AI Due Diligence for Transaction Lawyers: Testing Libra Discovery on an M&A Review

Due diligence is one of the most labor-intensive parts of an M&A transaction. Whether the data room contains ten material contracts or hundreds of agreements, the underlying exercise is often highly repetitive: identify the document, determine what kind of agreement it is, locate the provisions the deal team cares about, extract the relevant language, summarize it in a consistent format and flag anything that might affect the transaction. Multiply that across dozens or hundreds of contracts and it becomes easy to understand why due diligence can consume a substantial portion of a transaction team’s time.

The irony is that much of this work is repetitive without being trivial. A transaction lawyer may repeatedly search for change-of-control clauses, assignment restrictions, termination rights, consent requirements, exclusivity provisions, renewal mechanics, liability limitations and other provisions across a large document set. Locating those provisions is often mechanical. Understanding what they mean for the transaction is not. That distinction is precisely why due diligence is becoming such an interesting application for legal AI: AI can potentially absorb much of the locating, extracting and structuring work while leaving the lawyer to verify the result and determine its legal and commercial significance.

AI-assisted diligence itself is not entirely new. Sullivan & Cromwell notes in its 2026 guidance on the use of AI tools in M&A transactions that machine-learning tools have been used in transactions for years, particularly to identify and extract contractual provisions from virtual data rooms. What generative AI changes is the flexibility of that process. Instead of relying solely on narrowly trained extraction models, lawyers can increasingly define the questions they want answered and create structured review workflows around the needs of a particular transaction.

Recently, I tested Libra’s Discovery feature using an M&A due diligence exercise. My question was not whether AI could produce an impressive demonstration. It was much more practical: could I give the system a group of contracts, define the information a transaction lawyer would ordinarily collect in a diligence tracker, and get back a usable first-pass work product that I could actually verify?

What is Libra Discovery?

The simplest way to think about Discovery is as a smart spreadsheet for document review. Instead of opening agreements one at a time and manually entering findings into Excel, the lawyer uploads a collection of documents and defines the questions or fields that should be analyzed across that collection. Libra currently describes Discovery as a tool that can extract “structured, verifiable insights” from hundreds of documents for legal due diligence. The important words there are structured and verifiable.

Structure matters because transaction diligence rarely ends with a narrative answer. Deal teams need information that can be compared across documents. Which contracts require consent? Which contain a change-of-control provision? Which permit assignment? Which counterparties have termination rights? Which agreements renew automatically? A traditional due diligence tracker effectively converts a collection of unstructured contracts into a structured dataset. Discovery attempts to automate much of that conversion.

For an M&A transaction, the lawyer can create columns around the issues relevant to the deal and then apply those questions across the document set. Depending on the transaction, that might include change of control, assignment, termination, consent, exclusivity, renewal, governing law, limitations of liability, unusual obligations or simply whether a provision is missing. The result is presented as a table rather than a series of disconnected chatbot answers.

Why this matters for M&A lawyers

Consider what happens in a conventional contract diligence exercise. A junior lawyer receives a list of agreements and a diligence scope. The lawyer opens the first contract, searches for the relevant provisions, reads enough surrounding language to understand them, writes a summary in the diligence tracker and moves to the next agreement. The process repeats over and over. The work may be divided among several associates, which then creates another challenge: consistency. Different lawyers may describe effectively identical provisions differently, use different levels of detail or interpret the scope of a requested field differently.

AI changes the starting point. Instead of beginning with an empty spreadsheet and a folder full of contracts, the lawyer can potentially begin with a populated table generated from the documents. That does not eliminate review, but it changes the nature of the work. The first question is no longer simply, “Where is the assignment clause?” It becomes, “Has the system identified the correct provision, has it characterized it correctly, and what does this mean for our deal?”

That is a much more important transformation than simply saying that AI makes contract review faster. In a traditional workflow, significant lawyer time is consumed before substantive analysis even begins because the information first has to be located and organized. If technology can reliably complete a meaningful portion of that preparatory work, lawyer time can move further up the value chain toward identifying risk, understanding transaction consequences, deciding what requires disclosure or consent and communicating those issues to the client and deal team.

How I approached the test

For my test, I approached Discovery as I would approach the construction of a diligence tracker rather than as a generic AI chatbot. I first thought about the information I wanted to extract and how I wanted that information presented. That step matters. AI does not remove the need to design the diligence process. A poorly conceived set of questions will simply produce a well-formatted version of a poorly conceived diligence exercise.

I then uploaded multiple agreements and created the categories I wanted Discovery to analyze across the set. The objective was to see whether the system could turn those documents into something resembling a first-pass diligence spreadsheet. Instead of asking one question about one contract, I wanted the same analytical framework applied consistently across multiple agreements.

This distinction is important for transactional lawyers. The highest-value use of AI is not necessarily asking, “What does this contract say?” one document at a time. If that is all we do, we have essentially replaced Ctrl+F with a more sophisticated search box. The more interesting workflow is to define the review once and then apply that structure across the entire population of documents.

