Extending Legal Agent Bench to M&A Due Diligence

Our goal with Legal Agent Bench (LAB) is to create and distribute realistic, high-scale agent environments to evaluate agents’ ability to perform end-to-end legal work and support open model training and agent research. Today, we’re extending LAB to cover one of the most critical legal tasks: M&A due diligence.

Diligence underlies every merger and acquisition which collectively amounted to around $4.8 trillion dollars of economic activity in 2025. Diligence costs typically range from 1-4% of deal value, landing them between $50 and $200 billion per year. Much of that cost is spent reviewing a virtual data room (VDR), systematically working through a company's legal and financial history to identify risks, mitigate them, and confirm the deal lives up to expectations.

To develop LAB environments for diligence, we wanted to focus on emulating the depth and complexity of a real VDR. This requires building novel evaluative environments that scale both size and depth of evaluation. Historic benchmarks have either focused deeply on reasoning over relatively contained context or narrow task execution over larger datasets. Diligence requires a model to do both, reading hundreds of documents to identify many independent transaction risks.

To evaluate models on this problem, we created multiple synthetic VDRs spanning thousands of documents and embedded with issues spanning diligence specialties from tax to tech transactions. In an exemplary environment, agents are given a VDR with tens of millions of tokens of context requiring them to find and remediate dozens of issues with their work validated by hundreds of rubric criteria.

Throughout the rest of this post, we explain what diligence looks like, how LAB diligence emulates a key piece of it, and how we are building specialized agents to help conduct diligence.

What Diligence Looks Like

Let’s say you want to buy a company. You meet with their principals and reach a rough agreement on a few key terms like price, form of payment, and confidentiality of the negotiations. This agreement is memorialized in a term sheet spanning a few pages. By the time the deal actually happens, that term sheet is replaced by an acquisition agreement sprawling hundreds of pages and detailing the mechanics of the acquisition, the obligations of the parties, and what happens if things go wrong. The work that shaped these additional pages is diligence.

At its core, legal diligence has two parts. The first is understanding and ascribing an accurate value to the underlying business. The second is allocating the risk of acquiring, combining, and integrating the business successfully. Both of these tasks require developing an understanding of the business from its primitives: the commercial contracts, employment agreements, IP portfolio, tax and regulatory documents, and other legal agreements that establish its rights and liabilities.

This is where the VDR comes in. After agreeing to the deal in principle, key documents (often numbering in the hundreds or thousands) are organized by the parties and added to a VDR. Once opened, teams of lawyers representing various specialties will work systematically through the VDR identifying risks, gaps, and follow-up questions. This initial pass will generate follow-up requests, interviews with key people at the target, and additional negotiations about the deal and subsequent disclosures. These reviews happen at all possible speed, with lawyers routinely working hundred-hour weeks to wrap their heads around every aspect of the business under tight timelines.

The resulting analysis is compiled into a diligence memorandum. This memo influences the final negotiations and planning on:

1. Deal Price: How the company is actually valued, as many legal issues identified in diligence can impact that valuation
2. Deal Structure: The nature of what is purchased (equity or assets) and the form of the purchase
3. Representations, Warranties, and Indemnities: What a seller is made to guarantee is true and correct about the business and what risks they are required to pay for if they materialize
4. Disclosure Schedules: What known issues a seller is explicitly off the hook for
5. Conditions and Consents: The third party consents or regulatory approvals that are required to close the deal
6. Post-Closing: How the two companies actually integrate into a productive new entity and other actions that are required post-closing

Effective diligence is not just knowing the company factually. It’s layering judgment on top of that factual record to understand what actually creates value for the business, what risks exist to that value, and how to negotiate those risks into a final agreement satisfactory to all parties.

Building a VDR

LAB’s diligence environments test agents’ ability to identify and action issues at the scale of realistic VDRs. As an example, take the VDR of Sentinel Cloud Security, which is being diligenced for a potential acquisition by Helios Cloud Holdings in a deal loosely modeled on Google’s $32 billion acquisition of Wiz in terms of industry, deal size, and acquisition type.

Sentinel’s VDR is a filesystem categorized by the key types of documents required to validate its business. Across these categories are more than 3,500 documents ranging from commercial contracts to litigation materials. Collectively, these documents total around 45 million tokens of context. Diligence requires both aggregating these millions of tokens into a coherent story about Sentinel and then identifying issues within that context.

These issues may be direct, a key customer may have the right to terminate a contract upon a change of control and no consent to the proposed acquisition has been obtained. It may be a missing file, there is no proof that the company owns or leases certain key offices. Or it may require reasoning across a number of clues: the company has a risky view of copyleft licenses that risks exposing some of its key IP. That view can only be sussed out by reviewing a mix of product counsel memos, technical specifications, and taking an opinionated view of how they fit together under current copyright laws.

The amount of context and the shape required to understand it makes diligence a uniquely hard problem for current-state agents. They cannot keep tens of millions of tokens in context, and task-oriented compaction strategies prevent them from forming a clear global picture of the VDR. Lossiness in compaction also means that subtle, multi-document issues are not picked up on as their threads are not maintained clearly enough for models to connect the dots. In practice, these fundamental issues are exacerbated by frontier model biases towards efficiency using keyword search and selective reading strategy rather than exhaustively reviewing documents at scale.

Diligence Agents

In practice, diligence is solved by exactly this kind of brute force: dozens of lawyers across various practice areas who collectively review the VDR for thousands of hours. Different specialists consider the target’s IP portfolio, their employment agreements, equity and compensation plans, commercial contracts, and financial and tax records. Findings across each are then consolidated into a diligence memo which is used to shape the deal terms and closing strategy.

To successfully diligence a LAB VDR, one or more agents takes on all of these roles, identifying issues holistically and drafting a first pass diligence memo. This memo is then checked against a rubric containing ground truth findings and recommendations for each issue planted in the VDR.

These rubrics allow us to explore strategies at both the harness and post-training level for effectively shaping agents that can engage in diligence. Doing so requires solving novel technical problems including:

• Context Management: Agents must read and make connections across information stretching many times their context window. Novel approaches to memory and compaction are required to allow them to effectively parse and retain key information while identifying and tracing risks.
• Exhaustive Review: Most agents are trained to identify the relevant portion of a large data space such as the relevant function in a code base. Their bias is to search efficiently, not completely. Diligence requires reversing this intuition and teaching them to check, and double check, every possible issue.
• Contextualized Judgment: A change-of-control in a million dollar contract may tank one deal, it may be an inconvenience to another. Agents must learn what issues matter, how much they matter, why they matter, and how best to remediate them.

Agents that can do all of these are helpful for diligence. But, diligence is a team sport. To actually do diligence an agent must also be able to (1) source its findings to specific documents and be capable of explaining or defending them; (2) communicate recommendations clearly, including alternative approaches where multiple valid strategies exist; and (3) present it all at the right level of detail for different stakeholders.

We believe that agents will learn to be maximally effective participants in a diligence team the same way junior associates do: via incisive feedback from experienced practitioners. This is why our diligence environments are built not just for research, but as a data-safe way to collaborate with customers to train models using their feedback. Our customers are the ones trusted with billion dollar deals today; their agents will be the ones trusted with those same deals tomorrow.

What’s Next

In the coming weeks we will publish our research identifying strategies for effective diligence agents and initial results across a diverse set of VDRs. We will also release additional LAB extensions covering tasks ranging from enterprise search, to fund formation, to investigations and discovery.

In parallel, we will work on moving these worlds from research to production, showing how agents can be improved through natural language feedback and collaborating with our customers to refine custom models that solve hard problems the way that they do.