ChamberOne

AI in due diligence: speed gains and the risks of false comfort

By the ChamberOne team 4 minute read

Picture a data room with several hundred contracts and a timetable measured in days. A junior lawyer once spent a week building the schedule of change-of-control clauses. Today an AI tool can produce a first version in an afternoon. That is a real gain. The new risk is that a finished-looking table feels more reliable than it is, and the partner signs off on a comfort that nobody has earned.

Key points

  • Treat an AI extraction table as a first pass, not a finished review.
  • Sample against hand-review and note where the tool's mistakes cluster.
  • Agree the exceptions route and the lawyer-read categories before the review starts.
  • Give partners clause text and source references, and say what a person has read.

What the tools do well

Due diligence is repetitive by design. The same questions are put to every document: who are the parties, what is the term, can the contract be assigned, is there an exclusivity or a termination right on change of control. AI document analysis fits that shape. It reads each file and fills a row in a table.

Used this way it helps in three places: sorting an unlabelled data room into document types, extracting named fields into a spreadsheet, and drafting a first issue list from the extracted data. The gain is in the first pass. It gets people to the awkward documents sooner.

Where false comfort comes from

A table with every cell filled looks complete. But a blank, or a plausible answer, can mean several things. The clause may be absent. It may be present under an unusual heading. It may sit in a side letter that was uploaded as a separate file. The tool may have misread a scanned page. The row looks the same in each case.

Other sources of comfort that should be tested:

  • Files that failed to open or were skipped, which may never appear as an exception.
  • Poor scans and handwritten amendments.
  • Defined terms. A clause may say 'Restricted Party' and the definition, two schedules away, decides what that means.
  • Documents that refer to other documents not in the room.
  • Foreign-language contracts, where a machine reading may shift meaning.
What the table shows
  • Every cell filled
  • A blank or plausible answer in each row
  • Rows that look the same in every case
What may lie behind it
  • The clause is absent
  • The clause sits under an unusual heading
  • The clause is in a side letter uploaded as a separate file
  • A scanned page was misread
What a full table shows and what it may hide

Sampling before you trust the extraction

Before relying on any extraction run, test it. Choose a sample across document types, including the oldest, the longest and the worst scans. Have a lawyer read each one in full and fill in the same table by hand. Compare.

You are not looking for a pass mark. You are looking for the pattern of mistakes. If the tool is reliable on dates and parties but weak on assignment clauses in amended agreements, you know where human reading must be concentrated. Record the sample size and what you found. Keep it on the file; it shows later how the review was run.

Run a second sample once the issue list stabilises. Late-appearing issues often reveal that an earlier question was framed too narrowly.

Design the exceptions route first

The most useful feature of a review process is an agreed route for things that do not fit. Decide before you start who sees what.

  1. 1

    Every skipped, unreadable or empty-result file goes on an exceptions list that a person works through.

  2. 2

    Any document where the tool states low confidence, or where the answer is 'not found' on a key question, is read by a lawyer.

  3. 3

    Any document that touches the top issues for this deal, as agreed with the partner, is read by a lawyer regardless of the tool's output.

  4. 4

    Anything affecting price, conditions or warranties is escalated to the partner with the source passage, not a summary.

What to escalate to a partner

Partners do not want a list of two hundred observations. They want the five that change the deal. A useful report separates findings into those that affect valuation or structure, those that need a negotiated fix such as a waiver or consent, those that can be handled in the disclosure letter, and routine points for the record.

Whatever the category, give the partner the clause text, the document reference and a sentence on why it matters. If an AI tool suggested the issue, say so, and say whether a lawyer has read the clause. That is not an admission of weakness. It tells the partner how much weight to put on each line.

Telling the client what you did

Clients increasingly ask whether AI was used and whether the fee reflects it. Be ready to explain the method in plain terms: what the tool did, what lawyers read in full, and what the report does not cover. Check the guidance from your regulator or bar on disclosure and on supervising technology, and check your engagement terms and the data-room provider's terms before uploading confidential material to any third-party AI service.

Avoid language such as 'full review' unless a human has in fact read the full set. 'Reviewed with software assistance, with lawyer review of priority and exception documents' is more accurate and protects you if something later turns up.

A hypothetical afternoon in the data room

Suppose a three-lawyer team is acting for a buyer, with two days to produce a first report. They upload the supplier contracts and ask the tool for term, assignment and termination rights. By lunchtime they have a table of ninety rows.

The first thing they do is sort the table by missing answers. Eleven rows are blank. Opening the files, they find four scans too faint to read, two documents that are amendments to agreements already in the table, and five that really contain no assignment clause. Without that sort, the blanks would have read as 'no restriction', which is the opposite of what the team needed to know.

That small habit, sorting by what the tool could not answer, is worth more than any feature. Make it part of the routine for every run.

Keeping the review honest over time

Revisit the process at the end of every deal. Which categories of document produced the most errors in the tool's output? Which questions did the partner ask that the table could not answer? Write the answers into the team's checklist so the next deal starts from a better place.

Document analysis in ChamberOne

  • Upload a contract and it lists the document type, parties, key dates, clauses and citations, each linked to the paragraph it came from.
  • Risk flags on clauses are suggestions that a partner can override.
  • Accepts PDF, Word, Excel, CSV, HTML, text and markdown files.

What it does not do

  • It does not replace a lawyer reading the document; the analysis can miss or misread things.
  • It does not give legal advice or decide what to escalate.
  • Document text is sent to the AI provider chosen for the feature, so check the terms first.

This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.

Read next

AI in due diligence: speed gains and false comfort