AI Agents for Legal Document Review: What to Automate, What Stays With a Lawyer

AI agents for legal document review can do the first pass today. They sort NDAs against a playbook, pull clauses, flag diligence issues, rank discovery documents, and draft research with citations. A lawyer still owns every judgment that leaves the building: advice, privilege calls, materiality, redlines sent to the other side, and anything filed in court.

The ethics opinions and the sanctions record agree.

What can AI agents do for legal teams today?

Legal work splits into two jobs. One is finding text: which contracts have a change-of-control clause. The other is deciding what the text means for this client, this deal, or this judge.

Agents are good at the first job because the answer sits in the document, where anyone can check it. They are weak at the second because the answer depends on things outside the document: the client’s goals, bargaining power, local practice. So the useful legal agents are first-pass agents.

The Thomson Reuters Institute’s 2026 AI in Professional Services Report surveyed more than 1,500 respondents in 27 countries across legal, tax and accounting, corporate functions, and government. Organization-wide AI use nearly doubled, to 40% in 2026 from 22% in 2025. 15% of organizations had adopted some type of agentic AI tool, and another 53% were planning for it or considering it. Only 18% said their organization tracks AI return on investment.

Decision rule (mine): if a lawyer can check the agent’s answer against the source in under a minute, the task is an agent candidate. If checking takes as long as doing, it is not.

AI agents for legal document review: the task table

The automation levels are mine, not from a standard. Triage: the agent sorts items into lanes, and a lawyer reviews the risky lanes plus a sample of the safe one. Draft: a lawyer reviews all output before use. Assist: the agent finds material and a lawyer makes the call. None: no agent output reaches the decision.

An unaudited safe lane is where silent misses collect. For where to put approval gates in code, see human-in-the-loop design for agents.

What do the ethics rules require when lawyers use AI?

ABA Formal Opinion 512, issued July 29, 2024, is still the ABA’s main opinion on generative AI. Four duties shape agent design.

Competence. Lawyers need “a reasonable understanding of the capabilities and limitations” of the tool, and the amount of verification “will necessarily depend on the GAI tool and the specific task.” Its example: a lawyer using AI to summarize many long contracts need not necessarily re-read them all, if the lawyer first tested the tool on a smaller subset and found the summaries accurate.

Confidentiality. For self-learning tools that could expose client information to other users, informed consent is required before input, and “merely adding general, boiler-plate provisions to engagement letters” is not enough. State bars differ. Florida Bar Opinion 24-1 (January 19, 2024) says consent is not required when no confidential information is disclosed, and recommends it before using a third-party program when confidential information would be disclosed. It tells lawyers to research the program’s policies on data retention, data sharing, and self-learning. Retention is one input to the consent analysis, not the answer; see the zero data retention guide.

Supervision. Rules 5.1 and 5.3 require clear AI policies and training, and the opinion suggests marking AI-produced material in files.

Fees. Hourly billers “must bill for their actual time” and may not charge clients to learn a tool they use regularly.

Intake agents have their own guidance. Oregon’s Formal Opinion 2026-208 warns that an autonomous agent that sends an engagement letter could reasonably be read as forming an attorney-client relationship. It calls it prudent to make autonomous chatbot actions subject to lawyer approval. I am not a lawyer. Check your own jurisdiction.

Contract review, NDA triage, and clause extraction

The Vals Legal AI Report (February 27, 2025) scored four legal AI tools against a lawyer baseline on seven tasks.

On tasks with one right answer in the text, the best tool beat the lawyers. Data extraction scored 75.1 against a 71.1 baseline. Document Q&A scored 94.8 against 70.1. Summarization scored 77.2 against 50.3. On redlining, the lawyers won: 79.7 against a best tool score of 65.0. Redlining means picking a position and writing language that fits the deal. That is judgment, not lookup.

So the agent compares; it does not negotiate. Give it a playbook with three positions per clause: preferred, fallback, and walk-away. The agent reports which position each clause matches and quotes the clause. A lawyer decides what to send back. If the playbook is not written down, the agent will invent one from its training data.

