Blog№ 001

What AI can and cannot do in a data room

Vendors promise diligence in minutes. The published evidence says something narrower and more useful: AI is reliable when it is pointed at the right page, and a person still has to decide.

Most claims about AI in diligence come from companies selling it. The one you will see most often is that firms cut financial diligence time by 60 to 70%. We traced it. It comes from a vendor's web page with no source, no sample and no method.

There is better evidence, and it is more useful than the marketing. It says AI is reliable at some diligence work, unreliable at other work, and that the difference comes down to one thing.

The size of the job

A study by Bayes Business School and SS&C Intralinks looked at more than 900 deals between 2013 and 2023. The data room for a private company held 7,583 files on average. Diligence took 234 days.

Nobody reads 7,583 files. Analysts read the ones they think matter and sample the rest. That is the real case for software: it reads all of them.

Adoption has followed. Deloitte surveyed 1,000 senior US corporate and private equity leaders in 2025. Of those, 86% said they had brought generative AI into their M&A work. Their top two worries were data security, at 67%, and data quality, at 65%.

What it does well depends on where you point it

FinanceBench is a 2023 test of how well language models answer questions about public company filings. One result from it explains most of what you need to know.

Figure 1 · One model, the same 150 questions about company filings
Left to search every filing itself19% correct

13% answered wrong. 68% refused to answer.

Given the right page85% correct

15% answered wrong. None refused.

The model did not get smarter between the two bars. In the second run it was handed the page the answer was on. Most of the failure in the first run was refusal: the system could not find the page, so it would not answer.Islam et al., FinanceBench, 2023. GPT-4-Turbo, 150 questions per configuration.

Same model. Same questions. The difference between 19% and 85% is whether the system put the right page in front of the model.

So the dependable work is the work where the page is known:

  • Indexing. Listing every file by folder, page and table, and naming the files it could not read.
  • Extraction. Pulling a figure from a page it has been pointed to.
  • Tracing. Taking a number from the deck and finding the page in the data room that supports it, or does not.
Figure 2 · One number from the deck, traced to its source
Pitch deck · slide 9

Annual recurring revenue

$4.2M

Data room · Financials / MRR schedule · page 14

December MRR $325,000, annualized

$3.9M

Flag for the partner. Deck states $4.2M. Source supports $3.9M. Linked to page 14.

The deck and the data room disagree by $300,000. Software found the page. Whether the gap is a rounding choice, a timing difference or a problem is a question for the partner.

Tracing matters more than it sounds. When JPMorgan bought the student aid startup Frank in 2021 for $175 million, the founder had represented 4.25 million users. The Department of Justice later established that Frank had about 300,000. She was convicted of fraud in 2025. A claimed number and a real number sat a long way apart, and checking one against the other is exactly the kind of work software does without getting tired.

Where it fails

Open questions over many documents. A 2026 benchmark built on synthetic private equity fund spreadsheets found the best model reached 82.4% accuracy, about one error in every six questions. On the largest spreadsheet accuracy averaged 48.6%. The authors, who work at a firm in this field, concluded that no model on its own had an error rate low enough for unsupervised use.

Legal reasoning. Stanford researchers tested the leading AI legal research tools in 2024 on more than 200 questions written to probe their weak points. Each tool gave a wrong or unsupported answer between 17% and 33% of the time. These were purpose-built products from major legal publishers, not general chatbots. The test covered legal research rather than document extraction, so do not carry the figure across directly. Carry the lesson: a confident answer with a citation next to it can still be wrong.

Judgment. No benchmark measures whether the growth is durable, whether the founder is telling you everything, or whether the price is right. Those are the questions a fund is paid to answer.

Figure 3 · Where each diligence task sits
Software can do it and show its workA person has to decide
Index every file by folder, page and table
Pull a figure from a page it has been pointed to
Trace a number in the deck to its source page
List what a 90-page agreement says about change of control
Say whether the growth is durable
Decide to invest
This is our judgment, drawn from the studies above and from building these workflows. The further right a task sits, the more of it has to stay with a named person.

The rule that follows

Every answer should arrive with the page it came from. If the software cannot show its source, treat the answer as a guess.

Every conclusion should have a name on it. The software drafts the memo. An analyst edits it. A partner approves it before it reaches the investment committee.

This is also where regulators are looking. In March 2024 the SEC fined two investment advisers, Delphia and Global Predictions, a combined $400,000 for claiming to use AI in ways they did not. A fund that tells its investors AI is doing its diligence should be able to show what the AI did and who checked it.

Where software fits

This is how dailybuilt handles a fund deal. It indexes the data room by folder, page and table when access is granted, and lists any file it could not read. It traces every figure in the deck to its source and ranks the gaps for the partner, with each flag linked to the page it came from. It drafts the memo with every claim cited. Nothing reaches the investment committee until a named person approves it.

Sources