Ropes & Gray's In-House AI Agent for Deal Due Diligence

Background/Overview

On October 29, 2026, OpenAI published a legal case study on its official blog: The American law firm Ropes & Gray is using proprietary AI tools based on ChatGPT / Astra for Law to complete transactional due diligence, which traditionally requires a large number of junior lawyers to invest time. Founded in 1865, this law firm operates in various fields such as mergers and acquisitions, private equity, intellectual property, data privacy, antitrust, anti-corruption, employment, and taxation, handling a large volume of data room documents each year.

This is not a new SaaS product released by a SaaS company; it is an internally developed tool by the law firm. Ropes & Gray had already announced a partnership with OpenAI to develop an AI-powered due diligence tool on September 16, 2026, and OpenAI’s case study brought this initiative to the forefront again. For the legal industry, the significance lies not in “whether AI can read contracts,” but in “whether AI can free lawyers from the pile of documents and allow them to focus on judgment and advice.”

Howard Glazer, Partner and Head of Practice AI at Ropes & Gray, said bluntly in an interview: “Clients pay lawyers not for contract reading, but for judgments based on contracts.” This statement highlights the core objective of the project: to delegate repetitive, mechanical document reviews to AI agents, allowing lawyers to focus on more valuable work.

In traditional transactional due diligence, junior lawyers and legal staff often spend weeks searching through thousands of documents for abnormal clauses, cross-checking risk points, and compiling issue lists. This process is tedious, time-consuming, and highly dependent on human effort. Ropes & Gray’s approach essentially extracts the rule-based and patterned parts of this process to be completed by AI agents, while lawyers focus on review, judgment, and client communication.

Technical Principles/Architecture

According to public information, this tool does not simply send documents to a large model for summarization. Instead, it designs a custom agentic workflow based on the real workflow that lawyers follow after entering the data room. Its core tasks include identifying risks, highlighting abnormalities, and uncovering key issues early in the due diligence process.

Specifically, the workflow is roughly divided into three steps:

The first step is document ingestion and structuring. The files in the data room come in various formats—PDF, Word, Excel, emails, or scanned copies. The tool needs to extract these unstructured or semi-structured content into a machine-readable format, which then can be indexed and searched by the model. This step typically involves OCR, layout analysis, table recognition, and multilingual processing, forming the underlying foundation for all subsequent analyses.

Step two: Risk identification and cross-checking. The model will scan contracts, financial statements, litigation records, regulatory documents, etc., based on transaction types and areas of interest, and mark out terms that may pose obstacles to the transaction or affect valuation. For example, non-compete clauses in employment contracts, flaws in IP agreements, compliance risks in privacy policies, potential penalties in environmental permits, etc. The agent here does not simply perform a search but makes preliminary judgments based on domain knowledge.

Step three: Generate an issue-level diligence report. The system compiles the identified issues into a structured report, sorting them by topic and severity, allowing partners to quickly identify areas that require focused attention. The report also includes references for lawyers to review item by item.

This system is built using OpenAI’s Astra for Law / ChatGPT, developed through multiple iterative meetings between lawyers from Ropes & Gray, IT staff, and forward-deployed engineers from OpenAI. It is not a ready-made SaaS but a customized agent deeply integrated into the law firm’s internal processes. This means that its prompts, knowledge base, review criteria, and output format are all tailored around Ropes & Gray’s own business practices.

From a technical architecture perspective, it resembles a legal AI platform with a “human in the loop” model: at the bottom is the large language model, with modules for document parsing, retrieval-enhanced generation (RAG), risk classification, and report generation in between. Above all, there is the lawyer’s review, editing, and final approval. The agent’s role is to improve efficiency, not to replace final judgment.

Key Data/Case Studies

Currently, the quantitative information publicly disclosed is very limited. Ropes & Gray states that the issue-level diligence report, which previously took days or even weeks to complete, can now be done in just hours. However, regarding the ratio from “days/wks” to “hours,” the official has not defined sample size, number of pages in the report, or complexity, so it cannot be directly cited as a universal conclusion.

Known boundary information includes: the tool covers at least ten interdisciplinary fields such as finance, tax, IP, data privacy, regulatory, employment, antitrust, anti-corruption, international risk, litigation; the teams involved in development include lawyers, IT staff, and OpenAI engineers stationed there; the product is proprietary and not sold externally.

Publicly unavailable data includes: tool accuracy, manual review ratio, number of actual deployed lawyers, volume of documents processed, customer feedback, pricing model, and whether it has indeed reduced the demand for junior lawyer positions. All performance metrics come from Ropes & Gray or OpenAI official sources, with no independent third-party evaluations yet. Compared to other publicly available benchmarks in the legal technology field, this case seems more like an internal capability demonstration of the firm rather than an externally verified performance report. Readers should distinguish between “official publicity,” “media reports,” and “independent evaluations” when evaluating.

