AI review workflows are most useful when they combine selected source material, an organisation’s own way of working and an accountable expert who reviews the result. OpenAI’s latest customer case study shows why the strongest workflow is rarely a generic chatbot pointed at a folder of documents.
Sullivan & Cromwell, a global law firm, worked with OpenAI to develop Agreement Analyzer, a firm-specific tool for reviewing draft agreements. The legal use case is specialised, but the operating principles are relevant to businesses in the United States, the United Kingdom, India and other English-speaking markets.
This is not a generally available legal product, and the case study should not be read as legal advice. It is a practical example of how expert knowledge can shape an AI-assisted review workflow.
What expert-led AI review workflows include
According to OpenAI, Agreement Analyzer began as a single-user proof of concept built with Codex. Sullivan & Cromwell then worked with OpenAI to improve output quality, run parts of the analysis in parallel and support multiple lawyers using the tool.
The workflow brings together a draft agreement, precedents selected by the lawyer, negotiating playbooks, firm and client positions, market-practice information and ChatGPT Enterprise. It runs multiple review cycles from different perspectives, combines the findings and returns a memorandum with proposed tracked changes. A lawyer then reviews, adapts or rejects those suggestions.
That last step matters. The tool does not remove professional judgment from the process. It changes where people spend their attention: less time assembling first-pass comparisons and more time evaluating the issues that could affect the client.
1. Start with a defined decision, not a broad AI ambition
Agreement review is a clear workflow. It has a known input, a repeatable set of checks and a concrete output that a qualified person can inspect. That makes it easier to decide where AI helps and where human approval remains essential.
A small or mid-sized business can apply the same discipline outside legal work. Examples include comparing supplier proposals against an approved checklist, reviewing sales enquiries against qualification rules, checking project reports for missing evidence or preparing a first-pass summary of policy changes. The right starting point is a recurring decision with an identifiable owner.
Before building anything, write down the input, the questions the workflow must answer, the permitted sources, the expected output and the person accountable for accepting it. If those elements are unclear, automation will usually make the ambiguity faster rather than solve it.
2. Curated knowledge is part of the product
The case study describes a system grounded in selected precedents and playbooks. That is different from asking a model to rely on broad general knowledge. The organisation chooses which material should influence the result.
For AI review workflows in other businesses, curated knowledge might include current operating procedures, approved price lists, product documentation, customer-service policies, example deliverables and escalation rules. Someone must own those sources, remove obsolete material and decide which version is authoritative.
This work often determines the value of the finished system. A capable model cannot compensate for contradictory instructions, unnamed document owners or a folder containing five versions of the same policy.
3. Make evidence and uncertainty visible
OpenAI’s case study emphasises transparency. Users can inspect issues, explanations and relevant precedents before deciding what belongs in the final output. It also describes an adversarial pattern in which agents form conclusions and other agents challenge them, separating supported findings from dismissed or still-unverified issues.
In AI review workflows, that pattern is useful well beyond professional services. An AI-generated recommendation should show where the supporting information came from. When evidence is incomplete, the system should say so and route the item for review instead of presenting a confident answer.
A practical pilot can begin with three labels: supported, needs review and unsupported. Measure how often reviewers agree with those labels and which source gaps create repeated uncertainty. This produces better evidence for an investment decision than a polished demonstration alone.
4. Human review needs a real operating design
In AI review workflows, adding the words “human in the loop” to a diagram is not enough. The reviewer needs time, authority and a clear view of what changed. The workflow should show the original input, the AI suggestion, the cited source and any uncertainty in one place.
Teams should also decide what happens when the reviewer disagrees. Can they edit the result, reject it, request another pass or escalate it? Is that feedback captured so the process can improve? Who covers the queue when the usual reviewer is unavailable?
For global teams, including operations split between India, North America and Europe, the handoff must also account for time zones, access permissions and data-handling requirements. These are implementation choices, not details to leave until launch.
5. Prove the workflow with a narrow pilot
Sullivan & Cromwell’s system began as a single-user proof of concept. That sequence is a sensible model for smaller organisations: select one workflow, one accountable expert and a controlled set of source documents.
Run the pilot on representative historical cases before it touches live work. Track review time, the percentage of useful findings, unsupported suggestions, corrections made by the reviewer and cases that still require the original manual process. Compare the complete effort, including preparation and review, with the current baseline.
AI review workflows should begin with a narrow pilot that also limits permissions. The first version may only need read access to a controlled document collection and the ability to prepare a draft. Broader integrations and write access can wait until the team has evidence that the workflow is accurate, useful and governable.
What this means for business leaders
The most transferable lesson is that AI review workflows depend on more than model capability. The surrounding design determines whether the output is useful: authoritative sources, organisation-specific rules, visible evidence, clear uncertainty and accountable review.
Businesses do not need to copy a law firm’s system. They can use the same questions to evaluate their own opportunity:
- Which recurring review consumes expert time?
- Which sources should the AI be allowed to use?
- What evidence must accompany every finding?
- Which decisions always require human approval?
- What would a small, measurable pilot look like?
For many teams, the first valuable AI project will not be an autonomous agent. It will be a carefully bounded assistant that prepares better evidence for a person who remains responsible for the decision.
For another practical example of a controlled pilot, read our guide to what recent OpenAI changes mean for small businesses.
Plan an expert-led AI review pilot
trulyAI helps businesses design AI review workflows by mapping a recurring knowledge-work process, selecting the right source material and creating a controlled pilot with measurable review steps. If your team has a document-heavy workflow and a named subject-matter expert, we can help you identify a practical first use case.
Source: OpenAI, “Sullivan & Cromwell brings legal expertise into agreement review,” 15 October 2026.

Leave a Reply