How Law Firms Can Build Legal Tools with Codex: Inside Sullivan & Cromwell's Approach

When top law firms start using AI coding tools to build custom applications, the way legal services are produced is being redefined. A video released by OpenAI shows how Sullivan & Cromwell is using Codex to move contract review from "lawyers reading word by word" to a new workflow of "AI-generated memos + track changes markup."

From Draft Contracts to Intelligent Review: A Reusable Workflow

According to the OpenAI video, lawyers at Sullivan & Cromwell input the draft contract received from opposing counsel into an AI model, along with the most relevant precedents, a playbook representing the client's position, and market conventions. The model quickly outputs a memo listing all identified issues, complete transparency on the rationale for each decision, and a track changes markup for lawyers to review and revise. Lawyers then adjust it and deliver it to the client.

The key to this process is "auditability": rather than delivering a black-box conclusion, the AI acts like a junior lawyer, laying out the chain of reasoning and every edit for senior lawyers to review item by item.

Playbook in the legal context is not a sports tactics manual, but a negotiation-position document that a law firm or corporate legal department prepares in advance for a specific type of transaction, typically specifying which clauses must be insisted upon, which can be conceded, and which are red lines. For example, on the indemnity clause of an M&A contract, the playbook would set out the acceptable range and exceptions for the indemnity cap. Inputting the playbook together with precedents into the model is equivalent to encoding the firm's years of accumulated negotiation strategy as constraints on the AI's decisions, so that the suggested output does not deviate from the client's position. This also explains why the same draft contract yields different review results in different firms' hands—the difference comes mainly from the playbook, not the model itself.

The Role of Codex: Letting Lawyers Build Their Own Tools

The video emphasizes that the team uses coding tools like Codex to "produce bespoke applications." This means lawyers no longer depend entirely on IT department scheduling, but can quickly build review tools around specific transaction types.

"You never have enough time to do everything you want to do," a interviewee in the video says. The value of Codex lies in automating repetitive contract comparison and issue identification, freeing lawyers' time for deeper analysis and more sophisticated output.

Codex is a series of code-generation models launched by OpenAI, initially known as the underlying engine of GitHub Copilot, capable of generating, completing, and modifying code based on natural language descriptions. Unlike end-user-facing vertical legal software, Codex itself contains no legal knowledge; its value lies in lowering the barrier to "turning business logic into a runnable program." When a law firm uses it, a lawyer familiar with the business process needs to describe the requirements, the model generates a script or application skeleton, and then technicians or lawyers debug it themselves. This "low-code/generative programming" model drastically reduces the marginal cost of custom tools, but it also requires users to have basic software engineering literacy and to establish code review and version management mechanisms; otherwise, custom applications may become a new maintenance burden.

Dual Gains in Speed and Depth

According to the video, this approach is "much faster than a lawyer on their own could do this." But speed is not the only benefit—the interviewee emphasizes that it enables the firm to "do deeper analysis and produce more refined work product for clients."

This points to a trend: the role of AI in the legal field is not to replace lawyers' judgment, but to free lawyers from mechanical labor and focus their energy on strategic judgment and client communication.

Transparency by Design: The Trust Foundation for Legal AI Adoption

Legal work has extremely low tolerance for error, so the video particularly highlights "full transparency"—the model must explain the rationale for each decision and present suggested edits in the form of track changes.

This design keeps lawyers in the control loop at all times: the AI proposes suggestions, and lawyers review, adjust, and finalize. For law firms accountable to clients, explainability is not a nice-to-have but a precondition for adoption.

Track changes markup is a collaboration feature in word processing software such as Word that records every edit to a document and its author in the form of insertions, deletions, and comments. In legal practice, it is not only an editing tool but also a vehicle for accountability and negotiation communication: opposing counsel can see through the markup who proposed each change and why, and courts in some proceedings require submission of versions with tracked changes. When AI outputs suggestions in the form of track changes, it means its edits can be accepted, rejected, or further modified item by item, and coexist with human edits in the same document history. This both satisfies the firm's requirement for auditability and avoids the ambiguity of responsibility that "AI directly rewriting the contract" would bring.

Implications for the Legal Tech Industry

Sullivan & Cromwell's practice demonstrates the potential of combining "general-purpose coding tools + domain knowledge." Law firms need not wait for vertical legal AI products to mature; instead, they can use tools like Codex to encapsulate their own precedent libraries, playbooks, and business judgment into custom applications. For legal tech entrepreneurs, the focus of competition may shift from "who has a better model" to "who better understands how to embed models into lawyers' real workflows."