AI contract review: what to automate and where a lawyer still signs

A model can only flag a clause as off-standard if somebody has written the standard down. So the first piece of work in an AI contract review project is legal work: turning the playbook your lawyers carry in their heads into positions, fallbacks and deal-breakers for each clause. Once that exists, the table below is how I would split the rest. The numbers in brackets point to the sources at the foot of the page.

The split follows the ABA’s own line. A lawyer may use AI output as a springboard, an analysis to base advice on or a draft to work from, but may not leave it to the tool alone to advise a client, negotiate a client’s claims or do other work that needs a lawyer’s personal judgement [1]. The opinion is just as clear about why the middle column exists. Generative tools can produce unreliable or incomplete results, can combine accurate information in ways that turn out false, and some are prone to plausible answers with no basis in fact [1].

Whether a contract needs a lawyer at all is a decision for your general counsel. What AI changes is how quickly you can tell which tier a contract is in. I would set three tiers and let the tool sort contracts into them. Sorting is as far as it goes: it never approves.

Two rules hold in every tier. Nothing the tool writes goes to a counterparty until a named person has approved it, and the approval is recorded. The State Bar of California’s 2026 guidance says a lawyer must not deploy an agentic system in a way that lets it send client information out, through automated communications, filings or data transfers, without appropriate safeguards and human review. It adds that the more autonomy a system has, the more oversight the lawyer owes [2].

If contract managers or procurement staff use the tool, they do so under the legal team’s rules. ABA 512 says managerial lawyers must set clear policies on the permissible use of generative AI, that supervisory lawyers must make reasonable efforts to see that lawyers and nonlawyers comply, and that supervision includes training people on the tools they use [1].

ABA 512 describes the test itself, using contracts as the example. A lawyer relying on a tool to review and summarise many long contracts would not necessarily have to check every document by hand, if they had first tested the tool on a smaller set, compared its summaries with their own review and found them accurate [1]. That is an evaluation, and it is the step I would never let a project skip.

Build the test set from contracts you have already negotiated and signed, where your lawyers know what each clause says and what they accepted. Score the tool clause type by clause type, because a tool that handles governing law well can still be poor on limitation of liability. Count three things separately: clauses it missed, clauses it got wrong, and false alarms. Misses are the expensive error, because nobody goes looking for a problem the tool said was not there.

Set the pass mark before the pilot starts, and run the same test whenever the model, the prompts or the playbook change. The California guidance makes the same point: AI systems change through updates and model changes, so competence includes reassessing a system’s capabilities and risks from time to time, and when it is put to new legal tasks [2].

A vendor’s accuracy figure was measured on somebody else’s contracts. ABA 512 cites a June 2024 Stanford study that found the generative AI systems of established legal research providers hallucinate between 17% and 33% of the time [1]. That study measured research tools, not contract review, but it shows why a published figure cannot stand in for your own test. Public datasets help with the plumbing. The Atticus Project’s CUAD holds more than 13,000 labels across 510 commercial contracts, marked up under the supervision of experienced lawyers for 41 clause types [3].

Before anyone pastes a contract into a tool, somebody needs to know where the text goes. ABA 512 says that before lawyers put information relating to a client’s representation into a generative AI tool, they must evaluate the risk that it will be disclosed to or accessed by others, both outside the firm and inside it [1]. For self-learning tools, which can surface one client’s information in answers on another matter, it says the client’s informed consent is required first, and that general, boiler-plate provisions in an engagement letter are not enough [1].

Contracts carry a second layer: the other side’s information, often under a confidentiality clause of its own. Read that clause before the contract goes into any tool a third party runs.

The opinion’s baseline is to read the tool’s terms of use, privacy policy and related contractual terms, or to consult someone who has [1]. Drawing on earlier opinions about cloud computing and outsourcing, it lists what to establish about any tool that stores client information [1].

Walls inside the company matter as much as the vendor’s terms. A shared index across every contract lets a question on one matter pull text from another, which is the cross-matter disclosure both ABA 512 and the California guidance warn about [1, 2]. Give the tool each person’s existing document permissions, matter by matter, rather than one account that can read everything.

The privilege question has started to reach the courts, and the first answers do not agree.

In United States v. Heppner, the Southern District of New York held in February 2026 that a criminal defendant’s written exchanges with a publicly available AI platform were protected by neither attorney-client privilege nor the work product doctrine [4]. He had used the platform on his own, not at his lawyers’ direction. The court found that the platform was not an attorney, and that the exchanges were not confidential, because its privacy policy let the provider use inputs for training and disclose data to third parties, including governmental regulatory authorities. Sharing the outputs with his lawyers later did not make them privileged. The court left one door open: had counsel directed the use, the tool might arguably have acted like an agent of the lawyer [4].

On 10 February 2026, the day the Heppner ruling was first given from the bench, a magistrate judge in the Eastern District of Michigan reached a different result in Warner v. Gilbarco, a civil case in which the plaintiff was representing herself. The court refused to order production of her AI material and said that, even if it were discoverable, it would be protected as work product. It added that generative AI programs are tools, not persons, and that waiving work product protection takes disclosure to an adversary or in a way likely to reach one [5].

The facts differ: who used the tool, whether a lawyer was involved, and a criminal prosecution against a civil claim. Both are early trial-court rulings, and neither settles the law. What I take from them for a legal team is practical. Run legal AI work in tools the company controls, under terms that rule out training on your inputs and limit disclosure, at the direction of counsel, and keep personal accounts on public AI tools away from legal work. Whether that protects privilege in your jurisdiction is a question for your own lawyers.

This section describes the guidance and rulings as published on 29 September 2026. It is not legal advice.

The guidance is written for lawyers, but it sets the standard any contract review system will be judged against, so it is worth reading even when the buyer is legal operations or procurement.

Whether you buy a product or have one built, the questions are the same, and several come straight from what the bar guidance asks lawyers to establish.

1AYM builds the system around your lawyers, and the legal judgement stays with them. For a contracts team that means three pieces of work: encoding the playbook your lawyers set as rules a machine can check against, building the review pipeline in your own environment so every finding quotes its clause and every outgoing change waits for a named approver, and building the test set from your signed contracts so you know the error rates before anyone relies on them. That is our AI governance implementation work, scoped as a fixed statement of work.

Copyright in what we build, the playbook encoding and the test set included, is assigned to you where the contract needs it. US clients can contract through our US entity, and a fixed-scope build can start within a day of the scope being signed. The next step is a 30-minute call.