Agentic AI and the Future of Legal Work: Review, Reasoning and Responsibility

Ten minutes, and one footnote

A general counsel once showed me two things on the same afternoon. The first was a 400-page supply agreement her team had run through an AI review tool. In ten minutes it had flagged every auto-renewal, every uncapped indemnity and every change-of-control trigger. A year earlier, that would have been a week of an associate's time.

The second was a draft brief from outside counsel. Footnote 14 cited a case that, as far as anyone could tell, did not exist.

Same technology. Same month. One output saved her a week; the other could have cost her a sanction and her credibility with a judge. Like the scenes in my other pieces, this is a composite of conversations I have had with legal and technology leaders. But the tension in it is real, and it is the defining question for legal technology right now.

The state of play

Legal is no longer a laggard. Clio's 2025 Legal Trends Report found 79% of legal professionals use AI in some form, and estimated that up to 74% of hourly-billed legal work could be automated ^[1]. Thomson Reuters' 2025 Future of Professionals research found respondents expect AI to save five hours a week, about 240 hours a year ^[2].

But adoption and confidence are not the same thing. In the same Thomson Reuters research, only 22% said their organization had a visible, defined AI strategy, and half named accuracy as a concern ^[2].

The economics are what make legal different from most functions. Clio's report argues hourly billing can penalize lawyers who work faster with AI ^[1]. For an in-house team, that tension disappears: every hour saved is a cost saved. That is why corporate legal departments, not law firms, may be where agentic AI pays back first.

Review: where agents already earn their keep

Review is the beachhead. It is high-volume, pattern-heavy and checkable against a playbook, which is exactly where AI is strongest.

The most cited benchmark is "Better Call GPT," a 2024 study from Onit's AI Center of Excellence ^[3]. Senior lawyers set the ground truth on real procurement contracts, and the researchers compared junior lawyers, legal process outsourcers (LPOs) and large language models ^[3][4].

Source: ^[4], reporting on ^[3].

The accuracy story is more nuanced than the headlines. The best model nearly tied the LPOs on determining legal issues and beat junior lawyers on several measures, but LPOs still led on locating issues ^[4]. The study's own conclusion was augmentation, not replacement ^[4].

What makes this agentic rather than just generative is the chain. An agent can ingest the agreement, compare it to the playbook, redline deviations, route high-risk clauses to the right approver and log every decision. Each step is reviewable. That is the pattern to copy.

Reasoning: the footnote problem

Reasoning is where the risk lives. Finding a clause is retrieval. Deciding whether a precedent controls, distinguishing it, or knowing that the law changed last quarter is judgment, and judgment is where today's systems still fail in ways that look confident.

Stanford RegLab and HAI ran the first preregistered evaluation of commercial legal research AI, using more than 200 legal queries ^[5]. Purpose-built tools clearly beat general chatbots, but tools from certain commercial vendors still produced incorrect information more than 17% of the time, and in one case more than 34% ^[5]. Retrieval-augmented generation reduced the problem. It did not solve it.

The researchers flagged a subtler failure too: answers that state the law correctly but cite a source that doesn't support the claim ^[5]. In law, that may be worse than an invented case, because it survives a casual check.

The courts are keeping score. In Mata v. Avianca (S.D.N.Y., June 2023), Judge P. Kevin Castel fined two lawyers and their firm $5,000 for filing six fictitious ChatGPT-generated opinions and then standing by them ^[6]. That was the first famous case. It was not the last. Damien Charlotin's public database had logged 2,262 court decisions involving AI-hallucinated material as of November 2, 2026, including 1,533 in the United States ^[7]. The filers include 897 lawyers, 1,302 self-represented litigants and 33 judges ^[7]. Contract matters are the single largest legal field in the database, with 461 cases ^[7].

Responsibility: the part you cannot delegate

An agent can draft, compare, route and recommend. It cannot sign. Responsibility stays with a licensed human, and both regulators and courts are making that explicit.

In July 2024 the American Bar Association issued Formal Opinion 512, its first formal guidance on generative AI ^[8]. It maps AI use to six existing duties under the Model Rules:

Judge Castel said it plainly in Mata: using AI for assistance is not inherently improper, but ethics rules give attorneys a gatekeeping role over the accuracy of their filings ^[6].

In Europe, the AI Act lists AI used to assist judicial authorities in researching and interpreting facts and law, and in applying law to facts, as high-risk under Annex III ^[10]. The Digital Omnibus on AI, given final Council approval on June 29, 2026, moved the start date for those stand-alone high-risk obligations from August 2, 2026 to December 2, 2027 ^[10][11].

The playbook: review, reasoning, responsibility

The CIO's job is to give legal a system where speed and accountability reinforce each other. Five moves do most of the work.

1. Automate review aggressively. Start with high-volume, playbook-driven work: NDAs, vendor agreements, renewals and diligence. The cost and time gap in the Onit study is too large to ignore ^[4].

2. Gate reasoning with verification. Any output that cites authority or will reach a court, regulator or counterparty gets a mandatory citation check against the source. Given hallucination rates of 17% or more in leading research tools ^[5], verification is not optional.

3. Ground agents in your own knowledge. Your playbooks, executed agreements, precedent clauses and approval history are your edge. An agent grounded in them answers your question, not the internet's.

4. Make responsibility visible. Every agent action should carry a named reviewer, the sources used and a timestamp. That is how you evidence the supervision duties in Opinion 512 ^[8].

5. Rethink the fee and staffing model with legal, not for it. Opinion 512 expects fees to stay reasonable as efficiency rises ^[8][9], and Clio estimates up to 74% of hourly work is automatable ^[1]. Plan how junior lawyers will learn judgment when the first draft is no longer theirs to write.

Checklist for your next legal AI review

• Inventory every AI tool legal staff use, approved or not
• Confirm each tool's data terms against client confidentiality obligations
• Require source citations on every AI output used in legal work
• Add a mandatory citation-verification step before anything is filed or sent
• Test vendors with false-premise questions and record the results
• Log reviewer, sources and timestamp for every agent action
• Ask counsel to classify any court-facing AI under the EU AI Act before December 2027

A founder's perspective

When we built Contract Intelligence into Yavi.ai, the hardest engineering problem wasn't finding risky clauses. Models do that well. The hard problem was making every finding traceable: which clause, which playbook position, which source, which reviewer. In legal work, an answer without provenance is not an answer. It is a liability.

That is my core belief about agentic AI in law. Review will be automated faster than most firms expect. Reasoning will be augmented, carefully, with verification built into the workflow. Responsibility will never move. The winners will be the legal teams that let agents do the reading so lawyers can do the thinking, and that can prove, line by line, who decided what.

Go back to that general counsel's afternoon. The ten-minute review and the fabricated footnote came from the same technology. The difference was the system around it.