BRENNEN

Industries / Finance and business services / Finance and banking / Private equity and venture capital

Due diligence

Due diligence is the check an investor runs on a company before buying into it: its accounts, contracts, customers, technology, legal risks, and whether the deal needs a government's approval.

7

Milestones achieved, from 1933 to 2025.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1933

    Liable unless they checked

    Section 11 of the US Securities Act of 1933 lets buyers of shares sue over a false or missing material fact in a registration statement, the document filed to sell shares to the public. It names who can be sued: everyone who signed it, the directors, the accountants and other experts who certified parts of it, and every underwriter. Most can escape liability only by showing that they "had, after reasonable investigation, reasonable ground to believe and did believe" the statement was true.

    law.cornell.edu

    Credit

    the US Congress

  2. 2006

    Data rooms move online

    In a physical data room, the seller's documents sat in one room that each bidder's team visited at an allotted time. By 2006 CFO magazine reported companies running diligence instead through virtual data rooms hosted online by firms such as IntraLinks and Merrill Corporation's Datasite, which let several bidders review documents at once from their own time zones; the finance chief of Lifeline Systems said doing diligence on the internet sped up its sale to Royal Philips Electronics.

    cfo.com

    Credit

    IntraLinks; Merrill Corporation

  3. 2018

    Claims an outsider could check

    On 14 March 2018 the SEC charged Theranos, its founder Elizabeth Holmes and its former president Ramesh "Sunny" Balwani with raising more than $700 million from investors through "an elaborate, years-long fraud". Theranos told investors that its analyzer ran comprehensive tests from a few drops of blood, when most patient tests ran on modified or standard machines made by others; that the US Department of Defense used its technology in the field, which it never did; and that it would earn over $100 million in 2014, when it earned a little over $100,000.

    LessonEach false claim concerned something that could be confirmed outside the company: what a machine could test, who its customers were, and what it earned.

    sec.gov

    Credit

    the SEC

  4. 2021

    Screening as a deal gate

    The UK's National Security and Investment Act 2021 requires buyers to notify the government of certain acquisitions in 17 areas of the economy and lets it call in other deals for review. In the year to 31 March 2026 the government received 1,324 notifications, 15% more than the year before; it called in 4.4% of the 1,220 it reviewed, and of its nine final orders, eight let deals go ahead under conditions and one blocked a deal.

    assets.publishing.service.gov.uk

    Credit

    the UK government, for the count

  5. 2022

    Eight months, still missed

    Temasek, the Singapore state-owned investor, wrote off its $275 million investment in the crypto exchange FTX and FTX US after FTX collapsed in 2022. Temasek said it had spent about eight months on "an extensive due diligence process". A review by an independent team, reported in May 2023, found no misconduct by the investment team, but the team and senior management "took collective accountability and had their compensation reduced".

    LessonThe length of a review is no measure of what it finds.

    theblock.co

    Credit

    Temasek, for the review; The Block, for the report

  6. 2025

    Four million users that weren't

    JPMorgan Chase agreed to buy Frank, a startup that helped students apply for financial aid, for $175 million, relying on its founder's claim of 4.25 million users; Frank had about 300,000. When Frank's director of engineering refused to fake the data, the founder hired an outside data scientist to build a synthetic list, and an agreed third-party vendor confirmed to JPMorgan only that the list had more than 4.25 million rows. A jury convicted the founder, Charlie Javice, and Olivier Amar of fraud in March 2025; Javice was sentenced to 85 months in prison and ordered to pay $287.5 million in restitution jointly with Amar.

    LessonA count of rows is not a check that the people in them are real.

    fdicoig.gov

    Credit

    the US Attorney's Office for the Southern District of New York

  7. 2025

    Insurance takes the risk

    Representations and warranties insurance (RWI) pays the buyer of a company if the seller's written statements about the business, its representations and warranties, prove untrue. The American Bar Association's 2025 study of 139 public purchase agreements for private companies, signed and closed in 2024 and early 2025, found RWI mentioned in 63% of deals, up from 29% when the study first tracked it in 2017. In 46% of the deals that mentioned it, the policy was the buyer's only route to recovery for a breach of those statements, by the law firm Goulston & Storrs' summary of the study.

    goulstonstorrs.com

    Credit

    the American Bar Association, for the study

Watch for

  1. Test data rooms for agents

    Harvey, a legal AI company, described in July 2026 synthetic virtual data rooms it built to test AI agents on diligence. One, for a fictional cloud-security company, holds more than 3,500 documents and about 45 million tokens, the word pieces AI models read, with dozens of planted problems that range from a contract a change of owner would let a customer end to a software-licensing risk that shows only when product memos and technical specifications are read together. Harvey says agents cannot hold that much text at once and that their "bias is to search efficiently, not completely"; it named no models, published no scores and did not say the rooms would be released.

    Achieved whenThe data rooms and scoring rubric are released, or an outside team scores agents on them and publishes the results.

    harvey.ai

    Claimed by

    Harvey, July 2026

  2. Agents that find drug rivals

    Investors weighing a drug in development need a list of the competing drugs. A paper posted in August 2025 and revised in May 2026 describes an AI agent for this step and reports 83% recall, the share of known competitors it finds, against 65% for OpenAI Deep Research and 60% for Perplexity Labs, measured on five years of diligence memos from a private biotech venture fund. The authors write that "there is no accepted public benchmark for this task".

