Skip to content

Data

What 510 public contracts leave out

When a business contract lands on your desk, which protections can you assume are in it? We looked at a public dataset of signed commercial contracts to find out how often common clauses actually appear.

The data

CUAD v1 (Contract Understanding Atticus Dataset) from The Atticus Project, released under CC BY 4.0: 510 commercial contracts that US public companies filed with the SEC, grouped into 25 contract types. Law students labelled 41 clause categories in each one, with attorney review (CUAD README, “Labeling Process”).

Method

For every contract and category, we counted the clause as present when the labellers marked any text for it, and absent when they marked none — the rule the CUAD README itself uses. This is a count of labels, not a legal judgment about any clause. The script uses only Python's standard library and produces the same output byte-for-byte on every run. We changed nothing in the labels; the counts are ours.

What was there most often

ClausePresentShare
Governing law437 of 51085.7%
Anti-assignment374 of 51073.3%
Cap on liability275 of 51053.9%
License grant255 of 51050.0%
Audit rights214 of 51042.0%

Three protections that were often missing

  1. No liability-cap language in 46.1% 235 of 510. That does not mean liability was unlimited in those contracts; a separate, narrower category — wording that leaves liability uncapped for particular breaches — appeared in 21.8%.
  2. No right to leave without cause in 64.1%. Where the right exists, the labels do not say which side holds it.
  3. No stated warranty period in 85.3%.

Restrictions that were less common than people expect

Non-compete 23.3% · assignment of IP ownership 24.3% · change of control 23.7%. When one of these is in a contract you are sent, it is less standard than it may look.

By contract type

TypenCap on liabilityUncapped liabilityTermination for convenienceIP ownership assignmentNon-competeInsurance
Co-branding agreement2286.4%54.5%13.6%31.8%54.5%4.5%
Collaboration agreement2653.8%19.2%34.6%34.6%30.8%30.8%
Development agreement2962.1%34.5%55.2%55.2%44.8%37.9%
Distribution agreement3265.6%21.9%40.6%9.4%40.6%31.2%
Endorsement agreement2416.7%0.0%12.5%20.8%29.2%37.5%
Franchise agreement1540.0%6.7%20.0%53.3%66.7%66.7%
Hosting agreement2075.0%45.0%45.0%30.0%20.0%15.0%
IP agreement1735.3%0.0%23.5%88.2%5.9%5.9%
Joint venture agreement234.3%0.0%4.3%13.0%17.4%17.4%
License agreement3348.5%27.3%30.3%27.3%15.2%15.2%
Maintenance agreement3464.7%8.8%38.2%11.8%2.9%32.4%
Manufacturing agreement1752.9%41.2%17.6%17.6%5.9%58.8%
Marketing agreement1758.8%23.5%29.4%29.4%29.4%58.8%
Outsourcing agreement1866.7%16.7%61.1%16.7%0.0%61.1%
Service agreement2839.3%14.3%53.6%21.4%10.7%14.3%
Sponsorship agreement3161.3%25.8%16.1%3.2%16.1%35.5%
Strategic alliance agreement3250.0%18.8%56.2%18.8%25.0%28.1%
Supply agreement1872.2%38.9%38.9%16.7%5.6%72.2%

Types with fewer than 15 contracts in the dataset are left out: at that size one contract moves a percentage by 7 points or more.

Limits

  • These are contracts between public companies and their partners, filed with the SEC. They are not freelance agreements, leases or job offers, and the CUAD README says the set is “not comprehensive or representative”.
  • Absent is not always bad. A missing non-compete or minimum commitment is usually good news for the side that would have been bound by it.
  • The labels show whether a clause exists, not whom it favors.
  • We did not re-check the labels one by one.

How to read these numbers

The numbers are a reference point, not a rule: if a protection you care about is missing from your contract, that is not unusual — and asking for it is a normal part of negotiating a contract.

Source: CUAD v1, The Atticus Project (Hendrycks et al., 2021), CC BY 4.0 — https://creativecommons.org/licenses/by/4.0/. Counts derived from the published labels; no changes to the underlying data.