AI agents and contracts: what already works, what remains to be proven

An agent receives a goal, chooses steps and uses tools to carry them out: reading documents, querying a database, drafting, then checking its result. An assistant, by contrast, answers one request at a time. The line remains blurry, and each provider uses the term in its own way.

This ambiguity is not specific to law. In a June 2025 forecast, the firm Gartner describes a phenomenon of “agent washing”: existing assistants, chatbots or automation tools renamed as agents without any new capability. The first reflex is therefore to ask what the tool actually does without human intervention.

Offerings multiplied in 2026. Here is what the providers announce, without prejudging the results:

• Harvey announced in May 2026 a library of more than 500 agents, including contract review and due diligence, as well as a tool for building new ones, in early access.
• Legora describes an agent that plans, executes, checks and delivers work in a project workspace. The user sees the proposed plan and approves the deliverables, according to its presentation.
• Claude has offered legal plugins since February 2026. According to LawSites, four of them are available as agents that can be deployed via API. Anthropic’s repository presents them as producing drafts for a lawyer.
• ChatGPT has had a general-purpose agent mode since July 2025. According to Decrypt, it asks for explicit approval for purchases or access to sensitive data, and OpenAI itself flags the risk of prompt injection. It is not specific to law.

The presentations by Legora, Anthropic and OpenAI provide for some form of human review, at different points. The report we read on Harvey does not detail this point. These descriptions come from the providers or the specialized press, and are not a measure of performance.

The strongest evidence concerns specific tasks, evaluated on legal AI tools rather than on autonomous agents. The Vals AI report of February 2025 compared several specialized tools with a baseline of independent lawyers:

• Questions about a document. Best tool at 94.8%, against 70.1% for the baseline.
• Document summarization. Best tool at 77.2%, against 50.3%.
• Data extraction. Best tool at 75.1%, against 71.1%.
• Redlining. The lawyers keep the best score, at 79.7%, against 65% for the best tool in the report’s table.

These results concern dated products and depend on the study’s method. They do not prove that an agent can chain these tasks without error. Rather, they suggest an order of maturity: the narrower and easier to check the result, the more realistic AI support appears.

In practice, the most reasonable uses to test today share these traits:

• Triage of non-disclosure agreements. The task is repetitive, and each deviation is checked against a position defined in advance.
• Metadata extraction. Dates, parties and notice periods can be checked in the contract.
• Preparing an analysis. The tool compares the contract with a playbook, and the lawyer reviews the flagged deviations.
• Searching a contract repository. A list of contracts to review is easy to check, provided the data is clean.

We found no independent study measuring agents over a full contract cycle under French law. Three issues remain open:

• Final proofreading. The ContractScrub preprint, from August 2026, considers the results of the most advanced models weak on errors in contracts drafted by lawyers. The best one reaches an average recall of 0.75.
• Chaining steps. In a chain of actions, an early error can propagate. Task-by-task studies do not measure this effect.
• Negotiation. Deciding on a concession requires knowing the business relationship, the balance of power and current priorities, which are rarely written in a document.

Then there is the question of return on investment. In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, citing costs, unclear value and inadequate risk controls. This is an analyst forecast, general and not specific to law. It is simply a reminder of the value of starting with a specific use case.

An agent does not only carry the risks of a wrong answer. It can act, and its access widens the scope to be controlled. These points remain manageable if they are discussed before deployment:

• Access. An agent connected to email, the contract repository and e-signature has access to more data than an assistant. The principle of least privilege remains fully relevant.
• Incoming content. An external document may contain a hidden instruction. In September 2026, Morgan Lewis recommended treating these documents as untrusted, limiting connectors and having important conclusions checked.
• Traceability. You must be able to reconstruct what the agent read, decided and changed.
• Confidentiality. Each connector adds a data flow to review, especially if it depends on another provider.
• Binding the company. An automatic sending or acceptance raises a legal question of its own, to be reviewed with your counsel before any rollout.

The CNIL (the French data protection authority) provides a useful marker for deploying generative AI, listing among the examples of uses to prohibit the act of entrusting it with decision-making. This 2024 recommendation is aimed first at personal data, but the logic remains relevant.

A gradual approach makes it possible to learn without exposing the most sensitive contracts:

• Choose a task. Pick a frequent, narrow case with an identifiable right answer.
• Prepare a test set. Gather a few contracts whose difficulties you know, and also measure what the tool misses.
• Set the level of autonomy. Reading, proposing, editing a draft and sending are four distinct levels. Start with the first two.
• Keep an approval step. Nothing goes out without the agreement of a designated person.
• Compare with the current process. Count the review time, which is often missing from demos.
• Ask the provider for proof. Ask about the protocol, the action logs and the data shared.

The Docform assistant finds contracts based on criteria, and prepares analyses and answers that cite the source passage. The lawyer checks, corrects and approves. This approach does not eliminate possible AI errors: it leaves the decision to the teams. The assistant in detail is presented on the dedicated page.