AI summaries are now a routine part of reviewing a data room. A reviewer opens a 60-page supply agreement, asks for a summary and gets a tidy page of bullet points in seconds. For triage, that is a real gain. The risk is that the summary quietly becomes the review, and a bullet point that sounds right is copied into a diligence report without anyone opening the contract.
This guide is about the step in between: how to check a summary quickly enough that the time saving survives, but carefully enough that the report stays accurate.
Why do AI summaries of deal documents need checking?
Large language models write fluent text by predicting likely wording. Most of the time the result matches the source. When it does not, the error is presented with the same confidence as everything else, which is what makes it dangerous in diligence.
Deal documents are also an awkward fit for summarization. Their meaning often sits in exceptions, cross-references and definitions spread across the document. A summary that compresses “the supplier may terminate on 90 days’ notice, except during the first contract year and subject to clause 14.3” into “either party may terminate on 90 days’ notice” has lost exactly the part a buyer needs.
| Failure | What it looks like | Why it matters in a deal |
|---|---|---|
| Wrong number | A cap of $2.5m reported as $25m, or a term of 36 months reported as three years from signing when it runs from go-live | Valuation, liability and timing all rest on numbers |
| Lost qualification | ”Except”, “subject to” or “unless” dropped from a sentence | Turns a limited right into an absolute one |
| Misread defined term | ”Affiliate” or “Change of Control” read in its everyday meaning instead of the contract’s definition | Changes who is bound and what triggers a clause |
| Omission | A side letter, schedule or amendment not mentioned | The summary looks complete but is not |
| Blended sources | Details from two similar contracts merged into one summary | Assigns the wrong terms to the wrong counterparty |
| Invented detail | A clause, date or party that does not appear in the source | Rare but severe, and hard to spot without the source open |
What does a five-minute verification routine look like?
The routine below assumes the summary includes citations or page references, which most data room AI features aim to provide. If yours does not, the checks still work, but step one takes longer.
A five-minute check of one AI summary
- Open every cited passage
Click through each citation and confirm the passage says what the bullet says. A citation that points to the wrong page is a warning about the whole summary.
- Re-read every number against the source
Amounts, percentages, dates, notice periods and thresholds. Check units and currency, and check from when a period runs.
- Look up each defined term
Capitalized words in the summary usually have a contract definition. Read it. Pay most attention to definitions that limit or extend scope.
- Scan the source structure
Look at the table of contents or headings and ask whether any clause, schedule or annex is missing from the summary. Termination, liability, assignment and exclusivity are the usual gaps.
- Check for later documents
Search the folder for amendments, side letters and renewals. The summary may describe the original contract only.
Five minutes per document sounds slow next to a two-second summary. Compared with a full read of a 60-page contract, it is still a large saving, and it is the step that makes the saving safe.
Which parts of a summary deserve the most suspicion?
Not every bullet carries the same risk. Descriptive points such as the names of the parties, the governing law or the general subject of the contract are rarely wrong and rarely decisive. Spend the checking time on the points that drive the report.
| Summary element | Risk of error | Check |
|---|---|---|
| Parties and date | Low | Glance at the first page |
| Term and renewal | Medium | Confirm start trigger, length and renewal mechanics |
| Payment terms and amounts | High | Re-read every figure and its unit |
| Termination rights | High | Look for exceptions, notice periods and cure periods |
| Liability caps and indemnities | High | Read the full clause, including carve-outs |
| Change of control and assignment | High | Read the clause and the definition of control |
| Exclusivity and non-compete | High | Check territory, duration and scope |
| Confidentiality | Medium | Check survival period and permitted disclosures |
How should checked summaries be recorded?
The record matters because diligence work is reviewed, sometimes years later. A reviewer who relied on an AI summary should be able to show which parts were verified and against what.
A practical approach is to keep the AI output and the human check side by side. Paste the summary into the review notes, mark each high-risk bullet as confirmed, corrected or not checked, and add the page reference. If a bullet was wrong, keep the correction visible rather than overwriting it; patterns of error tell the team where the feature can and cannot be trusted.
Diligence reports should cite the source document and page, never “per AI summary”. If the report is shared with lenders or insurers, they will expect a human to have read the clauses it relies on.
Which data room platforms list AI assistance?
In our facts file, five of the 18 platforms list AI assistance: Ellty, Datasite, Ansarada, Box and Drooms. The facts file records whether AI features are listed, not how good their summaries are or how they cite sources, so the routine in this guide applies to all of them.
AI assistance across the 18 platforms
Ellty, Datasite, Ansarada, Box, Drooms
Average out of 10 for platforms that list AI
Average out of 10 for the other platforms
When evaluating a platform’s summaries, test them on your own document types. Upload three real contracts of the kind your deal involves, including one with a side letter and one scanned copy, and check the summaries against the routine above. Look at whether citations land on the right page, whether the tool says when it is unsure, and whether it respects folder permissions so that a reviewer cannot get a summary of a file they cannot open. The AI data room software hub lists the AI capabilities recorded for each platform and the questions to put to vendors.
Questions to ask about any data room AI summary feature
- Does each summary point cite a page or passage in the source?
- Does the summary respect the reviewer's folder and document permissions?
- Can it summarize scanned documents, and does it say when text quality is poor?
- Is customer data used to train the model, and where is it processed?
- Can summaries and questions be exported with the review record?
When should you skip the summary and read the document?
Some documents should be read in full regardless of what AI produces. The share purchase agreement drafts, the disclosure letter, key customer and supplier contracts above your materiality threshold, financing documents and any document already flagged in an earlier round all fall into this group. For those, use AI to navigate, for example to find every mention of a counterparty, and do the reading yourself.
Summaries earn their keep on the long tail: dozens of similar leases, standard employment contracts, routine supplier terms. There, a checked summary plus a sample read of a few full documents is a reasonable and defensible approach. The due diligence shortlist weights review features such as search and AI more heavily than the overall ranking.
Frequently asked questions
Can AI summaries be used in a due diligence report?
As a starting point, yes. Every point the report relies on should be verified against the source document and cited by page, not by the summary.
How long does it take to check an AI summary?
About five minutes for a typical commercial contract using a routine focused on citations, numbers, defined terms and omissions. High-value documents should still be read in full.
Do AI summaries see documents I am not allowed to open?
They should not. Ask each vendor to confirm that AI answers and summaries respect the same folder permissions as the documents, and test it with a restricted user.
Why do AI summaries miss side letters?
A summary usually covers the single document it was asked about. Amendments and side letters are separate files, so search the folder for them and summarize or read them separately.