For many legal departments, AI for lawyers has moved past the pilot stage. And importantly, that shift from pilot to practice turns AI governance into a professional responsibility issue. Axiom's on-demand AI governance CLE explored what happens once a legal team has chosen a tool. Axiom Legal Engineer David O’Hara and Draftpilot Founder and CEO Daniel van Binsbergen, a former corporate lawyer, walked through the conduct rules, court cases, and controls that matter most so lawyers can better implement AI governance in their own practices.
Why AI governance matters after a legal AI pilot
A pilot is often fairly easy to supervise. A small group of users works on a defined use case, and there’s enough visibility to review nearly every output.
Production use looks different, though. Hundreds of people may use the same tool across different matters and workflows, often without a central review queue and sometimes for use cases nobody formally approved. Many people make decisions, but no single team can see everything. Accountability becomes diffused.
And ownership stays where it was. The legal department or the attorney in charge still owns the quality, confidentiality, and defensibility of AI-assisted work product, even when a technology provider or the IT department helped implement the tool. Van Binsbergen shared that this surprises many lawyers, because supervision during a pilot happens almost automatically. Once the organization scales up, the question becomes how to scale the supervision, too.
Scaling AI responsibly means scaling oversight alongside adoption.
What AI governance means for lawyers and legal departments
In the program’s framing, an AI governance framework has four working parts. It starts with the existing rules of professional conduct and any state requirements layered on top. It applies the duties of candor, verification, supervision, and confidentiality to AI-assisted legal work, including work from agentic systems. It uses governance controls that reduce professional responsibility exposure and match human review to legal risk. It also builds documentation and escalation processes that can demonstrate reasonable supervision after the fact.
Proportionality keeps the framework workable. O’Hara explained that a regimen that collapses under its own weight protects nobody, and an abandoned system is worse than a modest one people can maintain. An escalation matrix can flag events that warrant formal intervention, such as fabricated authority, confidential content in an input, or an unsettled legal question. Effective AI governance keeps everything else proportionate to the risk.
The ethical duties lawyers must apply to AI use
AI ethics for lawyers rests on a point O’Hara returned to throughout the discussion: AI may assist with the work, and the duty stays with the lawyer. Nine ABA Model Rules frame that duty, and ABA Formal Opinion 512 applies them to generative AI.
Rule 1.1 requires competence, which O’Hara described as a practical understanding of what a tool can and cannot reliably do, including AI risks such as inaccurate or incomplete output. Knowing model architecture is optional. The question he posed is whether a lawyer can use AI responsibly in representing clients. Rule 1.3 holds that AI’s efficiency does not relax diligence. Rule 1.4 may reach AI use when a client asks or when the use is significant to the representation. Rule 1.6 requires reasonable efforts to protect confidential information, sometimes including informed consent before it goes into an AI tool. Rule 1.13 reminds in-house lawyers that the organization is the client. Rule 2.1 requires independent professional judgment, which no system can take over. Rule 3.3 requires AI-generated cases, quotations, and facts to be verified before filing. Rule 5.1 focuses on the system around the work, and the test is whether policies, training, access controls, and escalation are ensuring that AI systems and their users produce conforming conduct across a department. Rule 5.3 extends supervision to nonlawyer assistants, and the ABA applies it to AI tools and vendors.
Treating an AI assistant like a nonlawyer assistant is a good start, with one caution. A junior associate who invents case law can be investigated and corrected. A generative AI system can repeat the same error across hundreds of matters in a polished, authoritative tone.
The ABA rules are a baseline, and more than 25 state bars have issued guidance on generative AI. The core duties stay consistent, but emphasis varies. California has amendments approved for comment in March that would put AI-specific obligations into six rules of professional conduct, including verification of cited authority and a manager’s responsibility to establish AI policies. Florida’s advisory opinion says a lawyer may not delegate to AI any act that constitutes the practice of law or requires personal professional judgment. North Carolina’s 2024 opinion leans toward disclosure to clients when AI is used in any substantive way, and Axiom discloses AI use to every client before any AI is used.
The EU AI Act is a regulatory regime organized around risk and use cases. Van Binsbergen explained that a legal AI contract redlining tool falls comfortably outside the high-risk category, which includes uses such as HR candidate screening software. For other uses, the obligations center on AI literacy, with no certification and no minimum bar, only a general obligation to ensure that AI users know what the tool can do, its limits, and how to use it appropriately. Companies doing their own AI development should review their obligations separately.
How to verify AI-assisted legal work
Three authorities point to the same obligation. Verify AI-generated work before relying on it. Rule 3.3 applies when AI produces a false quotation or a nonexistent case, Rule 1.3 says time saved does not lower the standard of care, and Federal Rule of Civil Procedure 11(b) covers an attorney’s signature on a filing. In Park v. Kim, decided in January 2024, a reply brief cited a nonexistent state court decision that the attorney admitted ChatGPT had suggested. She had neither read it nor confirmed it. The Second Circuit said no special AI rule was needed, referred her to its grievance panel, and ordered her to give the decision to her client. Counsel must at least read the cited authorities and confirm they exist and support the stated proposition, whether an associate, a vendor, or an AI tool drafted the material.
O’Hara broke verification into five checks: whether the authority exists, whether the quotation is accurate and in context, whether the decision is still good law, whether the source supports the factual assertion attached to it, and whether the authority truly supports the legal proposition for which it is cited. That last check, characterization, is the one lawyers often skip. A case can be real, the quote exact, and the decision still good law, yet the case may have addressed a different issue or rested on narrower grounds. Van Binsbergen also called real authority with a missing conclusion the more dangerous kind of hallucination, so a link alone is not enough. Legal AI tools are making checks easier, and hallucinations are decreasing, but they are still going to be there.
