The AI Summary’s Silent Omission Is an Old Blind Spot Running Faster

What the vendors are actually selling

The pitch in the data room right now is speed with a conscience. Platforms like Drooms market document summarization, risk identification, numerical analysis, and source-cited answers as the core of the product¹².

The claim underneath is throughput. One due-diligence review piece describes AI summarising and comparing thousands of files in minutes, compressing work that used to take a team days³.

I have watched a summary like that land on my desk and read as finished. Clean headings, a tidy risk list, a citation after each line. Nothing about it invites a second pass.

The risk isn’t the invention – it’s the thing that was quietly left out

The failure most people brace for is the confident fabrication, the plausible answer that is simply wrong⁴. That is the one everybody screenshots.

The more dangerous one is quieter. A 2026 deep-dive on Harvey’s contract-review AI describes edits that pass through without a flag even though the output looks complete, a pattern the authors call a quiet failure mode and sum up as omission, not hallucination.

That is the failure that survives a skim, because a summary with a gap in it looks exactly like a summary without one.

The cost is not in the summary – it is at close

A missed clause is cheap to fix while the deal is open and costly once it has closed. A 2026 data-room review puts it plainly: a red flag found after close is far more expensive than one caught in diligence⁵. The examples it logs are the ones that hide well:

1. a change-of-control clause,
2. an unassigned patent,
3. customer concentration buried in a schedule⁵.

None of those announce themselves. Each is the kind of line a tired reader’s eye slides past.

The blind spot predates the model

Here is the part the AI framing tends to skip: people have always done this. A contract-analysis evaluation from Sirion states it directly – human reviewers still miss critical clauses under deadline pressure, which makes omission a baseline problem common to both human and AI review⁶.

A 2026 analysis of traditional document review goes further, describing a systematic bias toward errors of omission under load: reviewers miss clauses and approve items they would have caught at the start of the session. The sharp point is that these misses are not auditable failures, because the task still looks completed.

That last phrase is the whole problem, and it was true of humans long before it was true of software.

The stakes were already large before any of this

Bain’s global M&A research found that almost 60% of executives attributed deal failure to poor due diligence that did not identify critical issues⁷. That figure comes from a 2020 dataset, well before the tools discussed here existed⁸ – which is the point. The gap it measures is the human one.

Measured AI summarisation sits close to that human line, not off the chart

When the omission rate is actually measured, it does not look alien. On the ContractEval benchmark using the CUAD contract dataset, several open-source models show false “no related clause” rates above 25 to 30% – answers that wrongly assert no relevant clause exists⁹. That is omission at the clause level, quantified.

And on the summarization task itself:

77.2% · CoCounsel’s document-summarization accuracy on the Vals Legal AI Report benchmark, within three points of the 79.7% baseline set by practising lawyers on the same task (Vals Legal AI Report)

A trade analysis of that benchmark draws the honest conclusion: the tool rarely invents new facts, but it might leave out an important detail, and it is safe only if someone checks the summary against the original file.

It is a machine committing a familiar error. It is a familiar miss, benchmarked.

What Changed Was the Hedge

For decades, professional review carried its own scepticism on the page. Legal due-diligence precedents from the ISLP instruct authors to watermark preliminary work “Draft and Privileged” and to caveat any checklist as a guide, “not a definitive list”¹⁰.

Other guidance is built the same way. Firms are told to specify the scope’s limitations, state assumptions, and name sources; project-finance conventions add explicit lines that the firm reviewed only the material provided and made no independent investigation¹¹¹².

Those caveats were not decoration. They were the reader’s signal that the review had edges.

The generic disclaimer does not carry the same information

The AI version of that hedge has thinned out. GetAccept describes the usual answer from most tools as a small-print line – “AI can make mistakes. Please verify.” – though its own product attaches a confidence score to each extracted field¹³. A data-room vendor’s 2026 disclaimer says outputs may be incomplete or inaccurate and must be checked against source documents and professional advice¹⁴.

That is true, and it is generic. It applies to every output equally, which means it flags no output in particular. The old caveat told you where the review stopped; the new one tells you only that reviews can stop somewhere.

I should be straight about a gap here: I found no documented case that ties a disclosed financial consequence to an AI-specific omission. The mechanism is well described; the priced example is not yet on the record.

The strongest case for calm

The opposing argument is good, and it deserves its full weight. Vendors say the fix is built in: citations let a reviewer jump straight to the source passage, and if the system is confined to drafting while a person keeps final control, the exposure drops sharply¹⁵¹⁶.

The controls are shipping, too:

1. Harvey’s Agent Builder is designed with human-in-the-loop checkpoints as a core feature;
2. Luminance frames its system as “autonomous by default, human in the loop by choice”;
3. CMS’s rollout of Harvey sets defined thresholds where human review is mandatory, such as for court filings, legal opinions, or novel strategy.

The habit data backs it. Just 7% of dealmakers say they would act on an AI recommendation without human review¹⁷, and in Morae’s survey 89% of legal professionals believe AI work should be checked before use.

But the gap is not closed, and I won’t pretend it is

The same research that shows near-universal review habits also shows a shortfall. Around 58% of dealmakers say they apply human validation to AI due-diligence outputs – which leaves roughly 42% who, by that figure’s own logic, may not¹⁸.

And there is a gap I can’t fill: I found no vendor-neutral audit measuring how reliably these confidence flags and “not reviewed” markers are actually present, rather than silently absent. The control is asserted more often than it is measured.

The friction was always the point

Verifying a data-room summary by hand is not new work that AI has forced on us. It is the old friction the interface talked us out of doing – the second pass that a “Draft and Privileged” watermark used to prompt on its own.

The honest limit of my own argument: no reported case yet compares an AI omission’s outcome against a matched human miss. I am arguing that the failure mode is the same and faster, not that anyone has priced the difference.

The gap between 77.2% and 79.7% is less important than the real question: whether the reader treats the clean summary as a first pass or a verdict.

This is observation from my own desk, not advice. Nothing here is a recommendation to buy, sell, hold, or act on any deal or tool; the decision, and the verification, stay with you.