A Legal AI Agent: What It Does With a Contract in Three Minutes

A lawyer in a company with a steady flow of contracts spends most of the day not on hard questions but on the same routine: open the incoming supply contract, find the liability section, compare the penalty with what the company accepts, check jurisdiction, payment terms, termination. Twenty minutes for a simple contract, an hour and a half for a long one. Then the next.

That first read can be given to an agent. Not the decision, just the read. Below: what exactly it does, why it can be trusted with one thing and not another, and how to test it on your own contracts before paying for a project.

What the agent does

Input: a contract from a counterparty as Word, PDF or a scan. Output, a few minutes later, a review:

• Summary: parties, subject, amount, dates, payment terms.
• Deviations from your rules: each with a clause number, a quote from the contract and an explanation. “Clause 7.3: penalty of 0.5% per day, uncapped. Company policy: no more than 0.1% and no more than 10% of the amount.”
• What is missing: for example, no acceptance procedure or no limitation of liability.
• A draft list of proposed amendments for the clauses it flagged.

The lawyer gets a list of places to look at instead of forty pages. Checking each one takes seconds, because the clause and the quote are right there.

How it works, without the technical detail

The core of the agent is not “legal intelligence” but your own rulebook. Every legal team has one, even if it is not written down: what penalty we accept, which jurisdiction we agree to, which payment terms pass without questions and which go to the CFO.

The first step of the project is to write those rules down as a list. The agent then checks the contract against each rule separately: finds the relevant clauses, compares, draws a conclusion and always attaches the quote. The “no quote, no finding” rule is the main defence against made-up results: if the agent cannot point to a place in the text, the remark does not make it into the review.

A side effect that clients often value more than the agent: rules that lived in people’s heads are now written down, and a new lawyer gets up to speed in days rather than months.

What you cannot trust it with

• It does not make decisions. “This risk is acceptable because the client is strategic” is a human call. The agent shows deviations; it does not weigh them.
• It only sees what is in the rules. An unusual scheme an experienced lawyer would sense from the first paragraph will pass the agent by if no rule covers it. Non-standard and large deals are read by a person, in full.
• It makes mistakes on bad source files. A crooked scan, tables in appendices, references to documents that are not in the file. A good agent says “could not read section 4” instead of staying silent, but that needs separate testing.
• It does not follow case law by itself. Changes in legislation and court practice reach it through legal reference systems and rule updates. Someone on the team has to own that.

Responsibility for the review stays with the lawyer. The agent cuts reading time, not the number of signatures.

Where it pays off

Where there are many contracts of the same kind: supply, services, lease, construction, NDA. In one of our projects for lawyers, document analysis went from hours to minutes, and on a routine flow that is most of a team’s time.

Where there are five contracts a month and each is unique, you do not need an agent: setting up the rules will take longer than it saves.

A quick estimate: incoming contracts per month times the average time of a first read. If that comes to less than forty hours a month, it is probably too early.

Confidentiality

Contracts are trade secrets and often personal data. They cannot be sent to an external service as they are. There are two options: anonymisation before the text goes to the model, or an open-weights model on your own servers, where nothing leaves. How to choose and what the second one costs is covered in “AI agents and data law: when you need a closed perimeter”.

How to test before you buy

Two numbers are counted. The first is what share of the lawyer’s remarks the agent found. The second is how many extra remarks it added: if every contract comes with ten false alarms, nobody will use it. There is sometimes a third result: the agent finds something the human missed back then. It does not get tired.

With these numbers the decision takes one conversation: extend to other contract types, refine the rules, or stop. Any of the three is a normal outcome of a pilot.

What it connects to

The agent plugs into where contracts already live: document management systems, and, for the Russian market, 1C:Document Management, with Consultant+ and Garant as legal reference sources. The review arrives in the same place where the lawyer approves the contract, so there is no separate window to open.

If you want an estimate for your own flow, tell us during the audit which contracts come in most often and how many. The audit is free, and we give an estimate within one business day.