    Achieved whenAn outside party scores the agent on a public test set.

    arxiv.org

    Claimed by

    Vlad Vinogradov, Andrey Doronichev and co-authors, August 2025

  3. Fast lane for known investors

    The Committee on Foreign Investment in the United States (CFIUS) reviews foreign investments in US businesses for national-security risk. In May 2025 the US Treasury announced a "Known Investor Program", and since then a sample of the most frequent foreign filers has answered a questionnaire in advance, voluntarily and confidentially, to let CFIUS "more efficiently review the transaction". Treasury asked for public comment in February 2026; the program does not change what CFIUS may review, and Treasury states no targets.

    Achieved whenTreasury publishes the program's final terms and reports review times for participants against other filers.

    cfius.gov

    Claimed by

    the US Treasury, May 2025 and February 2026

Open

  1. 01

    Fake users, caught early

    A public test of the checks that a buyer, or a neutral vendor, can run on a target's customer list to tell real users from fabricated ones without the buyer seeing personal data. Build a set of invented customer tables: some generated as a realistic population with sign-up times, activity and email domains, and some fabricated in the ways seen in fraud cases, such as a list made by a script or a bought list relabeled as customers. Score each checker on how many fakes it catches at a fixed rate of false alarms. All data is invented.

    The brief
    Credit · unclaimed
  2. 02

    Check claims against public records

    A public test set of statements of the kind made in pitch decks, each about a fact that a public registry can settle: that a medical device has US Food and Drug Administration (FDA) clearance, that a patent was granted, or that a government body awarded a contract or grant. The statements are written for the test from real registry entries; half are true and half are altered, and the altered ones are labeled as invented, not attributed to any company. Score AI agents on whether they find the right record, reach the right verdict, and say "cannot tell" when the registries are silent.

    The brief
    Credit · unclaimed

Open 01 · the brief

Fake users, caught early

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of the checks that a buyer, or a neutral vendor, can run on a target's customer list to tell real users from fabricated ones without the buyer seeing personal data. Build a set of invented customer tables: some generated as a realistic population with sign-up times, activity and email domains, and some fabricated in the ways seen in fraud cases, such as a list made by a script or a bought list relabeled as customers. Score each checker on how many fakes it catches at a fixed rate of false alarms. All data is invented.

Why it is still open

In the Frank case the agreed vendor confirmed only that the list had more than 4.25 million rows. Research on detecting machine-made tables exists, but a 2025 study by Kindji and co-authors found detection on tables a detector has not seen "challenging", and it does not deal with customer lists in deals; no public benchmark for such checks was found.

fdicoig.govarxiv.org

Built with what exists

SDV (Synthetic Data Vault), an open-source Python library that generates realistic tables; Faker for invented names and addresses; scikit-learn; the detectors of Kindji and co-authors as baselines; Claude Fable 5.2 to write fabrication scripts in varied styles (see the Note "Three working devices").

  1. First step

    Generate 50 realistic and 50 fabricated tables with SDV and Faker, and test whether simple checks, such as sign-up time patterns and the mix of email domains, separate them.

  2. Done looks like

    At least 1,000 labeled tables from at least 10 fabrication methods released openly, with a leaderboard that reports each checker's catch rate at a 1% false-alarm rate, and at least one fabrication method held back to test checkers on fakes they have not seen.

Open 02 · the brief

Check claims against public records

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set of statements of the kind made in pitch decks, each about a fact that a public registry can settle: that a medical device has US Food and Drug Administration (FDA) clearance, that a patent was granted, or that a government body awarded a contract or grant. The statements are written for the test from real registry entries; half are true and half are altered, and the altered ones are labeled as invented, not attributed to any company. Score AI agents on whether they find the right record, reach the right verdict, and say "cannot tell" when the registries are silent.

Why it is still open

Theranos told investors that the US Department of Defense used its technology, which it never did. ClaimDB, a 2026 benchmark, tests claims against large databases, not against these registries, and found that more than half of 30 models scored below 55% accuracy and struggled to abstain. A 2025 paper on drug-diligence agents states that "there is no accepted public benchmark" for its task. No benchmark of pitch-style claims checked against public registries was found.

sec.govarxiv.orgarxiv.org

Built with what exists

openFDA's device 510(k) API; USPTO patent data through PatentsView; USAspending.gov award data; the EU's CORDIS research-grant records; ClaimDB's scoring code as a model; Mistral Large 4 run in-house (see the Note "A trillion open parameters") or Claude Fable 5.2 (see the Note "Three working devices").

  1. First step

    Write 50 statements about medical devices from openFDA 510(k) records, alter half, and test one agent on them.

  2. Done looks like

    At least 1,000 statements across four registries, each with the record that settles it, released openly; two people agree on at least 95% of 100 random labels; scores published for at least five agents, including how often each gives a verdict when it should abstain.