Failure mode: the silent miss. The dangerous output is “no issues found” on a contract where the problem clause sits inside a definition, an exhibit, or a scanned page. Before launch, build a test set of 30 to 50 past agreements where a lawyer marked every deviation, plant a few more, and measure misses. That is the subset test Opinion 512 describes. See how to write acceptance tests for agents, and read LLM vs OCR extraction if your files are scanned.

Due diligence and e-discovery: courts accepted machine review in 2012

In Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), Magistrate Judge Andrew Peck wrote that “computer-assisted review is an acceptable way to search for relevant ESI in appropriate cases.” It “is not perfect,” but “the Federal Rules of Civil Procedure do not require perfection.” The opinion describes counsel setting cut-off scores “after sampling and quality control tests.” In 2015, in Rio Tinto v. Vale, the same judge wrote that “it is now black letter law” that courts will permit a producing party to use technology-assisted review.

What transfers to agents is the method: prove the ranking by sampling what the machine set aside. For a diligence agent, have a lawyer review a random sample of the documents it did not flag. If the sample turns up a missed material term, widen the flags or review that document type by hand.

What does not transfer is the output type. Classic technology-assisted review scores documents. A generative agent writes sentences about them, and a sentence can misdescribe a clause the agent ranked correctly. So every agent summary in a diligence memo needs a pinpoint quote to the source page. No quote, no entry.

Why do AI legal research tools still invent citations?

Because retrieval does not guarantee grounding. The tool searches a case database, then a model writes from what it found. Two things go wrong. The search misses and the model invents a plausible case from memory. Or the search hits a real case and the model misstates what it holds.

The Stanford study “Hallucination-Free?”, published in the Journal of Empirical Legal Studies in 2025, measured both on 202 queries. A response counted as hallucinated if it was incorrect or “misgrounded,” citing a source that did not support it. Lexis+ AI hallucinated on 17% of queries and was accurate on 65%. Westlaw AI-Assisted Research hallucinated on 33% and was accurate on 41%. Ask Practical Law AI hallucinated on 17% but was accurate on only 19%, because 62% of its answers were incomplete. GPT-4 hallucinated on 43%. These were 2024 versions, so read the rates as a warning, not a current score.

In Vals’ October 2025 legal research study of 200 questions, three legal AI products scored 78% to 81% accuracy and ChatGPT 80%, measured where the products returned a response. Some questions got no response. Vals also found that when ChatGPT could not find a specific judgment or law, it would “still rely on its knowledge to provide an answer.” That is an answer with no located source, the kind of output the protocol below is built to stop.

Damien Charlotin’s AI Hallucination Cases database listed 2,190 decisions as of October 15, 2026, 870 of them involving lawyers. In Mata v. Avianca (S.D.N.Y. 2023), the court imposed a $5,000 penalty and noted there is “nothing inherently improper” about a reliable AI tool, but lawyers have “a gatekeeping role.”

In Lacey v. State Farm (C.D. Cal. 2025), about nine of 27 citations in a brief from two firms were wrong, and “no attorney or staff member at either firm apparently cite-checked” the research. The firms were ordered to pay $31,100. In April 2026, Sullivan & Cromwell apologized to a bankruptcy judge for AI errors in a motion. The firm had AI policies and mandatory training, yet wrote that “the Firm’s protocols were not followed here.”

A policy is not a gate. See the hallucination budget for how much error each workflow can absorb.

A citation-verification protocol with pass/fail steps

Run this on every authority in any memo, opinion letter, or filing that leaves the team. One fail blocks the document.

