Knowledge Management at a Law Firm with Google Drive and Codex

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

Our firm, LegalAgent, provides corporate legal services as a law firm for the age of generative AI. As I have mentioned in my writing about the AI Native Law Firm, whether a law firm has organized its data in a form that generative AI can easily handle has a major influence on whether AI can be put to practical use in day-to-day work.

A firm holds a wide variety of information: contracts, revised drafts from counterparties, exchanges on Slack, practice guides, guidelines and more. In actual practice, however, it is rare for all of this to be gathered neatly in one place.

Scattered Information and the Challenges of Using AI

In reality, information tends to be scattered: some documents are attached to emails, exchanges remain in Slack, the latest contract sits in Google Drive, and past reference files are saved in a separate folder. It is not unusual for templates to be kept by individuals, for case-specific judgment criteria to exist only in the head of the attorney in charge, and for books and guidelines to be saved as PDFs without being connected to the daily flow of matter work.

If people alone are doing the work, a certain amount of inefficiency can be absorbed through the memory and experience of the person in charge and by checking with colleagues.

When you try to build AI into legal practice, however, scattered information becomes a major obstacle. AI can refer only to information that exists in places it has access to. Background information in an attorney's head cannot be referred to unless it is put into words. Information remaining in Slack or past matter folders may also be overlooked unless the connections and the scope of search are set appropriately. That is why, in aiming to be an AI Native Law Firm, simply having a high-performance model is not enough.

What needs to be organized first is an environment for centralized management of information. Centralized management here does not simply mean saving files to Google Drive. It means organizing the information handled in legal practice, by units such as matter, client and contract type, in a format that AI can easily read. In my view, the key to knowledge management at a law firm is creating a state in which both people and AI can later refer to the same context.

In legal work, even for contracts with identical clauses, the conclusion can change significantly depending on the background of the transaction. For example, when reviewing a service agreement, the comments to be offered differ depending on whether the contract is for a proof of concept for a new business or the renewal of an existing transaction, on the balance of power with the counterparty and on whether there have been past disputes.

If the counterparty is a major customer and concluding the contract quickly is strongly desired for business reasons, negotiating on a narrowed set of key issues is more suitable than sending back many minor revisions. Conversely, if the company is extremely dependent on the counterparty's performance of the work and the impact of any damage would be significant, it is appropriate to carefully check the limitation of liability and information management clauses.

This kind of material for judgment cannot be obtained from the text of the contract alone. Appropriate analysis becomes possible only when surrounding information such as the background of the matter, the business, past policies and exchanges with the people in charge is available. With an operation like conventional AI chat, in which a person has to find and attach those materials every time and explain the background in writing, the preparatory workload is too great. This is where the design of the folder structure becomes important.

Access Control for Google Drive and the Design of Matter Folders

Our firm uses Google Drive to manage matter documents. Currently, when Codex handles documents on Drive, we use the Google Drive API through an MCP we built ourselves. We do not use a local mount via Google Drive for desktop as the route through which AI retrieves and updates documents.

In actual work, starting from the matter folder, the AI checks the levels of the hierarchy needed for the task in turn and retrieves the target files. When editing, it keeps the retrieved original as is, creates a working copy, and proceeds after the save location and scope of changes have been determined. We manage separately the permission for AI to search for candidate documents and the permission to modify the files it finds.

This mechanism allows us to search for candidates such as past contracts with the same counterparty or templates close to the contract type at hand. However, we do not adopt an operation in which the entire shared drive is scanned without limit. Searches are carried out in line with the scope of the matter and the permissions, and an attorney reviews whether the candidates presented are appropriate and up to date.

Even if the burden of manually selecting and attaching files is reduced, this does not mean that data is no longer sent externally. We decide what can be given to AI in light of our confidentiality obligations to clients. In addition to the service's terms of use, we also check opt-out settings for training use and data retention periods. Care is also needed on the point that having access permission does not necessarily mean that documents from other matters may be diverted to unrelated processing.

Consolidating Communications and Organizing Reference Materials

Our firm places great importance on the design of matter folders. Knowledge management at a law firm may bring to mind a dedicated database, but in practice all work moves matter by matter. Daily work, such as review requests from clients and consultations on fundraising, proceeds matter by matter. We distinguish documents received from files in progress, and make it possible to trace the counterparty's revised drafts and exchanges on Slack and email from the same matter. If the past cases used for reference are also recorded, the matter folder functions as the starting point for the work.

Keeping the explanation given at the time of the request and subsequent exchanges in the matter folder is especially effective. Background information such as "this clause is non-negotiable," "we want to build a long-term relationship with the counterparty" or "this time we want to prioritize the deadline" is not written in the text of the contract but is essential for deciding the review policy. If such information remains scattered across Slack and email, people other than the person in charge cannot check it later, and AI cannot grasp the context of the request. We therefore keep the full text of the exchanges, the dates and times they were sent and received, and information about attachments in the matter folder.

Our firm's practice is to save exchanges on Slack and email to the matter folder. By setting up the available Slack and Gmail integrations and a scheduled execution mechanism, we have built a flow that retrieves designated exchanges and stores them in the prescribed location. Storage in Google Drive uses the API through our own MCP. Simply enabling an integration does not mean that all messages are automatically saved unconditionally. We store the records needed for handling matters after designing in advance the target channels and emails, the scope of storage and the treatment of confidentiality.