The feature I cared about most: citations

The feature that stood out most in my testing was not the speed of extraction. It was the ability to trace the answer back to the underlying contract. Libra emphasizes verifiability across its product, and Discovery is much more useful to a lawyer when an extracted finding is accompanied by the source language and the location in the document from which that finding was derived. That gives the reviewer a path back to the evidence instead of presenting the AI’s conclusion as a black box.

This is critical in legal work because a plausible answer is not enough. Imagine a diligence tracker stating that an agreement requires consent upon a change of control. Before that conclusion is communicated to a client, a lawyer needs to see the provision itself. Does the clause apply to a direct change in ownership or only an assignment? Is there an affiliate exception? Is consent required, or merely notice? Does the counterparty have a termination right? Is there another provision elsewhere in the agreement that modifies the result? A citation makes verification possible; it does not make verification unnecessary.

I also liked that, in my testing, a blank result was not always presented as an unexplained empty cell. The system could provide context and supporting references for why it reached a result. That matters because “no provision found” can itself become a legally meaningful conclusion. A lawyer needs some ability to assess whether the provision is genuinely absent or whether the AI simply failed to identify it.

The most useful legal AI does not merely give the lawyer an answer. It makes the answer easier to verify.

Can AI replace lawyer review in due diligence?

No, and I do not think that is the most useful way to frame the opportunity. AI-generated diligence findings should not simply be accepted and forwarded to a client because they appear in a neat spreadsheet. The ABA’s guidance on AI and technology in M&A practice specifically emphasizes the importance of careful human review, including the risk that generative AI can produce inaccurate or nonsensical outputs. The same guidance points to confidentiality, supervision, data ownership and workflow design as issues lawyers need to consider when adopting these systems.

But saying that a lawyer must review the output does not mean the technology has failed to create value. Transaction lawyers already work through layers of review. A junior associate may prepare a tracker that is reviewed by a senior associate, who identifies issues for the partner or client. The important question is whether AI can reliably accelerate the first layer of that process without compromising the quality of the final work product.

The potential efficiency is therefore not “AI does the diligence and the lawyer goes home.” It is that the lawyer begins from a materially more advanced starting point. If the system has already identified candidate provisions, extracted relevant language and placed the results into a consistent structure, the lawyer can spend more time checking edge cases, understanding commercial significance and deciding what actually belongs in the diligence report.

The real value is the first pass

This is where I think lawyers should be careful about the language used around legal AI. “Automation” can create the impression that an entire professional task disappears. In diligence, the more realistic near-term model is accelerated first-pass review. AI can perform much of the repetitive work required to surface and organize information, while lawyers remain responsible for confirming the result and determining its significance.

The second workflow does not eliminate the lawyer. It removes several of the steps that consume lawyer time before the most valuable part of the exercise begins. That can be particularly significant when a deal contains dozens or hundreds of agreements and the same questions need to be answered repeatedly.

AI may also improve consistency across the review

Speed is the most obvious selling point, but consistency may be almost as important. Human diligence teams are not perfectly consistent. Different reviewers use different language, emphasize different details and occasionally interpret the scope of a diligence question differently. A structured AI workflow can apply the same question and requested output format across every document in a set.

That does not guarantee that every answer is correct. It does, however, give the team a more standardized first-pass methodology. This is one reason Thomson Reuters describes AI-powered due diligence as useful not only for identifying obligations and exposures but also for applying repeatable review frameworks across document populations. Consistency matters when a deal team ultimately needs to compare contracts rather than simply understand each one in isolation.

What this changes about the role of the junior transaction lawyer

This is where the technology becomes more interesting than a simple productivity tool. Due diligence has traditionally been one of the ways junior transaction lawyers learn contracts. By reading hundreds of agreements, associates gradually learn what change-of-control language looks like, where assignment provisions hide, how termination rights vary and which provisions partners care about. If AI increasingly performs the first extraction, law firms will need to think carefully about how associates continue developing that pattern recognition and judgment.

At the same time, there is an opportunity to improve the work juniors actually do. Instead of spending the majority of an evening locating clauses and copying them into Excel, a junior lawyer could spend more of that time checking whether the extraction is right, comparing provisions across agreements, identifying unusual language and thinking about how each finding affects the transaction. That is arguably better training — but only if firms deliberately teach associates how to interrogate AI output rather than simply trust it.

Good AI diligence starts with good diligence design

One of the most important lessons from using tools like Discovery is that AI does not rescue a poorly designed review. Transaction lawyers still need to decide what matters. Which contracts are in scope? Which provisions should be reviewed? What should constitute a red flag? When should an issue be escalated? What should the output look like? Which findings require specialist review from employment, tax, IP, privacy or regulatory lawyers?

In other words, the diligence checklist increasingly becomes a form of workflow specification. The better the lawyer can express the legal and commercial questions in a structured way, the more effectively an AI system can assist. This is closely connected to legal engineering: taking legal judgment that previously existed in the minds of experienced lawyers or in an informal checklist and expressing it as a repeatable system that technology can help execute.

