What does Sullivan & Cromwell use AI for in agreement review?

Lawyers supply agreements, selected precedents and client or firm positions, and the Agreement Analyzer reviews the agreement against them, preparing a memorandum and tracked changes. Codex was used to develop the tool, alongside ChatGPT Enterprise.

Where do lawyers stay involved in Sullivan & Cromwell's review?

The source says lawyers inspect the supporting precedents and then accept, adapt or reject the changes. No independent accuracy study is disclosed. As a proposed check, qualified lawyers would adjudicate seeded clause issues and acceptable alternatives separately from the generating agents.

How strong is the evidence for Sullivan & Cromwell's tool?

Evidence is early-stage. The published account says the Agreement Analyzer grew from a single-user proof of concept to simultaneous multi-lawyer use. It discloses no deployment volume, accuracy results or measured time gain, and an adversarial architecture should not be counted as deployed.

What would an enterprise need to build something similar?

This is Nuvepro's proposed design, not the firm's. It would use a controlled document workspace, versioned drafts, playbooks and precedents, immutable source versions for rerunning review tasks, and held-out negotiation situations that test whether client-specific tradeoffs are preserved.

Limited published evidence · Legal services

AI agreement review: Sullivan & Cromwell example

Sullivan & Cromwell uses ChatGPT Enterprise, with Codex used to develop the Agreement Analyzer, to review agreements against precedents and negotiating playbooks and prepare memoranda and tracked changes. Lawyers inspect sources and accept, adapt or reject each change.

Agreement review against precedents and negotiating playbooks; memorandum and tracked-change preparation

Sullivan & Cromwell

Published account: Implemented Agreement Analyzer expanded from single-user proof of concept to simultaneous multi-lawyer use; agent-specific runtime architecture remains unproven.

A published third-party example, not a Nuvepro customer case or evidence of a partnership.

The workflow at a glance

Sullivan & Cromwell reported work

Illustrative summary of the published account

Reported work components. Their arrangement does not imply a fixed execution order or simultaneous work.

Lawyer inputs
Agreement, precedents and positions
Selected firm and client material grounds agreement analysis.
Review artifacts
Memorandum and tracked changes
Analysis prepares lawyer-facing review material.
Lawyer decisions
Accept, adapt or reject
Lawyers inspect sources and decide which changes to use.

Published account: Implemented Agreement Analyzer expanded from single-user proof of concept to simultaneous multi-lawyer use; agent-specific runtime architecture remains unproven. No disclosed model/tool loop for every implemented step, deployment volume, accuracy results or measured time gain. See the source account and limitations below.

What the source reports

Lawyers supply agreements, selected precedents and client/firm positions, then inspect sources and accept, adapt or reject changes. Do not count the potential adversarial architecture as deployed.

Reported technology: ChatGPT Enterprise; Codex used to develop Agreement Analyzer. Scope and outcomes are source-reported.

What the enterprise environment would need

The following is a proposed project design for an enterprise doing similar work. Published details are marked; the remaining systems, data and scenarios would be agreed for your project.

Proposed practice environment

Review agreement terms against client instructions

Illustrative project design to scope with your enterprise.

1 · Set up the environment
Negotiation document set
Fictional agreement versions, selected precedents and client instructions.
2 · Preparation and agent contributions
Draft review and redline
Identify clause issues with supporting precedent and acceptable alternatives.
3 · Independent checks
Independent legal adjudication
Qualified reviewers check issue coverage, cross-clause effects and cited authority.
4 · People decide acceptance
Matter lawyer accepts draft
Review client-specific tradeoffs and unresolved authority before accepting practice scope.

Only within the accepted practice scope

Practice output

A reviewable evidence package

A source-linked memorandum, proposed redline and issue register.

If a check fails → withhold acceptance.

Conflicting precedents or absent client authority produce an unsupported clause change. Withhold acceptance, repair, independently recheck and return to owner review.

Failure or owner rejection returns to review

Hold
→ 2 · Revise work
→ 3 · Rerun independent checks
→ 4 · Owner review

Operating boundary. Controlled document drafts only; no client advice, executed agreement or inferred adversarial-agent deployment.

Proposed, not an executed Nuvepro project or a description of the named company’s implementation. Owners agree the data, rules, responsibilities and acceptance criteria for their enterprise.

Systems and tools: Inferred: controlled document workspace, precedent repository and tracked-change tooling; avoid unnecessary enterprise replica.
Data and state: Inferred: versioned draft, playbook and precedent set plus clause interactions form relevant state.
People and handoffs: Published lawyer review/adapt/reject; inferred: different client instructions need separately scoped reviewers.
Rules and authority: Inferred: client authority, confidentiality and source-backed interpretations must be explicit.
Normal and exception cases: Inferred: conflicting precedents, absent authority, cross-clause interactions and legitimate alternative positions.
Reset and replay: Not published. Proposed: retain immutable source/draft versions and rerun defined review tasks.

What needs checking

Published: Lawyers inspect supporting precedents and decide on changes; no independent accuracy study disclosed.
Proposed: Qualified lawyers adjudicate seeded clause issues and acceptable alternatives separately from generating agents.
Published: Firm/client playbooks and selected precedents ground the workflow.
Proposed: Review whether held-out negotiation situations preserve the client-specific tradeoffs that matter.

Review environment fidelity, human capability and agent performance separately. The enterprise’s accountable owners decide acceptance of the shared workflow and any operating use.

Questions about this example

What does Sullivan & Cromwell use AI for in agreement review?: Lawyers supply agreements, selected precedents and client or firm positions, and the Agreement Analyzer reviews the agreement against them, preparing a memorandum and tracked changes. Codex was used to develop the tool, alongside ChatGPT Enterprise.
Where do lawyers stay involved in Sullivan & Cromwell's review?: The source says lawyers inspect the supporting precedents and then accept, adapt or reject the changes. No independent accuracy study is disclosed. As a proposed check, qualified lawyers would adjudicate seeded clause issues and acceptable alternatives separately from the generating agents.
How strong is the evidence for Sullivan & Cromwell's tool?: Evidence is early-stage. The published account says the Agreement Analyzer grew from a single-user proof of concept to simultaneous multi-lawyer use. It discloses no deployment volume, accuracy results or measured time gain, and an adversarial architecture should not be counted as deployed.
What would an enterprise need to build something similar?: This is Nuvepro's proposed design, not the firm's. It would use a controlled document workspace, versioned drafts, playbooks and precedents, immutable source versions for rerunning review tasks, and held-out negotiation situations that test whether client-specific tradeoffs are preserved.

Shape this around your enterprise

Bring a similar workflow and the people responsible for it. Together, define the records and systems to use, the result your team will build or configure, and how its owners will check it. Nuvepro scopes the simulation, environment and preparation. Agree any existing-agent evaluation, implementation assistance and production responsibilities separately.

Discuss this work Reimagine this workflow