Industries / Finance and business services / Legal and professional services
Contracts
Drafting, negotiating, signing and keeping track of agreements between parties.
5
Milestones achieved, from 2000 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2000
Electronic signatures made valid
The US Electronic Signatures in Global and National Commerce Act, approved on 30 June 2000, says a contract "may not be denied legal effect, validity, or enforceability solely because an electronic signature or electronic record was used in its formation."
Credit
US Congress
2016
Code as contract, drained
The DAO was a pool of investors' money run by self-executing code on the Ethereum blockchain, a so-called smart contract. On 17 June 2016 an attacker used a flaw in that code to divert about 3.6 million ether, a third of the fund. Most users then adopted a "hard fork", a change to the blockchain that moved the funds to a recovery address; a minority refused.
LessonCode that runs exactly as written can still do what no party meant.
Credit
account from the US Securities and Exchange Commission's Report of Investigation, July 2017
2021
An open contract answer key
The Atticus Project released CUAD, 510 commercial contracts with more than 13,000 labels across 41 types of clause, marked by law students under the supervision of experienced lawyers and free to use under a CC BY 4.0 license. It was accepted at the NeurIPS 2021 conference, and later tests such as ContractEval (2025), which scored 19 AI models, are built on it.
Credit
The Atticus Project
2024
Model law for automated contracts
The UN Commission on International Trade Law (UNCITRAL) adopted a Model Law on Automated Contracting at its 57th session in New York. A model law is a template that countries can copy into their own law, and this one covers contracts made and carried out by automated systems, including AI and smart contracts.
Credit
UNCITRAL
2025
Lawyers still lead at redlining
In tests that Vals AI published in February 2025, lawyers scored 79.7% at redlining, marking proposed changes to a contract, against 65.0% for the best AI tool, Harvey Assistant. The AI tools beat the lawyers at other tasks, such as answering questions about documents, where the best scored 94.8% against 70.1%.
Credit
Vals AI
Watch for
Agents that strike deals
In a one-week trial in December 2025, Claude agents negotiated and closed 186 deals worth just over $4,000 on behalf of 69 Anthropic employees, who traded their own goods. People whose agents ran a weaker model got worse prices and did not notice.
Achieved whenan outside group runs a similar trial and publishes the outcomes.
Claimed by
Anthropic, April 2026
Open
01
Two agents, one NDA
A public test in which two AI agents negotiate a mutual non-disclosure agreement (NDA, a contract to keep each other's information secret) for opposite sides. Each agent gets a written playbook of what its side may accept. Lawyers grade each final draft blind, beside drafts negotiated by pairs of people with the same playbooks. No real client is involved.
The briefCredit · unclaimed02
A fresh contract answer key
A new set of 100 commercial contracts filed with the US Securities and Exchange Commission after current models were trained, labeled with CUAD's 41 clause types by CUAD's method, with part of the set kept private so no model can be trained on the answers.
The briefCredit · unclaimed
Open 01 · the brief
Two agents, one NDA
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test in which two AI agents negotiate a mutual non-disclosure agreement (NDA, a contract to keep each other's information secret) for opposite sides. Each agent gets a written playbook of what its side may accept. Lawyers grade each final draft blind, beside drafts negotiated by pairs of people with the same playbooks. No real client is involved.
Why it is still open
Harvey's June 2026 benchmark of 500 contracting tasks scores single steps of a negotiation and lists negotiation against a live counterparty as future work. Anthropic's trial had agents trade goods, not contract terms.
Built with what exists
Common Paper's Mutual NDA, a free standard form written by a committee of lawyers and licensed CC BY 4.0, whose one-page cover sheet holds the terms that get negotiated, such as purpose, term and governing law; two agents built with a framework such as the Claude Agent SDK; a shared document that records each change. (see the Note "Many agents, one budget")
First step
Write two opposing playbooks for the NDA's cover sheet, have two people negotiate it once, and keep the result as the first reference draft.
Done looks like
50 negotiations across 5 pairs of playbooks, each final draft graded blind by three lawyers on a published rubric, with the share of drafts that stay inside both playbooks reported beside 10 drafts negotiated by people.
Open 02 · the brief
A fresh contract answer key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A new set of 100 commercial contracts filed with the US Securities and Exchange Commission after current models were trained, labeled with CUAD's 41 clause types by CUAD's method, with part of the set kept private so no model can be trained on the answers.
Why it is still open
CUAD has been public since 2021, and newer tests still use it: ContractEval (2025) scores 19 models on CUAD contracts, which models may have seen during training. I found no held-out set of newer contracts.
Built with what exists
EDGAR full-text search, the SEC's free search of company filings, where material contracts appear as exhibits; CUAD's published labels and categories as the guide; law students supervised by a lawyer, as CUAD used; current models from Anthropic, Google, OpenAI and open-weight makers to score.
First step
Pull 20 contracts filed in 2026 from EDGAR, label them with CUAD's categories, and time it to estimate the hours the full set needs.
Done looks like
100 contracts, each labeled by two people with their agreement rate published, and scores for at least five models on both CUAD and the new set, side by side.