What lawyers still need to watch carefully

There are obvious limits. AI can miss a provision, misunderstand the interaction between clauses, mischaracterize unusual drafting or give a confident answer where the document is ambiguous. Diligence also involves questions that cannot be resolved by extracting text from individual agreements. A transaction lawyer may need to understand how several documents relate to each other, whether an amendment changes an earlier agreement, whether the data room is complete, whether a schedule has been omitted, or whether a seemingly ordinary provision creates unusual risk in the context of the specific deal.

Lawyers also need to think about confidentiality and platform governance before uploading transaction documents. Libra currently states that it is ISO 27001 certified, GDPR compliant and hosted on servers within the European Economic Area, but those product statements do not replace a firm’s own security, confidentiality and vendor-review process. Firms should understand what data is uploaded, how it is processed, retention policies, contractual protections, access controls and whether the particular use complies with professional obligations and client requirements.

The ABA’s M&A technology guidance similarly emphasizes that efficiency cannot be separated from the lawyer’s duties around confidentiality, supervision, competence and client information. The practical lesson is straightforward: the more useful these tools become, the more important it becomes to build them into a governed legal workflow rather than treating them as an individual productivity shortcut.

There are also limits to what contract extraction can tell you

This becomes particularly clear in complex or cross-border transactions. A recent Foley & Lardner analysis of AI in cross-border M&A makes an important point: accelerating the middle of the diligence process does not eliminate the difficult work at either end. Lawyers still need to design the scope of the review, understand the transaction, determine which legal regimes matter and then interpret the extracted information in context. Faster document review can compress the process dramatically, but it does not turn diligence into a purely computational exercise.

That distinction should shape expectations. If an AI system can analyze a hundred contracts faster than a team of associates, that is enormously valuable. But the transaction lawyer still needs to determine whether a particular consent condition threatens the closing timetable, whether a restrictive covenant affects the buyer’s integration plan, whether a termination right creates revenue risk, or whether an unusual liability provision should influence the purchase agreement or valuation.

Beyond M&A: the same model applies to other document-heavy work

Although I tested Discovery using an M&A diligence exercise, the underlying workflow is much broader. Any legal task that requires lawyers to ask the same set of questions across a large collection of documents is a candidate for structured AI review. The technology does not necessarily need to know that it is “doing M&A.” It needs a document population, a well-designed review framework and a clear output structure.

• Commercial leases: extract rent, renewal options, assignment provisions, change-of-control restrictions and termination rights.
• Employment agreements: identify compensation, restrictive covenants, notice provisions, change-of-control rights and termination terms.
• Vendor contracts: compare liability, indemnity, data protection, renewal and termination provisions.
• Software licenses: identify usage rights, audit provisions, assignment restrictions, sublicensing rights and renewal obligations.
• Regulatory reviews: extract recurring obligations or identify whether documents comply with a defined standard.
• Portfolio reviews: turn hundreds of agreements into structured data that can be compared across the entire contract population.

That is why Discovery-style tools interest me more than isolated contract chat. They point toward a model in which lawyers do not simply interact with AI one document at a time. They can begin to treat entire document populations as datasets and apply a consistent legal review framework across them.

The bigger shift: from reading documents to interrogating document sets

For decades, the basic unit of legal document review has been the document. Open the agreement, read it, summarize it and move to the next one. AI creates the possibility that the unit of analysis becomes the document set. Instead of asking what one agreement says, a lawyer can ask which of 100 agreements contain a particular risk, which counterparties have consent rights, which contracts deviate from the standard position or which documents are missing a provision entirely.

That is a meaningful shift in how transactional work can be organized. It transforms due diligence from a sequence of individual reading tasks into a structured data exercise with lawyer verification layered on top. The spreadsheet has always been the bridge between contracts and deal analysis. Tools like Discovery attempt to automate the creation of that bridge.

The future of due diligence may be less about asking lawyers to read every document from a blank page and more about giving them a cited, structured map of the document set and asking them to determine what matters.

Final thoughts

Libra’s Discovery feature provides a strong example of where legal AI is becoming genuinely useful for transaction lawyers. The attraction is not that the system magically replaces diligence. It is that it can take a repetitive, document-heavy process and give the lawyer a structured first pass across multiple agreements, with citations that make the results easier to verify.

That can materially change the economics and experience of diligence. Lawyers can spend less time locating the same categories of information across agreement after agreement and more time deciding whether those findings matter to the deal. But the quality of the outcome will still depend on the lawyer: defining the scope properly, building sensible review questions, checking the extracted information, recognizing what the system missed and applying legal and commercial judgment.

For me, that is the more credible way to think about AI in transactional practice. The goal is not to remove the transaction lawyer from due diligence. It is to remove as much unnecessary friction as possible before the work that actually requires a transaction lawyer begins.

---

Last updated: August 2026. This article reflects my testing and analysis of the product and is provided for educational purposes only. It is not legal advice.