It is worth mentioning that the official proposal specifically emphasizes the transformation of lawyers’ work methods through tools, rather than just efficiency metrics. This indicates that Ropes & Gray values the change in workflow more than simple cost reduction. For readers, this means we cannot simply evaluate its value by “how many hours saved,” but rather by how much it liberates lawyers from low-value labor.

Impact/Significance

The most direct impact on the legal technology industry is that top law firms are beginning to integrate AI agents into core business processes, rather than keeping them as pilot or demonstration projects. Due diligence is one of the most labor-intensive and standardized parts of M&A transactions, and Ropes & Gray’s approach sets a precedent for peers—if agents can assist top law firms with due diligence, other firms and legal departments will also be forced to reevaluate their technology investments.

From the perspective of lawyers’ work, it may change the career development paths of junior associates. In the past, junior lawyers accumulated experience by reading extensive contracts and preparing due diligence reports. If agents take on document organization and initial draft generation, how junior lawyers learn risk identification skills will become a new training topic. This may also prompt law schools and legal training institutions to introduce AI-assisted collaborative training earlier.

From a product perspective, it demonstrates the choice between “self-developed vs. purchased” solutions. Ropes & Gray’s decision to collaborate with OpenAI instead of purchasing general legal technology SaaS shows that in highly specialized fields with sensitive client data, off-the-shelf products rarely meet law firms’ requirements for accuracy, confidentiality, and depth of workflow. In the future, more large law firms may choose to deeply collaborate with tech companies to build their own proprietary tools.

From the client perspective, the reduction in due diligence costs and improved quality consistency may change the bargaining power in M&A transactions. Although lawyer fees are unlikely to decrease significantly due to AI in the short term, faster delivery speeds and reduced risk of omissions will change clients’ expectations for legal services.

Controversies/Risks

First is data security and confidentiality. Due diligence transactions involve a large amount of confidential financial, business, and legal information. Ensuring that this data does not enter model training or is not accessed by other clients is a question both law firms and clients must answer. Although OpenAI has data protection commitments for corporate clients, specific implementation details still depend on contracts and audits. For cross-border transactions, data export rules in different jurisdictions also increase compliance complexity.

Second is liability attribution. If the risks identified by AI are overlooked or misjudged, the ultimate responsibility still falls on lawyers and law firms. This means that the output of agents must be manually reviewed and cannot simply be considered “AI has already checked.” Howard Glazer’s words also hint at this: AI is a supplementary judgment, not a replacement. In the foreseeable future, lawyers’ professional responsibilities will not be reduced due to the use of AI.

Third is the exaggeration of risks in promotional messaging. The official highlight is “compressing days/weeks into hours,” but there is a lack of transparency. The complexity of different transactions varies greatly, and simple time comparisons can easily lead readers to overestimate current capabilities. As readers, we need to distinguish between “some processes can be accelerated” and “they have already fully replaced human work.” Before more independent third-party evaluations appear, any conclusions about efficiency and accuracy should be approached with caution.

The fourth factor is the model’s potential for error. When generating responses, large language models sometimes arrive at seemingly reasonable but actually incorrect conclusions with high confidence. In scenarios where tolerance for errors is extremely low, such as legal due diligence, an error labeled as “identified risk” can be more dangerous than missing information. Therefore, the output of an agent must not only be reviewed but also traceable: it must be clear which document, page, or clause each conclusion originates from.

Conclusion/Prospects

Ropes & Gray’s collaboration with OpenAI is a landmark example of AI agents entering the core workflow of high-end professional services. It demonstrates how agent-based workflows can be implemented in legal due diligence: instead of letting AI make judgments for lawyers, it allows AI to complete repetitive and mechanical document processing first, leaving room for lawyers to make judgments.

In the short term, these tools will first become widespread within large law firms and corporate legal departments, mainly used for initial drafts, risk pre-assessment, and document indexing. In the medium term, they may reshape the work content and training methods for junior lawyers. In the long term, as model capabilities improve, the role of agents in professional services will evolve from “draft generators” to “collaborative judgment-makers,” but manual review and ultimate responsibility will remain an essential foundation in the legal industry.

What deserves most attention in the future is not which law firm uses AI, but how these tools balance efficiency, confidentiality, responsibility, and cost. Only when technical capabilities, institutional design, and professional ethics all advance simultaneously can AI’s value in high-risk professional services like law be truly realized. For practitioners and observers, a healthy criterion for judgment is whether AI gives lawyers more time to make judgments, rather than providing them more reasons not to.

Reference Sources

• OpenAI official blog: Ropes & Gray uses OpenAI to accelerate due diligence
• Ropes & Gray official press release (2026-09-16)
• Law.com report (2026-09-23)