Harvey's new AI agents are outperforming lawyers

Prompting an AI to do legal work usually feels like micromanaging a nervous first-year associate: if your instructions aren't completely airtight, you get junk back.

Legal tech unicorn Harvey is flipping that dynamic. On March 11, the company introduced its next-generation AI agents, packaging them directly into a new feature called Assistant Workflows.

Scheduled for general availability at the end of March, the rollout marks Harvey’s biggest push yet to turn raw reasoning models into purpose-built legal assistants.

Flipping the prompting burden

Harvey breaks down its tech into a clear hierarchy. Individual models like GPT-4o or o1 are just basic building blocks. Combine them with search and internal databases, and you get a linear model system that basically runs on rails.

Agents are supposed to be different. Harvey defines them by three specific traits: planning out multi-step tasks, adapting their strategy when intermediate results shift, and interacting with humans along the way.

In practice, this means lawyers no longer need to craft elaborate prompts. When initiating a workflow, the agent actively questions the user to pull out necessary case context before generating drafts.

Harvey also added visible thinking states. Instead of staring at a blank loading screen, legal teams can watch the agent’s intermediate reasoning steps, model calls, and citation logic unfold in real time.

Beating lawyers at transcript analysis

The initial rollout covers core transactional work, litigation, and financial services. To back up the marketing claims, Harvey ran blind benchmarks comparing its workflows against human lawyers on tasks that typically take an hour or more.

For structured tasks, like drafting a 1.01 disclosure or extracting key dates to build a chronological timeline, Harvey’s outputs were rated equal to human lawyers.

Things got more interesting with unstructured analysis. When tasked with analyzing court transcripts, human evaluators actually preferred Harvey’s work over attorney baselines.

The lawyer evaluators noted that the agent produced outputs with significantly more depth, specific detail, and immediate utility for client work.

Why it matters

For legal professionals, this fundamentally changes the user experience. You no longer have to spend twenty minutes engineering the perfect prompt; the system acts like a proactive colleague that conducts the intake interview itself. If you handle heavy discovery or routine transactional filings, it is definitely worth testing when it rolls out at the end of the month.

Across the industry, this is a clear sign that vertical AI has outgrown generic RAG wrappers. In high-stakes fields like corporate law, single-turn prompts are simply too fragile. Harvey is showing that the real moat lies in orchestrating multi-step agentic workflows that can pause, rethink, and ask for clarification.

We think Harvey’s most clever move is exposing the agent's internal thinking states. In an industry obsessed with liability and billable hours, showing the work builds trust far better than a black-box answer.

That said, high praise in blind evaluations carries risks. If junior-level drafting becomes this seamless, the temptation for senior attorneys to skim rather than rigorously verify could open up brand-new liability traps.