Deal team reviewing contract extraction results from a data room

AI Due Diligence: What to Automate and What Stays Human in Deals

AI can speed document review and red-flag search in M&A, but confidentiality, privilege, data-room terms and model error set limits. Here is where to draw the line.

By Shayan Salehi 8 min read

Image: ColdAI

AI due diligence works best as a first-pass reader: it can sort a data room, extract terms from hundreds of contracts, and surface candidate red flags for a person to check. It does not replace the judgment that prices a deal, and it adds duties around confidentiality, privilege and error. Used well, it speeds the search. Used carelessly, it can leak the target's information or put a confident wrong answer into your valuation model.

What AI does well in a data room

A data room is a large body of documents with repeated structure: customer contracts, supplier agreements, employment terms, leases, licenses, board minutes. Pattern-heavy, high-volume reading is where language models earn their keep, because the task is to find and label specific provisions across many similar documents.

The tasks that suit a first pass are narrow and checkable:

  • Contract extraction: pulling the term, renewal mechanics, termination rights, assignment and change-of-control clauses, exclusivity and liability caps into a table.
  • Red-flag search: finding documents that contain a named risk, such as a most-favored-customer clause, a non-compete, or a consent requirement on a sale.
  • Gap detection: comparing the index of what the target uploaded against the diligence request list.
  • Consistency checks: comparing a figure in a management presentation with the same figure in a contract or ledger extract.
  • Summaries: producing a short, source-linked summary of a long document for a reviewer who then reads the cited passage.

Each of these has a defined output and a way to verify it. That matters, because the value of the tool depends on how cheaply a reviewer can confirm what it says.

What stays human

Some work should not be delegated, however good the extraction looks. The reason is not that the model is useless. It is that these decisions carry consequences that belong to a named professional.

Valuation judgment is the first. A model can list the customers whose contracts allow termination on a change of control. A deal team decides how likely those customers are to leave, what the loss does to the forecast, and whether the price or the structure should change.

Materiality is the second. A reviewer decides whether a finding is a rounding issue or a reason to walk away. The third is negotiation: turning a finding into a specific indemnity, escrow, condition or price adjustment. The fourth is reading what is missing. A model summarizes the documents it is given. An experienced reviewer asks why a document the business should have is not there.

Legal conclusions sit with counsel. The SRA's Risk Outlook on AI in the legal market, published on 20 November 2023, states that firms remain responsible and accountable for the outputs of any AI they use, and that responsibility cannot be handed to an IT team or an external provider. It also says AI should not be trusted to check its own accuracy.

Model error is a diligence risk in its own right

Language models produce plausible text, and plausible is not the same as correct. NIST's Generative AI Profile, NIST AI 600-1, published in July 2024, defines confabulation as the production of confidently stated but erroneous or false content, and notes that users may believe it because of the confident delivery.

There is evidence from legal work. A preregistered study by Stanford and Yale researchers in the Journal of Empirical Legal Studies evaluated AI legal research tools from LexisNexis and Thomson Reuters. It found that the providers' claims of eliminating hallucinations were overstated: the tools hallucinated between 17 percent and 33 percent of the time, though less than a general-purpose chatbot. The study tested legal research, not contract extraction, so its rates do not transfer to your use case. The lesson that does transfer is that a vendor's description of a tool is not a measurement of it.

Build your own measurement. Before relying on a tool for a deal, test it on a sample of documents where a lawyer already knows the answers. Record how often it misses a clause, invents one, or attaches the right clause to the wrong contract. Misses matter most in diligence, because a clause the tool did not find is a clause nobody reads.

Controls that reduce the exposure are simple:

  • Require every extracted item to cite the document and page, and have a reviewer open the source for anything that drives price or structure.
  • Sample the documents the tool marked as clean, not only the ones it flagged.
  • Keep the extraction prompts and tool versions in the deal file so that a result can be reproduced.
  • Treat a tool's silence as no information, not as confirmation that a risk is absent.

