Codex for Legal Work: Differences from Claude Code and Use in Word Practice

Hello, this is Legal Agent.

Depending on the matter, Legal Agent delivers most work within one business day, sometimes the same day. People often ask how a law firm manages that turnaround. I use many tools to work efficiently, and lately Codex has become the one that fits daily practice best. This article covers how we actually use it.

A general work tool, not only a coding assistant

Codex is usually associated with software engineering: writing code, fixing bugs, running tests. That is still the easiest use case to picture, but in practice Codex has become a core tool for law firm work too.

How it compares to Claude Code

Claude Code gets a lot of attention for business use as well, and it is excellent. Even so, Codex currently fits my own work better, for three reasons.

First, the interface is simpler. Claude's product has several entry points, Chat, Cowork, Code, with overlapping roles that an engineer can sort out but a non-engineer like me finds less obvious. With Codex, any task, reading a local folder, fixing a Word file, running research, can be asked from the same place, which matters for a law firm where most users are not engineers.

Second, GPT-5.5, the model behind Codex, is genuinely strong for Japanese legal drafting. Claude Opus 4.7 is excellent too, but for my use GPT-5.5 produces more finely grained analysis and more natural Japanese than earlier GPT models managed, and Fast mode keeps the pace close to Claude.

Third, usage limits in practice let me keep working longer stretches without interruption, which matters in text-heavy legal work.

Codex works inside the task, not just in a chat window

A generative AI workflow usually starts as a chat: paste a clause, ask a question, read the answer, then go back to Word to edit. That is useful but limited, since reading an answer does not finish the job; someone still has to decide what to change, how to record it in track changes, and how to separate a counterparty-facing comment from an internal note, while weighing the deal and the relationship with the other side. What makes Codex interesting is that it does not stop at answering: it reads a local folder, finds the right files itself, copies them, edits and compares documents, and checks the differences as it goes.

No more manually attaching files

With earlier tools, a lawyer had to choose which files the AI should read and upload them one at a time, which looks routine but is quietly heavy: a person first searches the folder, recalls the relevant past matter, and confirms the current version before AI ever sees it. Codex changes that. Told to find a similar past matter, check for a past contract with a counterparty, or copy a file as today's working version, Codex searches local folders and surfaces the material itself, removing much of that upfront searching across a firm's past contracts, templates, commented Word files and due diligence material.

A custom MCP to edit Word while it stays open

We built our own tool so Codex can instruct Word while the document stays open and edit it directly. Contract review means adding track changes, placing counterparty-facing comments, keeping separate internal notes, and finishing in a state ready to send out. Codex on its own struggled here: review needs the smallest possible edit, intact track changes, and clause numbers and table structure that do not drift, and ordinary Codex tended to disturb a Word file's structure. Our MCP lets Codex operate Word directly, handling body text, track changes and comments together, in the file the lawyer is actually viewing.

Once that was in place, the experience changed. Codex edits Word directly rather than only describing a fix, so the lawyer can watch and give follow-up instructions such as softening a comment that reads too strong, and edit several files at once. At that point Codex stops resembling a chat and starts resembling a working environment that reaches across accessible folders and handles Word's body text, track changes and comments together.

Automating comparison versions

Corporate legal work often needs a redline built with Word's compare feature, tedious since a person must select a folder and identify files by hand each time, manageable for one file but a chore for ten at once. We built another MCP so Codex decides which files to compare and produces the redline itself on request, which has saved real time.

Example: contract review

Under the old process, a lawyer saves the contract from email, searches for a similar past matter, checks the template folder, and reviews past comments and the counterparty relationship before opening Word to edit, comment and carefully preserve the revision history; a single shifted clause number means re-checking every cross-reference. With Codex, one instruction, review this services agreement from the contractor's side using our past exchanges with the client, the right base template and the relevant skill checklist, is enough for it to search the folder, read the contract, check the differences, and prepare a review approach, Word edits and a first draft of the points to confirm. The lawyer still checks everything at the end, but spends less time file-hunting and comparing clauses, and more on the business background and where the negotiation should land.

Example: legal due diligence and consultations

Due diligence involves a large volume of material: corporate registries, shareholder registers, investment agreements. Existing DD-efficiency services usually assume files are uploaded one by one, itself a burden in a data room that keeps changing. With Codex, the upload step disappears: pointed at the folder holding disclosed material, it can build a document inventory, flag what is missing, read key sections, reference past template reports, and draft the issue list, question list and report. Legal Agent also built an MCP for Excel, so follow-up question lists can be handled through Codex too.

The same applies to legal consultations. A business team might ask whether a campaign or certain terms are fine; a lawyer answering that also checks past similar consultations and whether the matter went to management before. Asking Codex to find similar past consultations and draft a first-cut answer changes both quality and speed. Legal Agent has also digitized in-house reference books, statutes and guidelines, within what is lawfully permitted, so AI can search them, and doing research inside Codex avoids the fragmented workflow of a separate book-search tool.

The lawyer's judgment still matters

This might sound like Codex makes a lawyer unnecessary. I do not think that follows; if anything, the more Codex is used, the more the lawyer's judgment matters. AI can find past matters and templates, compare clauses, and draft comments, but deciding which risk to accept, how far to negotiate, and what needs to go to management stays with the lawyer. Codex is not a substitute for that final call; it exists to find material and prepare a first draft so the lawyer can spend time on judgment instead.

A shift in how a law firm operates

In the traditional model, a junior lawyer searches for precedent, drafts a first version, and a senior lawyer reviews it, which has real training value. But if every matter still requires manual file-hunting and comparison in the same way, that part can change: Codex searches for precedent, proposes candidates and prepares a working file, while the lawyer judges whether the candidates are right and sets the direction. That is not simply adding headcount, it is building an operation that assumes AI from the start, which I see as one form an AI Native law firm can take.

What Legal Agent is working toward

At Legal Agent, lawyers and AI Agents work together to support contract review, legal consultations and M&A, aiming to build an AI Native law firm that genuinely contributes to clients' growth.