All articles
    Deal Technology
    8 min read••Vetting Vault Team

    Why AI Can't Read Your Data Room (and What Fixes It)

    AI is only as good as the context around each file. See why folder-based data rooms leave it guessing, and how request-first structure gives it something to work with.

    Every deal team is being told that AI will speed up due diligence. Point a model at the data room, ask it whether revenue ties to the bank statements, and get an answer in seconds. Then someone tries it on a real data room and the answer comes back vague, wrong, or confidently built on last year's P&L.

    The model usually isn't the problem. Modern AI reads a 60-page lease or a year of bank statements without trouble. What it can't do is work out what each file is supposed to be. And a folder-based data room, which is what most AI data room features are bolted onto, throws that information away at the moment a document is uploaded.

    What AI needs to do due diligence work

    Reading a document and doing diligence on it are different jobs. To answer "does 2025 revenue tie to the bank deposits?", an AI has to know several things before it reads a single number. It needs to know which file is the 2025 P&L, and that it's the version the seller stands behind rather than a draft. It needs to know which files are bank statements, which months each one covers, and which account they belong to if the business banks in more than one place. And it needs to know whether all twelve months are actually there, or whether three are still outstanding.

    None of that lives inside the documents. It lives in the diligence request list: "Monthly P&L, January through December 2025", "Bank statements, all operating accounts, last 24 months". That's the context that makes a file mean something. The question is whether the system holding the files still has it.

    Animated · Illustrative example
    Same files, same question. Only one AI can answer it.

    An AI reads each file, then gets asked to reconcile revenue. On the left the files sit in folders; on the right each one arrived against the request it answers.

    Folder-based data room
    Financials_FINAL_v3.pdf
    scan0042.pdf
    Q3_bank_stmt_revised.pdf
    From Bob - email attachments.zip
    IMG_2231.jpg
    Request-first
    Financials_FINAL_v3.pdf
    scan0042.pdf
    Q3_bank_stmt_revised.pdf
    From Bob - email attachments.zip
    IMG_2231.jpg
    Does 2025 revenue tie to the bank deposits?
    Low confidence
    Waiting for the question…
    High confidence
    Waiting for the question…
    Show as a table
    FileWhat a folder tells the AIWhat the request tells the AI
    Financials_FINAL_v3.pdfLooks like a P&L. Which year? Is v3 the final one?1.2 Monthly P&L, Jan–Dec 2025
    scan0042.pdfA scanned tax form. No year in the name.1.1 Tax returns, last 3 years
    Q3_bank_stmt_revised.pdfA bank statement for part of a year. Revised from what?1.3 Bank statements, Jan–Dec 2025
    From Bob - email attachments.zipAn archive with no description. Could be anything.1.3 Bank statements, Jan–Dec 2025
    IMG_2231.jpgA phone photo of a table of names and amounts.1.4 AR aging, current
    The extracted text is identical on both sides. What differs is the context around each file: what it was asked for as, which period it should cover, and what's still missing. Without that, the AI has to guess, and a careful one will say it can't tell.

    Why folder-based data rooms leave AI guessing

    A traditional virtual data room is a file system with permissions. Somebody on the sell side builds a folder tree (01 Financials, 02 Legal, 03 HR), documents get dropped into it, and a separate spreadsheet tracks what was asked for and what has arrived. That works for people, because people carry the context in their heads and in their inbox. An AI reading the data room sees only what's on disk, and what's on disk was never designed to explain itself. Four gaps show up in almost every folder-based room:

    File names written for nobody
    "Financials_FINAL_v3.pdf", "scan0042.pdf", "IMG_2231.jpg". The name is often the only label a file has, and it's usually ambiguous or blank.
    Versions without a winner
    Two copies of the P&L with different numbers on page two, and nothing to say which one the seller stands behind.
    Folders that say where a file landed
    A bank statement filed under Legal is still a bank statement, but now nothing connects it to the financial requests it answers.
    No idea what's missing
    The tracker that says "bank statements: 9 of 12 received" lives in someone's email. The AI reconciles nine months and calls it a year.
    The quiet failure is the dangerous one
    An AI that says "I can't tell" is annoying. An AI that picks the wrong P&L and reconciles it against a partial year, with a confident summary on top, is worse. Missing context doesn't always produce an obvious error. It often produces a plausible one.

    Adding AI to a folder-based room can help with search and summaries of single documents. But cross-document work, the part of diligence that takes the hours, depends on knowing how the files relate to each other and to the questions the buyer asked. That information was never captured.

    Side by side
    What an AI knows about one file

    The text is the same either way. Everything around it comes from how the file was collected.

    Context
    Folder-based
    Request-first
    The file's text
    Extracted
    Extracted
    What it was asked for as
    Only a file name
    The request it answers
    The period it should cover
    If the name says
    In the request: Jan–Dec 2025
    Which section of the deal
    The folder, if filed right
    Section and project
    Whether the item is complete
    In a separate tracker
    Request status
    What's still missing
    Invisible
    Open requests

    Request-first: every file arrives with its question attached

    The alternative is to stop treating the request list and the data room as two things. In a request-first data room the request list is the structure. The seller doesn't file documents into folders; they answer requests. Whatever they upload is attached to the request it answers.

    That single change gives every document context a folder can't. A phone photo called IMG_2231.jpg is now known to be the current AR aging, because that's the request it was uploaded against. The period it should cover is written into the request itself. It belongs to a section and a workstream. And the request carries a status (open, submitted or complete) and an owner, so the system knows not just what has arrived but what hasn't.