Confidentiality, privilege and data-room terms

The data room is the target's confidential information, shared under a non-disclosure agreement and the data-room rules. Putting those documents into an AI tool is a disclosure to whatever that tool's provider can see, store or train on. Read your NDA and the data-room terms before you upload anything, and ask whether they permit processing by a third-party AI service, whether they restrict copying, and what they say about retention and deletion.

Professional rules say the same thing about client information. ABA Formal Opinion 512, issued on July 29, 2024, addresses lawyers' use of generative AI. According to published summaries of the opinion, it says a lawyer needs the client's informed consent before entering information relating to the representation into a self-learning tool, and that boilerplate in an engagement letter is not enough. The opinion also says lawyers must have a reasonable understanding of a tool's capabilities and limitations, and must verify output before relying on it. Opinions of this kind are not binding law, and state rules differ.

The Solicitors Regulation Authority in England and Wales published a warning notice on AI misuse on 17 August 2026. It says client information should go into AI tools only with contractual, technical and organizational safeguards, that data should stay in a secure environment, should not be used for model training unless explicitly authorized, and should not be kept longer than necessary. It warns that putting client material into public AI tools risks breaching confidentiality and waiving legal professional privilege.

Privilege is a legal question for counsel, and the answer depends on the jurisdiction. The practical point for a deal team is to ask counsel before using any AI tool on privileged material, such as legal advice about the target's litigation. Use enterprise tools under contract, with training on your data switched off, rather than consumer services.

Supervision belongs in the same policy. The SRA notice says supervisors remain accountable for work done by others and that firms need governance, systems and controls covering AI risks. A junior analyst who runs a tool and pastes the output into a memo has not completed a review. Name the person who is accountable for each AI-assisted workstream, and require that person to sign off on the sample that was checked.

Terms to read in the tool contract

The vendor's contract decides much of the confidentiality risk. Before approving a tool for deal work, ask for written answers to a short list of questions: whether inputs and outputs are used for training, how long they are retained, who at the vendor can access them, where they are processed, and whether the vendor uses sub-processors. Ask what happens to the data when the contract ends, and whether you can delete a deal workspace on demand.

Also check how the tool is configured. A tool that is acceptable with one set of settings can be unacceptable with another, so record the approved configuration and have IT lock it.

Personal data in the data room

Data rooms contain personal data: employee files, customer lists, director details. In the UK and EU, using an AI tool on that data is processing under data protection law. The ICO's guidance on using AI and personal data treats AI as generally a high-risk technology and advises a data protection impact assessment before processing begins.

The ICO's response to its generative AI consultation also separates data protection accuracy from statistical accuracy, and says the appropriate level of accuracy for a model is linked to the purpose it is used for. Both developers and deployers must consider how outputs will be used. For a deal team, that means accuracy requirements should be higher when an output feeds a decision about a person, such as a key-employee retention plan, than when it feeds a general search.

Minimize what the tool sees. Redact or exclude personal data that the diligence question does not need, and keep employee-level files out of general-purpose workflows. Check where the provider processes data, and whether the processing is covered by the transfer rules that apply to the target's data.

Where to start

You can set up a defensible workflow in a few weeks, before the next live deal:

  1. Write a one-page AI use policy for deals: approved tools, prohibited tools, what may be uploaded, and who signs off.
  2. Add an AI clause to your NDA template review: does it permit processing by approved tools, and under what conditions?
  3. Choose one task, such as change-of-control clause extraction, and test it on a closed deal where the answers are known. Record misses and false positives.
  4. Define the review step: who opens the source, what gets sampled, and what is logged in the deal file.
  5. Brief counsel on the tool and agree which categories of material never go into it.
  6. Run a data protection impact assessment for any workflow that touches personal data.

Start with the test on a closed deal, because it replaces a vendor's claims with your own error rate. Once the miss rate is known, you can decide where the tool saves time and where a person must read every page.

Tags:ai-due-diligencemergers-acquisitionsdata-roomcontract-reviewconfidentialitylegal-privilegemodel-error

Your starting point

What would you like to move forward?

Start with the outcome that feels closest. We'll help give the next step a useful shape.

Let's find your next move

One conversation. A clearer direction.

Tell us what you want to improve. A few sentences are enough to start.

shayan@coldai.org