What to Decide Before Bringing AI Agents into Legal Work

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

AI agents can be a powerful help in the groundwork for legal work. They can assist with reading contracts, drafting proposed revisions, researching related information, and drafting emails and documents.

However, when it comes to using them in practice, simply introducing the tool does not make things run smoothly. Unless you decide in advance which tasks to entrust to them, which information to allow them to access and where attorneys or legal staff will check their work, responsibility may become unclear.

Dividing Up the Tasks to Entrust

Legal work covers a wide range of tasks. Contract review, first-pass checks of non-disclosure agreements, searching internal rules, referring to past decisions, drafting explanations of revisions for business partners, taking in day-to-day legal consultations, and checking board materials and M&A data rooms all carry different weights of judgment.

When introducing AI agents, the first step is to divide up the target tasks concretely. For example, checking standard clauses in non-disclosure agreements and summarizing contracts are typical tasks for which AI makes it easy to move the groundwork forward. That said, even for a non-disclosure agreement, a person verifies the accuracy of the output and assesses it according to the nature of the transaction and the importance of the information. On the other hand, judgments on core business contracts, capital policy, representations and warranties in M&A and the like should not be completed on the basis of AI output alone.

In this way, divide up the tasks and decide which groundwork the AI handles and which final judgments humans make. There is no need to automate every step from the start.

The Scope and Management of Information the Agent Can Reference

If an AI agent can reference the necessary related materials, it becomes easier to convey the premises for a judgment. If it can reference past contracts, review standards, past negotiation histories and the like, it becomes easier to obtain output that follows your company's policy.

On the other hand, carelessly widening the scope of reference creates information management problems. There is a risk of mistakenly quoting confidential information from another matter, a risk of confidential data being displayed to employees without viewing rights, and a risk of personal or confidential information accumulating in conversation logs.

For this reason, the data disclosed to the agent is managed separately by department, matter and access rights. Legal judgments depend on context, so even the same wording can lead to different conclusions depending on the balance of power with the business partner and the course of negotiations. Simply having it read a large volume of materials does not lead to appropriate judgments.

Designing the Points at Which Humans Check

The key to building it into practice is at what stage a person checks. Decide who verifies revised clauses generated by the AI, what level of checking is inserted before anything is sent outside the company, and how high-risk matters are detected and routed to specialists.

In legal practice, prose that reads fluently and prose that is legally sound are two different things. Because AI output tends to be naturally worded, there is a tendency for deficiencies and oversights to be hard to notice. Governing law, limitation of liability, ownership of intellectual property rights and similar matters in particular require specialist review that takes individual circumstances into account. AI output is used as an aid to judgment, not as a final decision.

How to Keep Conversation Logs and Audit Trails

When running AI agents in legal work, keeping records of usage is also important. If you can trace who gave what instruction when, what answer was obtained and how the final decision was made, it becomes easier to explain matters inside and outside the company from an internal control perspective as well. At the same time, decide on viewing rights and retention periods for the confidential and personal information contained in the logs.

In contract review and legal consultations, there are cases in which the appropriateness of a judgment is questioned later. Recording how humans examined the AI's proposals, and which revisions were adopted and which were rejected, makes it easier to fulfill the legal function's accountability.

It is necessary to review usage regularly, identify the tasks for which it is used heavily and the areas where incorrect answers tend to occur, and continuously revise the operating procedures.

An Opportunity to Rethink How Legal Work Is Organized

Introducing AI agents goes beyond adding one more tool. It is a good opportunity to rethink the flow of legal work itself, from taking in consultations to approvals and the division of roles with outside attorneys.

In many workplaces, consultations are concentrated in the legal department, and responses tend to rely on individual experience and searches of past emails. Decide which groundwork to entrust to the AI and which tasks a person checks, and put in order the procedure from consultation to answer. It also becomes an opportunity to rethink work that relied on individual memory and searches of past emails.

At LegalAgent, we use AI agents as tools that allow attorneys and legal staff to devote time to individual judgments. The final legal judgment is made by a person.