How AI reads a contract
The model receives the whole contract, with annexes, specifications and amendments, and splits it into terms: subject, price and payment, deadlines, acceptance, liability, termination and jurisdiction. Each term links to its clause, so a lawyer can open the source text in one click.
Then the AI compares the contract with a reference. The firm sets the reference: the client’s standard contract, past versions with this counterparty or a position on disputed terms that a partner approved earlier.
The AI marks an unclear term as unclear and does not interpret it for the lawyer. Under Article 431 of the Russian Civil Code, a court first takes the literal meaning of the words, and if that is unclear, it compares the term with other terms and with the purpose of the contract as a whole. So an ambiguous wording is a reason for the model to ask a question, not to guess.
What the lawyer gets as output:
1. A table of terms with links to the clauses of the contract and annexes.
2. A list of deviations from the firm’s template, each marked: acceptable, discuss, unacceptable.
3. Questions to the client where facts are missing: who signs the specifications, and whether other contracts with this counterparty have a liability cap.
The firm’s checklist as the base
A general request to “find the risks” gives a general answer. A working result comes from a checklist the firm already keeps for its contract types: supply, works, lease, license, services.
Sample checklist lines for a supply contract on the buyer’s side:
1. The supplier’s penalty for late delivery is not lower than the buyer’s penalty for late payment.
2. Ownership and the risk of accidental loss pass at acceptance, not at shipment.
3. The period for quality claims is not shorter than the incoming inspection period at the buyer’s warehouse.
4. Disputes go to the court at the buyer’s location.
5. The cap on the supplier’s liability does not apply to an intentional breach.
Each checklist line becomes a review rule. For each line the AI answers “met”, “breached” or “term missing” and quotes the contract clause, so the lawyer sees not a summary but a check against the firm’s own rules.
The checklist lives in the firm’s knowledge base, and a partner edits it, not the contractor. A rule added after a lost dispute goes into the review of the next contract the same day.
The list of objections
If the counterparty sent the draft, a reply on other terms is not an acceptance under Article 443 of the Russian Civil Code: it counts as a refusal to accept and a new offer. In practice, such a reply takes the form of a protocol of disagreements, a table with the counterparty’s wording, the client’s wording and the reasons.
The AI fills in the protocol from the review results: it takes a clause, offers wording from the firm’s template and writes the reason in one or two sentences. The lawyer removes the points the client is not ready to argue about and softens the tone where the relationship with the counterparty matters more than the term.
For contracts that one party is required to sign, Article 445 of the Russian Civil Code sets the deadlines: the protocol of disagreements is sent within 30 days of receiving the offer, unless the law sets other deadlines or the parties agree on others. When the AI sees such a case, it puts the deadline into the contract card, and the lawyer confirms that the rule applies.
When you need to write your own draft rather than fix someone else’s, a different process works: drafting a contract from the firm’s templates.
Where AI makes mistakes
The model makes mistakes in predictable places, and the lawyer’s check is built around them. In a Stanford RegLab and HAI study in May 2024, legal research services gave wrong answers in more than 17% of cases, and one of the services in more than 34%.
Typical mistakes in contract review:
1. Missing a term that sits in an annex or in a document the contract only refers to.
2. A confident reference to a rule or a court decision that does not exist or says something else.
3. A wrong clause reference: after the counterparty’s edits the numbering shifted, and the model checked the old number.
4. A judgment without the deal context: a term is dangerous for one client and normal for another.
Hence two process rules. The AI does not cite a rule or a decision without the source text next to it. The lawyer checks every “term missing” and every proposed wording before it goes to the counterparty.
Why not to upload a contract to a public chat
A client’s contract contains trade secrets and often personal data: signatories, representatives, employees. A public AI service stores requests under its own rules, and the firm does not control who sees them and when.
On 2 May 2023, Samsung banned generative AI for its staff after engineers uploaded internal source code to ChatGPT. For a law firm, such an incident means disclosing client information.
On 17 February 2026, Judge Jed Rakoff ruled in United States v. Heppner that the defendant’s chats with a public AI service were not protected by attorney-client privilege. The court cited the service’s policy, under which prompts and answers are collected and may be disclosed to third parties. More in the review: can you upload documents to ChatGPT.
In our projects, personal data is processed on servers in Russia under 152-FZ, Russia’s personal data law, and the contract text does not go abroad or to services without a data processing agreement. The AI has a separate account with rights only to review contracts, its actions go into a log, and signing and negotiations stay with the lawyer. The page on AI for a legal department shows how to build this into the department’s work.