AI agents for contract drafting and review: a fast first draft, and a documented reason not to skip the lawyer
An agent can draft a contract clause or flag a risky term in seconds. A study of legal hallucinations in large language models shows exactly why that speed still needs a legal review before anything gets signed.
Contract work — drafting a standard clause, comparing a received contract against a company’s usual terms, flagging an unusual condition — is a genuinely strong fit for an AI agent: the source material is text, the patterns are learnable, and a first draft in seconds beats a blank page. It is also one of the areas where the cost of an unnoticed error is highest, which makes it worth being precise about what "AI-assisted" should mean here.
What the research on legal hallucinations found
A study by Matthew Dahl and colleagues at Yale and Stanford, Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models, published in the Journal of Legal Analysis in 2024, tested general-purpose models on specific, verifiable questions about real federal court cases and found legal hallucinations — confidently stated but factually wrong legal claims — in a striking share of responses, ranging from 58% with GPT-4 to 88% with Llama 2 depending on the model. The study also found that models often fail to correct a user’s incorrect legal assumption when asked a leading question, and that they cannot reliably tell when their own answer is one of these hallucinations.
What this sets as the realistic scope for a contract agent
None of this rules out using an agent for contract work — it rules out using it unsupervised for anything that carries legal weight. A first-draft clause, a summary of a long contract’s key terms, a flag on language that differs from a company’s standard template — all genuinely useful, all still requiring a lawyer’s review before signature, precisely because the study shows an agent cannot reliably signal when it has gotten a legal point wrong.
- →Use the agent for drafting and flagging, never for a final legal opinion or a decision to sign without review.
- →Treat any specific legal claim generated by the agent (a citation, a regulation reference, a precedent) as unverified until a person checks it against the primary source.
- →Keep the agent’s scope to your company’s own template and clause library rather than general legal research, where the hallucination rate measured in the study was highest.
Free resource
The self-assessment grid: 20 tasks AI can automate
Sales, admin, support, operations: the 20 tasks AI agents already handle in SMEs — with, for each one, the tell-tale sign that your team is concerned.
Read next
AI agents & automation
Claude Code switches to autonomous mode by default: the lesson for how you supervise your own AI agents
10 August 2026·5 min read
AI agents & automation
OpenAI Presence: even OpenAI won’t sell its AI agents self-service — what it means for your SME
1 August 2026·5 min read
AI agents & automation
Perplexity’s Personal Computer lands on Windows: what an AI agent that touches your files means for your SMB
30 July 2026·5 min read