At Harvey, we have no doubt that 2025 will continue to be a transformative year for AI. This transformation is propelled by the development of reasoning models, whose stronger reasoning capabilities enable a more complex class of AI agents. In 2024, we outlined some of our early impressions of reasoning models—such as planning and reflection—and how they would enable agents for knowledge work.

The first of these next-generation agents are now available in Harvey, making our platform more transparent, more meaningfully interactive, and more equipped than ever to solve the hardest knowledge work problems. With the help of Harvey agents, our clients can stay focused on their highest-impact work.

There are four key terms that collectively define AI systems at Harvey. The first are models. Individual models, such as GPT-4o or o1, are our AI building blocks—given a single prompt, they produce a single response.

When multiple models are combined with task-specific tools and knowledge sources—such as search tools, RAG databases, or other function calls —we describe that as a model system. Model systems tend to be rigid one-way trips, working linearly through preset model calls and informational handoffs to produce a result.

Agents present a more flexible and powerful approach to solving problems than model systems. We define agents as model systems that can plan, adapt, and meaningfully interact with humans to complete a task.

The first client-facing instantiation of agents in the Harvey platform is in workflows. Workflows consist of one or more agents that collectively produce a meaningful, specific work product. They define an overall goal, from reviewing a filing to drafting an MSJ, and use a set of interfaces to enable human and agent interactions to reach that goal. Agents are the means; workflows are the ends.

Today we introduce workflows in Assistant, designed to proactively guide users through real-world tasks and yield high-quality outputs. We’re starting with a set of workflows across transactional, litigation, financial services, and other high-impact use cases, and will continually add more to support the tasks our customers commonly perform.

Together, Assistant workflows make the Harvey platform dynamic, anticipate user needs (plan), refine their approach based on the inputs they receive (adapt), and provide structured outputs that can be iterated on (interact). By taking a proactive role in managing complex tasks, they help professionals navigate multi-step processes with efficiency and accuracy.

Assistant workflows introduce several key advantages that make interactions with Harvey more intuitive and effective:

As part of our commitment to expert quality, we are introducing custom evaluations to compare workflow outcomes directly against human lawyers. These evaluations aim to establish whether workflows produce human-quality work on common tasks typically requiring an hour or more of lawyer time. Our initial benchmarks cover three task types: structured drafting, unstructured drafting and analysis, and data extraction and structuring. On each of these, Harvey workflows perform at or above human lawyer-level and save substantial time on common, high-value tasks.

Overall, we find that on more objective tasks—such as building chronologies or drafting based on a well-defined precedent template—Harvey performs at human level, facilitating these tasks and leaving only review and finalization to the end user. On more unstructured tasks, such as analyzing transcripts, Harvey creates work products that outperform lawyer baselines. Lawyer evaluators routinely preferred Harvey outputs in blind reviews, citing greater depth, detail, and specificity, all of which set them up to more quickly and effectively deliver value to their clients.

Assistant workflows will be generally available in Harvey at the end of March. We hope that these workflows also serve to highlight key use cases Harvey can solve, both today and in the future. We will continue expanding workflows to support a broader range of professional tasks, integrating feedback from our customers, and refining agent capabilities. With workflows as the foundation, agents will also be embedded through Assistant, Vault, and the rest of the platform, helping professionals collaboratively tackle complex work with greater ease and precision.