The cases show penalties growing from fines to broader accountability. In a 2023 case, counsel submitted six fabricated ChatGPT decisions and faced a $5,000 penalty. The court saw nothing inherently improper in using a reliable AI tool. The violation was abandoning the gatekeeping role and defending the fake authorities once the court questioned them. In a later case, the court publicly reprimanded three attorneys defending a former Alabama corrections commissioner, removed them from the case, and referred them for bar review. Consequences like these invite reputational damage on top of financial penalties. AI can assist with legal research, but the duty to verify stays with the lawyer.
Review should scale with risk. Higher-risk work needs rigorous attorney review, while lower-risk work may use playbooks, sample sets, or escalations, which helps mitigate risk without a one-size-fits-all rule. Rule 2.1 adds the danger of automation bias, since a fluent, confident answer feels authoritative even when it is wrong. A reviewer who starts from the AI’s conclusion and makes small adjustments has effectively delegated judgment that the rule requires the lawyer to exercise.
Nonlawyer users raise the stakes. Van Binsbergen described nonlawyer personas that keep the chat from answering legal research questions and escalation rules that send unfamiliar redlines to legal. And process gaps matter, too. If a business analyst drafts a regulator response in an approved tool and sends it to legal, the signer owns the certification, but a review step that assumes someone upstream checked the authority has been skipped. Policy must name an owner for each review step.
How to protect confidential information and assess AI vendors
Rule 1.6 requires reasonable efforts to prevent inadvertent disclosure, and enterprise and consumer platforms can handle data very differently. From a six-point due diligence checklist, Van Binsbergen named three questions as most important:
- Can the provider train on your data? He called that the big killer for privilege, which is why free tools are a big no.
- How long does the provider retain the data? Ask about zero data retention and have confidentiality terms in place with your direct vendor.
- Where is the data located? Ask whether the vendor moves data outside the US and whether you can choose where it sits.
Outside counsel guidelines are another lever legal departments already control, yet many say nothing about AI. O’Hara named four provisions: disclosure when AI is used, clear verification standards, defined billing treatment for AI-assisted work, and confidentiality obligations that flow down to every AI provider the firm uses.
Data governance also reaches prompts and outputs. Assume every prompt, output, and audit log could eventually be read by someone outside the original team, and do not assume the technology creates any protection for attorney thinking. Limiting retention protects confidentiality, but other obligations may require keeping the same information, so decide deliberately what the tool retains.
How to govern agentic AI in legal workflows
Agentic AI changes the review model. An agent can plan and execute steps across multiple systems, and its output may go out before anyone approves it, such as an email sent or data transferred without review. Reviewing the final product comes too late, so supervision moves upstream to how execution is designed. O’Hara outlined five controls:
- Define which systems, data, matters, and environments the agent can access
- Specify the actions it is permitted to take, working from a list of what is permitted
- Require human approval before any consequential AI-generated content is shared externally
- Preserve a reviewable record of instructions, decisions, actions, and exceptions
- Make sure completed actions can be stopped, corrected, or rolled back
Regulators are paying attention. In 2025, the California Supreme Court directed the state bar to consider whether additional professional conduct guidance was warranted for agentic AI.
A practical AI governance checklist for legal teams
Whether you run a law firm or a legal department, the program points to these steps for implementing AI governance:
- Check your own jurisdiction’s rules, since the ABA opinion is a baseline.
- Define permitted AI uses across your practice areas and the matters that require legal review.
- Assign an owner to every review step.
- Run all five verification checks on every citation, quotation, and legal proposition before anything reaches a court.
- Match review depth to task risk, and set escalation rules for anything outside approved positions.
- Set access by user group, including what nonlawyers are allowed to do.
- Ask each vendor about training on your data, retention, and data location, and add AI terms to outside counsel guidelines.
- Set access, permitted actions, approvals, logging, and reversibility before an agentic tool goes live.
- Train people and keep documentation proportionate to risk, so policies produce conforming conduct across the department.
- Consider consulting an experienced artificial intelligence lawyer for help developing and implementing AI governance.
AI Governance for Lawyers FAQ
What does AI governance mean for lawyers and legal departments?
It means the policies, training, access controls, review steps, escalation processes, and documentation that let a legal team show it supervises AI use reasonably. O’Hara ties it to candor, verification, supervision, and confidentiality, with human review matched to legal risk.
What ethical obligations apply when lawyers use generative AI?
The existing duties apply: competence (Rule 1.1), diligence (1.3), communication (1.4), confidentiality (1.6), independent judgment (2.1), candor to the tribunal (3.3), and supervision (5.1 and 5.3). State bars may add requirements.
Does ABA Formal Opinion 512 require lawyers to disclose AI use to clients?
O’Hara tied client communication to Rule 1.4, which may include AI use when a client asks or when the use is significant to the representation. State guidance varies. North Carolina leans toward disclosure for substantive use, and Axiom discloses AI use to every client before any AI is used.
Who is responsible for reviewing AI-generated legal work?
The lawyer is. The legal department or attorney in charge owns the quality, confidentiality, and defensibility of the work, and the signing attorney certifies a filing no matter who or what drafted it. Name an owner for each review step so no one assumes someone upstream already checked.
How should lawyers verify AI-generated citations, facts, and legal analysis?
Run five checks on every authority: existence, quotation, currency, assertion, and characterization. Read the cited sources, confirm they support the stated proposition, and apply independent judgment to the analysis.
What should a law firm or legal department include in its AI use policy?
Define permitted uses, identify matters requiring legal review, and prevent AI output from becoming de facto legal advice. Cover training, access controls, escalation, vendor diligence, deliberate retention choices, and documentation proportionate to risk.
Listen to the full discussion by attending the CLE on-demand for up to 1.0 ethics credit.