Why I Founded LegalAgent, a Law Firm for the Generative AI Era

Hello, I'm Noriaki Asato, Representative Attorney at LegalAgent.

LegalAgent is a startup that is rethinking how corporate legal work is done through collaboration between generative AI and attorneys. We do not merely develop and provide AI tools; we run an "AI-BPO (BPaaS)" business in which our attorneys themselves use the agents we have developed in-house and deliver the legal work itself, such as reviewing contracts and drafting proposed revisions, to our clients.

Specifically, we have set an example fee of JPY 10,000 (excluding tax) per document for the review of an NDA (non-disclosure agreement). Our service is based on responding within one business day as a rule, while we confirm fees and turnaround times individually according to the nature and volume of the work.

Legal Agent, a Law Firm for the Generative AI Era

In this article, I look back on the background to my launching a law firm that takes on the work itself, the trial and error involved in using generative AI, and our in-house development efforts.

Structural Problems I Faced in Practice

I previously worked on M&A and other corporate legal matters at Anderson Mori & Tomotsune, and then practiced at AZX Professional Group, a firm specializing in startup support. At both firms I learned a great deal about practice and gained valuable experience as an attorney.

At the same time, through my daily work I became strongly aware of the structural burdens that traditional legal practice carries. In many matters, a pyramid-style division of labor is used, in which junior attorneys and staff read through large volumes of documents to prepare first drafts and research memos, and senior attorneys then revise them. In my own experience, too, there were times in large M&A due diligence (DD) when teams of 10 to 15 people read through materials late into the night, day after day.

From the client's point of view as well, legal services tended to have problems in three respects: cost, speed and the stance of the advice. There was the burden of fees under hourly billing, the speed problem of keeping the business waiting because it took about a week to complete a review, and a conservative tendency to stop at pointing out abstract risks and to find it hard to go as far as concrete options for moving the business forward.

Client companies asked us not only to tell them which risks to avoid but also to show them options for moving the business forward. However, I kept thinking that the heaviness of work in which many people are involved in stages, and a stance of formalistic risk avoidance, were getting in the way of meeting that request.

Providing the Legal Work Itself, Not Just Tools

Around the time I was grappling with these issues, I had the opportunity to try GPT-4o. Its understanding of context and the flexibility of its output had improved compared with earlier models, and I strongly felt that it had the potential to change the way legal practice is conducted. The starting point for founding the firm was the thought that, if AI were used appropriately, a small number of attorneys might be able to provide fast legal support at reasonable prices without relying on sequential review by many people.

In recent years, AI tools for contract review have become widely available. While the significance of improving work efficiency across the industry is great, there are aspects that are hard to resolve by providing software alone. That is the fact that the final review and judgment must be carried out on the user's side. Even if the AI extracts the issues, judgments such as "how exactly to revise the clause," "how to proceed with negotiations with the counterparty" and "whether this is acceptable as company policy" are made by the in-house legal team or the attorneys engaged. Even if there is a function that helps revise clauses, someone is still needed to review it.

From a business perspective, too, the breadth of the market a company faces differs between merely selling software and taking on the legal work itself. We chose a business model in which, rather than stopping at providing tools, we use AI to take on legal practice itself and are paid for the finished deliverables.

This is the model known as BPaaS (Business Process as a Service). We do not merely provide software as a tool; we carry out the business process itself. Attorneys carefully review the drafts prepared by AI and, after ensuring legal and practical consistency, deliver them as finished products. The Word add-in tool we developed is already being provided to outside law firms and corporate legal departments as well, but our main focus is on our attorneys using this system to provide services for which they take responsibility.

From the client's perspective, the process is simple. Clients ask us to review a contract through the chat tools they use every day, such as Slack or Teams, and share the background circumstances and their wishes. Behind the scenes, AI and attorneys work together, and we return a Word file reflecting tracked changes and comments. We value an experience in which clients can make requests following their usual procedures and receive finished products, without being forced to learn new tools or write prompts.

Trying Full Automation and Switching to an Interactive Editor

That said, we did not arrive at our current way of working from the outset. When we first started, I thought the entire workflow could be fully automated with AI, and I tried designing it using the no-code tool Dify. For documents such as non-disclosure agreements, we built up detailed rules such as "if a certain clause exists, go to branch A; if not, go to branch B."

As a result, we ended up creating nearly 200 large-scale workflows with layer upon layer of branches and decision conditions. However, once we actually started operating them, maintaining them became extremely difficult. This was because every time a law was amended, an interpretive standard changed, or exceptional circumstances arose in an individual matter, the enormous number of workflows had to be revised by hand. Legal practice always involves exceptions, so the very design of trying to handle every situation with branching programs prepared in advance was unworkable.

So we changed our policy of aiming for full automation and switched to a "Human-in-the-loop" approach, in which a human gradually raises the quality of the work while interacting with the AI.

In software development, editors such as Cursor are popular, in which you give instructions in a chat while looking at the code and make revisions while checking the diffs. I thought that if this interactive feel could be reproduced in Word, legal work could also proceed smoothly, and so we developed in-house a dedicated add-in that runs inside Microsoft Word.

Screen of the Word add-in

In this system, an AI chat panel sits permanently beside the Word screen. When an attorney gives an instruction such as "please revise the definition of confidential information," the AI picks up the context, rewrites the clause directly, and leaves tracked changes and comments with minimal differences. The attorney verifies the draft prepared by the AI against the source materials and the law and finishes it into a form suitable for practice. As a result, attorneys are no longer tied up with drafting from scratch and can concentrate on examining the content and making judgments.

Transforming the Organizational Structure and Our Future Vision

As operations premised on generative AI take hold, the organizational structure of law firms also changes. It is a shift from the traditional pyramid, in which "juniors do the groundwork and veterans revise it," to an inverted pyramid in which AI does the preparation and experienced attorneys review the content and decide the direction.

With AI supporting preparatory work such as research and first drafts, attorneys can focus on the role of verifying the validity of what is presented and making judgments consistent with business policy. Of course, accurate verification by humans is a prerequisite, rather than accepting the AI's output uncritically.

I see this system as a major opportunity for attorneys with strengths in particular fields and for small law firms as well. I think that, by putting in place tools and operational workflows that support practice, the range of matters that can be handled by a small team will expand in some cases, even for matters that used to require many people. With an eye to collaborating with specialist attorneys and expanding our organization in the future, we are building the foundations of a law firm suited to the generative AI era.

The Role of Professionals and Building Our Team

The business model we practice has the potential not only to revise costs but also to change the very way legal services are delivered. I think that the stance of professionals taking responsibility for the work while making use of the power of AI, rather than simply handing over a tool and leaving it there, will spread not only in legal services but also to other professional fields such as tax and consulting.

At LegalAgent, we are looking for people to build a new legal practice suited to the generative AI era with us. For the status of openings in each role, including attorneys and engineers, please check our current careers page. We welcome conversations, such as casual interviews, with anyone interested in what we are doing.

Legal Agent, a Law Firm for the Generative AI Era