When a review request arrives from a client on Slack, the message text and attachments are kept in the matter folder's records. The same applies when additional preconditions arrive by email. Organized this way, we can instruct Codex: "Please refer to this matter folder and organize the client's request and additional conditions." The burden of handover when the person in charge changes is also reduced, and AI can proceed with work reflecting the past context.

Equally important is organizing reference materials. A law firm has a large number of books, statutes and practice guides. Image-only PDFs and materials whose location is unknown can be difficult to use as is for search and citation. After confirming the rights holder's permission, the license agreement and the scope permitted by law, we carry out OCR processing and conversion to text, and store the materials classified by contract type and area of law. Note that the mere fact of having purchased a book does not permit every kind of reproduction, including sharing within the firm and use with AI. Our firm systematically stores and uses statutes, public guidelines, internal research memos and type-specific checklists.

For example, when reviewing a privacy policy, we can compare the provisions of the Act on the Protection of Personal Information, the guidelines of the Personal Information Protection Commission and internal drafting memos in the same working environment. For a fundraising matter, we can check the investment agreement, the shareholders' agreement and issue memos from past Series A matters in parallel. As a result, AI does not merely state generalities but can proceed with work while accurately comparing matter documents with reference materials.

Designing Skills That Carry Out Work Rules

To put knowledge to use in practice, the design of skill folders is also essential. Knowledge such as checklists is hard to put to use if it is simply stored in the form of memos. With a procedure in which a person opens and rereads the materials every time, applies them to the contract and drafts comments, knowledge is not naturally built into the work.

We therefore take the approach of packaging knowledge in a form that AI can execute. Our firm defines skills for each type of work, such as contract review, Word revisions and legal due diligence. A skill is not limited to a single prompt; it defines the work procedure, such as which materials to read in which order, which checklist to apply and in what format to write comments. It also includes how to handle tracked changes in Word, the folders to refer to and the folders that must not be touched. Instructions in a skill never override access permissions defined in advance.

In contract review, the points of focus differ depending on whether the client is the disclosing or receiving party, or the outsourcing or service-providing party. By setting up skills for each position, Codex can carry out the work in line with the checklist and comment policy suited to, for example, the service provider's perspective, without drifting into generalities.

Knowledge Management as Work Design, and Legal Outsourcing

I think that at law firms going forward, knowledge management will become work design itself. Until now, knowledge management has been largely a repository where people later search for templates and research memos.

In an environment that combines generative AI and Codex, however, knowledge management becomes the foundation for AI to carry out work. If matter folders are in order, AI can easily refer to the documents of the case; if exchanges are recorded, it understands the context of the request; and if templates are available, it can find the appropriate source file. If internal judgment criteria and playbooks exist, it can reflect each company's standards, and if statutes and guidelines are organized, it can identify issues on a reliable basis.

The more thoroughly a firm has organized its information, the easier it is to build AI into its practice. Conversely, in an environment where information is scattered, it is difficult to draw out AI's capabilities no matter how much model performance improves. When AI cannot find the materials it should refer to and cannot grasp the latest version or past history, oversights and mistaken inferences based on fragmentary information are likely, and it becomes difficult to consistently maintain the quality that legal practice demands.

Building on API access to Google Drive and the design of matter folders, our firm makes records from Slack and email available for reference. We also systematize statutes and guidelines and maintain an environment in which they can be used in line with the procedures set out in our skills.

This effort also bears on the nature of legal outsourcing. Even if you ask outside attorneys for contract review or legal advice, if the company's business, past decisions and relationships with counterparties are not shared, the result tends to be a formal outside answer. In that case, the in-house legal team has to reread the comments and re-judge them in line with the company's risk tolerance.

On the other hand, if matter folders, templates, company-specific judgment criteria and skills are centralized, outside attorneys can make judgments informed by past context. With Codex organizing and presenting the underlying information, attorneys can support decision-making at a resolution close to that of in-house legal. Developing this kind of knowledge environment is also important for using a collaborative system of outside attorneys and AI as much as needed, when needed, before hiring a dedicated legal staff member.

At our firm, attorneys well versed in generative AI and corporate law work together to support contract review, fundraising, M&A and more. For our support combining knowledge management and AI operations, please see the following page.

• Legal Outsourcing

Frequently asked questions

What does centralized management mean in knowledge management at a law firm?

It means organizing information in a format that AI can easily read, in units such as matters, clients and contract types. Rather than simply saving files to Google Drive, it is necessary to create a state in which both humans and AI can later refer to the same context. Appropriate AI-assisted analysis is likely to become possible only once surrounding information such as the transaction background and history is in place.

How is access controlled when Codex handles materials on Google Drive?

We use the Google Drive API through an MCP we built ourselves and manage permissions separately. For safe operation, we separate the permission to have AI search for candidate materials from the permission to modify files. We do not use local mounts as a route, we avoid unrestricted scanning of the entire shared drive, and we search within the scope of each matter, with attorneys checking the content.

What is a "skill" for making use of AI?

It is a concrete work procedure defined for each type of work. It goes beyond a single prompt, specifying which materials to read in what order, which checklist to apply and in what format to write comments. By setting up skills according to the position taken, Codex is likely able to carry out work from the appropriate perspective.