1. Existence. A human pulls the full text from an official source or commercial database, not the tool that generated the citation. Pass: full text retrieved. Fail: not found, or snippet only.
2. Citation accuracy. Check case name, court, year, reporter, volume, page, and pinpoint. Pass: all match. Fail: any mismatch.
3. Quotations. Find every quoted phrase, word for word, at the pinpoint. Pass: exact match. Fail: paraphrase shown as a quote, or text not found.
4. Support. Read the cited passage. Pass: it supports the proposition exactly as written, as a holding, not a dictum or a dissent. Fail: it does not, or supports only a weaker claim. This is the misgrounding check.
5. Still good law. Run a citator check. Pass: no reversal, overruling, or negative treatment on the cited point. Fail: any.
6. Adverse authority. Opinion 512 lists “a failure to include controlling legal authority” among errors lawyers must correct before filing. Pass: a search for contrary controlling authority is done and logged. Fail: no search on record.
7. Independence. Pass: someone other than the person who prompted the tool signs off. Fail: self-check only.
8. Record. Pass: the log shows checker, date, source, and result for each authority, and AI-generated text is marked in the file. Fail: any gap.

Gate rules. Filings need 100% pass. If two or more authorities in one document fail, re-check every authority in it. Store the log with the matter; the audit trail guide covers fields and retention.

Worked example: NDA triage time savings (hypothetical)

Say an in-house team receives 80 NDAs a month, and a lawyer spends 30 minutes on each. That is 2,400 minutes, or 40 hours.

An agent sorts each NDA against the playbook. Assume 50% green (all terms match), 35% yellow (fallback positions), and 15% red. The lawyer spends 5 minutes per green, 15 per yellow, and 30 per red. One in ten greens gets a full audit, adding 25 minutes each.

1,080 minutes is 18 hours, a saving of 22 hours a month (55%). Validating on 50 past NDAs at 30 minutes each costs 25 hours once, so payback is about 1.1 months (25 ÷ 22).

If only 25% land green, 45% yellow, and 30% red, the total is 100 + 540 + 720 + 50 = 1,410 minutes, or 23.5 hours. Savings fall to 16.5 hours (41%), and payback stretches to about 1.5 months. If yellow reviews take the full 30 minutes in the base case, the total is 200 + 840 + 360 + 100 = 1,500 minutes, or 25 hours. Savings fall to 15 hours (37.5%). Measure green share and yellow review time in month one.

One catch for firms: under Opinion 512, those 22 hours are not billable by the hour, and the same flat fee with and without the tool may be unreasonable. In-house teams get capacity back. Firms need a pricing conversation first.

Bottom Line

Agents do the first pass in legal work. Lawyers keep the judgment and the signature.

Start with NDA triage or clause extraction, validated on a lawyer-coded test set as Opinion 512 describes.

Put the citation protocol in place before an agent drafts a single research memo.

FAQ

Is technology-assisted review the same as generative AI document review?

No. Technology-assisted review trains a classifier on lawyer-coded documents and scores the rest. Generative AI writes about documents. In Rio Tinto, the court noted the Da Silva Moore defendant volunteered to hand over every non-privileged seed set document it reviewed.

Which legal AI research tool hallucinates least?

The Stanford rates (17% for Lexis+ AI and Ask Practical Law AI, 33% for Westlaw AI-Assisted Research) concern 2024 versions and cannot rank today’s products. A February 2026 study tested Westlaw AI and Lexis+ AI on state unemployment insurance statutory surveys and found 58% and 64% accuracy, below a custom research tool at 83%. That covers one statutory task, not hallucination overall. Test on your own question types.

Can a law firm use an AI agent for client intake on its website?

Yes, with controls, under Oregon Formal Opinion 2026-208. It notes that more than one state has said a chatbot that does not identify itself as one could be misleading. It also says to enter prospective clients into the conflict system promptly, because duties to prospective clients apply even if no representation forms.

How should a firm bill for AI tools?

Opinion 512 says it depends on the tool’s function and cost. A tool that works like equipping a practice is overhead absent advance disclosure to the client; its example is an AI grammar checker in word processing software. A third-party service charged per use to review thousands of contracts for one client can ordinarily be billed as an expense. Florida Opinion 24-1 adds that if a lawyer cannot determine the actual cost tied to a client’s matter, the lawyer may not prorate the tool’s periodic charges.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect