The Data Room Reads Itself Now: AI in Due Diligence, and What It Still Cannot See

The data room used to be a reading problem. Four thousand documents, six weeks, three junior analysts, and a partner who wanted the summary on Friday. Whatever got missed got missed, and the reps and warranties were there to cover the difference.

That constraint has largely dissolved. Machine review can now ingest an entire data room, extract contract terms across hundreds of agreements, flag non-standard clauses, reconcile revenue schedules against customer contracts, and produce a structured summary in hours rather than weeks. On the tasks it does well, it does them better than a tired associate at eleven at night, and it does them consistently.

The risk that has emerged is subtler than error. It is displacement. A comprehensive-looking machine output produces a sensation of thoroughness, and that sensation reallocates human attention away from the parts of diligence that were always the hard parts. The findings that kill deals have rarely been in the documents. They have been in the gaps between them.

What machine review is genuinely good at

Four categories, and they are not trivial.

Extraction at scale. Pulling every change of control provision, every exclusivity clause, every auto-renewal date, every price escalator across four hundred customer contracts. A human team does this by sampling. A machine does it exhaustively. Exhaustive beats sampled, and the difference is not marginal on a business with concentrated contract risk.

Comparison against a standard. Give the system your firm's position on twenty commercial terms and it will tell you where this target's contracts deviate. This is tedious work performed perfectly, which is the ideal machine task.

Completeness checking. Cross-referencing the customer list in the CIM against the billing extract against the contract folder, and surfacing the customers that appear in one and not the others. Humans are poor at this because it requires holding three lists in working memory. Machines are excellent at it.

Timeline construction. Assembling a chronology from board minutes, filings, and correspondence, which is slow manual work that produces disproportionate insight once assembled.

The four blind spots

One. The absent document

A model reasons about what it was given. It does not reliably reason about what it was not given, because absence has no representation in the input.

The experienced diligence lead notices that there are no board minutes for two quarters in 2024. That there is no written agreement with the second largest supplier. That the customer contract folder contains a hundred and forty agreements for a hundred and fifty-two named customers. That every reference the seller offered is a customer and none is a former employee.

These absences are the most reliable predictor of a problem in the entire diligence process, and they are precisely what a completeness-scored machine output obscures. The system reports on the room it was shown. The question that matters is what was not put in it.

Two. Contradiction that requires domain judgment to see

Machines are improving at flagging direct numerical contradictions, and that is useful. What they miss is the contradiction that only reads as a contradiction if you know the industry.

Headcount in the service delivery function fell eleven percent while revenue grew nine percent, and the CIM presents this as operating leverage. To a model, both facts are consistent with the narrative. To someone who has run a service business, the combination raises an immediate question about whether delivery quality has been quietly degraded and whether the churn shows up eighteen months out, after close.

That is not a data problem. It is a pattern held in a practitioner's head, and there is no clean way to hand it to a system that has not carried the number itself. The related failure inside a portfolio company, where sources disagree and get averaged rather than escalated, is the same structural problem examined in the single source of truth audit.

Three. What people say rather than write

The most valuable half hour in most diligence processes is unstructured, undocumented, and happens away from the data room. A supplier who mentions in passing that the payment terms were renegotiated twice last year. A departing controller who is careful in an oddly specific way. A customer who is complimentary about the product and evasive about renewal.

None of this is in the room, so none of it is in the model. It also does not survive being written down, because the moment it becomes a documented finding it becomes a negotiating position, and people stop offering it.

Four. What the numbers mean

A model can tell you that gross retention is 86 percent. Whether 86 percent is good depends on the category, the contract structure, the customer size distribution, what the competitors run at, and whether the business has been deliberately shedding unprofitable accounts. That context is judgment, and the reason it matters is set out in customer concentration theatre.

The displacement risk, stated plainly

Here is the sequence that concerns me, and it is already visible.

A fund deploys AI diligence and the review cycle shortens from six weeks to three. The efficiency is real and gets celebrated. In the next deal, the timeline is compressed to match the new capability, because that is what efficiency gains are for. The three weeks saved do not get reinvested in supplier calls and former-employee references. They get given back to the process.

The team now has a more complete read of the documents and less time in the field. Since roughly seventy percent of material findings originate outside the documents, total diligence quality has fallen while every visible metric has improved.

The correct response is not to slow down. It is to be explicit that the hours released by machine review are earmarked, before the deal starts, for the work that machines cannot do. Write it into the diligence plan as a line item. Otherwise the saving is simply taken by the calendar.

Four questions for the next diligence kickoff

What has the seller not given us, and have we asked for it by name rather than by category.

Of the hours machine review saves us on this deal, which specific field activities are they funding, and who owns them.

Which findings in the machine output would we have reached anyway, and which are genuinely incremental. The second number is the honest measure of what the tool contributed.

Who on this team has operated a business in this category, and have they been in the room, or have they been sent a summary. A model output read by someone without domain pattern recognition is a completeness report, not a diligence finding. The wider case for running an operating read before signature rather than after is the pre-close operating veto.

The version of this that works

The funds getting real value from AI diligence are not the ones running the biggest models. They are the ones that redrew the boundary deliberately: machines take extraction, comparison, reconciliation, and completeness, and every hour that releases goes into absence-hunting, third-party conversation, and operating diagnosis.

That is a genuinely better diligence process than existed three years ago. It is better because the humans stopped doing work that machines do properly, not because the machines started doing work that humans do properly. Keep the boundary in the right place and the technology is a straightforward gain. Let it drift and you will run faster processes with thinner conviction, and you will find out which one mattered somewhere around month fourteen.

VCI Institute in collaboration with Mohamad Chahine
Published 22 October 2026

About the VCI Institute

The VCI Institute is a nonprofit dedicated to building practical capability and shared standards for value creation in private equity. The Institute publishes operator-grade frameworks and runs certification programmes for operating partners, portfolio company executives, and value creation analysts. You can see what each programme actually covers before deciding. Analysis published here draws on the Institute's certification curricula and on structured review of mid-market transaction patterns rather than on any single proprietary dataset. Where a figure is directional rather than measured, it is described as such.

Further material is available in the Institute's Insights library and its free resource library of templates, checklists, and case snapshots.

© 2026 VCI Institute. All rights reserved. The frameworks, terminology, and analysis presented in this article are the intellectual property of the VCI Institute. Reproduction or derivative use without written permission is prohibited. Citation with proper attribution is welcomed.