what deployment looks like when a firm treats the model as a workflow rather than a writing aid. The firm is developing a proprietary transactional diligence AI tool in partnership with OpenAI that automates the first pass of commercial contract diligence by grouping documents by theme and issue, and it refined the tool's rubric by comparing outputs with human-authored diligence reports from completed deals. The firm aims for the tool to produce detailed, issue-level diligence reports in hours rather than weeks, and it utilizes GPT-6 Astra in its transactional diligence workflows. Lawyers review all interim summaries and source documents before delivering work to clients, and the tool produces summaries and reports in the firm's signature format. m:…9017

On the plaintiff side, the constraint is capacity, not drafting. Plaintiff law firms are currently bottlenecked by labor constraints, forcing them to reject approximately 96% of labor and employment cases, according to a discussion on the a16z show with EVE chief executive Jay Madheswaran. A law firm in Los Angeles increased its capacity from 100 to 500 clients per attorney using the EVE platform, and EVE is developing AI agents capable of handling end-to-end legal workflows, including voice-based outbound calls to insurance and healthcare companies. The platform allows firms to respond to discovery requests immediately, which the discussion frames as neutralizing traditional defense delay tactics, and a single legal case can involve approximately 10,000 individual actions or touches throughout its lifecycle. EVE says it does not train its models on attorney-client privileged data, and that it is currently limited by engineering capacity rather than market demand. m:…8716

The pattern across all four threads is the same: the draft is solved enough to be useful, the workflow is where the value now sits, and the quality bar that matters, filing-ready work or client-ready diligence, is still enforced by human review. Firms that treat AI as a faster first pass with review built in, as Ropes & Gray describes, are deploying. Firms waiting for a model that clears the filing bar on its own are still waiting. m:…2687