    Once files are tagged that way, extracting their text produces something closer to a structured knowledge base than a pile of PDFs. The system knows that these twelve files are the 2025 bank statements and this one is the P&L, so it can line them up. It also knows which requests are still open, which means an AI can say "September's statement hasn't arrived" instead of silently reconciling eleven months.

    This is also why request-first collection is easier for the other side of the table. A first-time seller doesn't need to learn a folder taxonomy. They see a list of plain-English requests and answer them one at a time. The same structure that helps the counterparty is the structure that makes the AI useful.

    What AI can do once it knows what each file is

    With that structure in place, the useful work isn't summarizing one document. Single-document summaries save a few minutes of reading. The hours in diligence go into checking documents against each other, the way an analyst would, and that's where knowing which file is which pays off:

    Revenue vs. deposits
    Monthly P&L revenue against deposits on the bank statements for the same months.
    Tax returns vs. books
    Gross receipts on the return against revenue on the P&L for the same year, which surfaces cash-basis and timing differences.
    AR aging vs. the balance sheet
    Does the aging total match receivables on the balance sheet at the same date?
    Completeness
    Are all the months in a monthly series actually there, or is one quietly missing?

    The same structure makes it practical to pull data out of documents into a workbook: resumes into a key-employee table, leases into a rent schedule, twelve statements into one month-by-month deposit summary.

    Illustrative example
    Reported revenue vs. bank deposits: a $210K gap, mostly in September

    Monthly revenue from the 2025 P&L next to deposits from the twelve monthly bank statements.

    P&L revenueBank deposits
    Show the numbers
    MonthP&L revenueBank depositsDifference
    Jan$360K$355K−$5K
    Feb$372K$372K$0K
    Mar$401K$401K$0K
    Apr$418K$418K$0K
    May$430K$430K$0K
    Jun$425K$420K−$5K
    Jul$410K$410K$0K
    Aug$415K$415K$0K
    Sep$404K$214K−$190K
    Oct$398K$398K$0K
    Nov$390K$380K−$10K
    Dec$397K$397K$0K
    Total$4.82M$4.61M−$210K
    A gap isn't proof of anything. It can be a customer paying late, a deposit split across accounts, or revenue booked that never arrived. The point is that someone has to ask, and the check is only possible once the system knows which file is September's statement and which is the P&L.

    Notice what the reconciliation above doesn't do: it doesn't decide what the September gap means. A customer may have paid late, a deposit may have gone to a second account, or revenue may have been booked that never arrived. AI is good at finding the question and pointing to the two files that raise it. A person still has to ask the seller.

    Where AI still needs a person

    Structure fixes the context problem. It doesn't make AI output trustworthy by default, and nobody signs a purchase agreement on the strength of a chat answer. The teams that get real value from AI in due diligence treat it like a fast, tireless junior analyst: useful for finding things and assembling them, never the last word. Three habits follow from that:

    Demand sources
    Every figure should point back to the file it came from so a reviewer can check it. An answer without a citation is a lead, not a finding.
    Watch the confidence
    Scans, photos and unusual formats extract imperfectly. A good system says when it's unsure rather than smoothing over it.
    Keep judgment with the team
    Whether an add-back is legitimate, or a gap is a deal issue, is a professional call. AI gets you to that call faster.

    How this works in Vetting Vault

    Vetting Vault is built request-first, so every file is attached to the request it answers and its text is extracted and searchable. That's the structure the rest of this article argues for, and it's what the AI works from. There are two ways to put it to work:

    Deal AI, inside the deal
    Deal AI sees the request list with each request's status and attached files. It can combine P&Ls, reconcile revenue or build a table from bank statements, and every answer cites its files with a High, Medium or Low confidence. It can suggest changes, but nothing changes until you click Apply.
    AI Connect, for your own AI
    Connect Claude, ChatGPT or Grok with AI Connect and it does the work, not just reads: builds a request list from a document, chases open items, moves files. It uses your permissions, asks before anything risky, and logs every change under your name.

    Questions to ask about any AI data room

    Nearly every data room vendor now advertises AI, and the demos all look alike: a document goes in and a tidy summary comes out. The difference shows up on a real deal, with a hundred files and a request list that's two-thirds done. Whatever platform you use, these questions separate AI that can do diligence from AI that can only read:

    1. Does the system know what each file was requested as, or only its name and folder?
    2. Can the AI tell what's still missing, or only what's been uploaded?
    3. Does every answer cite the files it used, so you can check them?
    4. Does it say when it's unsure?
    5. When it takes an action, does it use your permissions and record what it did?

    If the answer to the first two is no, the AI is working without the context a junior analyst would get on day one. For more on how the two models differ, see What is a Data Room?

    Frequently asked questions

    Can AI replace a quality of earnings review?

    No. AI can pull figures together, reconcile them across documents and flag gaps much faster than a person can, which shortens the work. Deciding whether an adjustment is legitimate, and putting a name behind the conclusion, is still the job of the accountant or deal team.

    Does AI work on scanned documents and photos?

    Only as well as the text that can be read from them. Scans and phone photos have to be converted to text first, and the quality varies. That's one more reason an AI should cite the file it used: a person can open the source and check the number.

    What is a request-first data room?

    One where the diligence request list is the structure. Each document is uploaded against the request it answers, so the system knows what every file is supposed to be, which period it should cover, and which requests are still open.

    Is it safe to use AI on confidential deal documents?

    Ask the vendor where documents are processed, who can see AI output, and whether AI can be turned off for a deal. An AI that works inside the deal's own permissions shows each person only what they could already open.

    Try AI on a data room that knows what every file is

    Sign up free and open the sample deal: a request list with real-looking documents already attached. Ask Deal AI to reconcile revenue, check what's missing, and see which file every answer comes from.

    Try the Sample Deal