AI for contract review: what it catches, what it misses, how to check it

A founder pastes a supplier agreement into a chat tool and asks if it is safe to sign. The tool answers in confident, tidy paragraphs. Some of what it says is right. Some of it may be wrong in ways the founder cannot see. AI for contract review is useful, and it is easy to misuse. This guide explains how it works, what it does well and how to check what it tells you. It is general information, not legal advice.

How it works under the bonnet

There are two broad designs. A general chat tool reads whatever you paste and answers from that text plus whatever it absorbed in training. A document-grounded system first retrieves the relevant passages from your files, then writes its answer from those passages and points to them.

The second design matters for legal work. When an answer is tied to a passage, you can open the passage and see if the answer holds. Good AI document analysis software makes that a single click.

Neither design reads a contract the way a lawyer does. A model works with patterns in language. It has no view of the deal, the parties or what happens if the relationship breaks down.

What it does well

The strong side is volume. Tasks that take a person an afternoon of careful reading take a model seconds, and it does not get tired by page thirty.

• Finding clauses. Termination, liability caps, governing law, renewal and confidentiality terms are pulled out of a long document quickly.
• Summarising. A forty-page agreement becomes a one-page outline that tells you where to read closely.
• Comparing. Two versions are set side by side and the changed wording is listed.
• Checking against a standard. Give it your usual terms and it shows where the other side’s draft departs from them.
• Searching by meaning. A question about “ending the contract early” finds clauses that never use those words.

What it misses

The weak spot is what is not on the page. A model notices a missing clause only if you told it the clause should be there. That is why a written checklist, often called a playbook, improves results so much.

It can also miss interactions. A small change in the definitions section may alter the meaning of a clause thirty pages later. A person who knows the deal will spot that. A model often will not.

Then there is the context it never had: what was agreed on the phone, how much bargaining power you hold, what is normal in your industry. A clause can be unusual and still be the right one to accept.

The invented citation problem

Language models predict plausible text. A case citation is a pattern, and a model can fill that pattern with a court, a year and party names that never appeared together in a real judgment. The result looks exactly like a genuine reference.

You cannot detect this by reading the output. The only reliable test is to find the authority in an official source and read it. Even a real case can be cited for something it does not say.

Grounded systems reduce the risk because they answer from retrieved documents. They do not remove it. A model can still summarise a real passage badly.

Confidentiality and client data

Pasting a contract into an online tool sends it to someone else’s server. Before you do, read the provider’s terms for three things: where the data is stored, whether it is used to train models, and when it is deleted. Read the contract terms, not the marketing page.

Professional secrecy and data protection law still apply when the reader is a machine. If you are unsure, strip names and identifying details from the text, or use a system that runs on hardware you control.

Legal AI for startups: a sensible routine

A young company signs many low-value, repetitive contracts and cannot send each one to a law firm. Legal AI for startups works best as a filter that decides what needs a lawyer’s time, and a routine like this keeps it honest.

• Write down your standard position on the clauses you care about.
• Let the tool compare every incoming contract against that list.
• Read the flagged clauses yourself, in the original text.
• Send anything high in value, long in term or hard to exit to a lawyer.
• Before paying for a tool, run it on a contract whose problems you already know.

AI contract review shortens the first read and gives a lawyer a better starting point. It does not give legal advice, and it carries no responsibility for the signature. The person who opens the source and makes the final call is still a lawyer. If you need a system built around your own archive and kept on your own hardware, that is the kind of work we do at Webitro.