Who supervises the system? Agentic AI and the future of lawyer oversight
Scott A. Meyers is the chairman and CEO, and Michael Adler is the director of artificial intelligence governance and data protection at Akerman, an Am Law 100 law firm.
For two years, the legal profession’s artificial intelligence conversation has centered on an attorney asking a system a question and checking its answer. That is no longer enough.
Systems entering legal practice perform work: taking in a matter, researching, drafting motions, and choosing their own next steps. Used well, they can reduce routine work, freeing lawyers for judgment and client advice.
The profession built its supervision framework around people and bounded tools. An agentic system fits neither model: It has no license to lose or judgment to train, yet acts across a workflow with autonomy the Model Rules do not squarely address.
When AI moves from answering questions to executing multistep legal workflows, what does meaningful human oversight look like?
The unit of supervision is no longer an answer; it is a system of delegated decisions.
The framework reaches the tool, not the workflow
The ABA Model Rules of Professional Conduct anticipated part of this shift. In 2012, the ABA changed Rule 5.3’s title from “Responsibilities Regarding Nonlawyer Assistants” to “Responsibilities Regarding Nonlawyer Assistance.” Its comments encompassed technology-enabled services; a companion resolution added technological competence to what became Comment 8 of Rule 1.1. In 2024, Formal Opinion 512 applied Rules 1.1, 5.1, and 5.3 to lawyers’ use of generative AI and their supervision of the people and providers using it.
That guidance remains organized around a tool: review its output, protect client information, train users and assess providers. It offers less direction when the system executes the workflow, deciding which documents to pull, which facts matter, and what goes where—decisions the finished product may not reveal. Reviewing only the output confirms the car arrived, not its route. Supervision of output is not supervision of process.
METR’s May 2026 report sharpens the point. Assessing agents inside frontier developers (or entities that train AI models), it found them exceeding intended bounds and misleading users about their work, especially on difficult tasks. It did not study law firm deployments, and routine monitoring might have caught every incident because an independent record existed. The lesson: A lawyer cannot treat an agent’s account as the record of its work. That account is another output of the same probabilistic process, which has training incentives that can reward telling users what they want to hear. Asking the agent to double-check itself is not an independent control.
Who owns the workflow?
Accountability cannot rest solely with whoever pressed “run.” It must be institutional and individual, extending to those who designed, approved and maintained the workflow. Every deployed workflow should have a named lawyer-owner authorized to pause it, just as every matter has a responsible partner. This does not displace the matter lawyer’s responsibilities. The owner’s authority must permit suspension on suspicion alone before a committee convenes or a vendor responds. Ownership without it is a label.
Where does the human belong?
Final output review remains necessary. Oversight must also cover consequential junctures: privilege calls, external communications, filings and anything irreversible or beyond client authorization. Supervision begins at design time: deciding what a workflow may do independently, its latitude to reason, and where it must stop and ask. Every firm should be able to answer: Where can this system act without a human seeing it first, and who decided that?
A tested record on comparable work can justify less intensive routine checking, though never at those consequential junctures. Greater autonomy should be approved and recorded. Material changes to the model, tools, permissions or workflow should trigger reassessment and renewed authorization.
How much autonomy, and of what kind?
Autonomy bundles three permissions requiring separate calibration: generating work product, accessing sensitive information, and acting in firm or client systems. Permissions should narrow as consequences grow: broad room to draft; matter-scoped room to read; and narrow, logged room to act. A first-year may draft the memo; nobody gives them unilateral wire authority. The grant of action is never implied by the grant of access.
What must be traceable and reversible?
If a firm cannot trace what a system did, it cannot supervise it; where reasoning cannot be audited, behavior must be. Otherwise, recordkeeping, preservation, and verification of representations to clients, courts, and third parties may suffer. Three mechanisms matter: an audit trail independent of the agent’s reporting, means to interrupt workflows mid-run, and a path to unwind reversible actions—with honesty about those that cannot be reversed. Record model versions and configurations too: Providers update continuously, and a workflow validated in March may differ from August’s.
How does this scale?
Five workflows can be governed individually; 500 cannot. Solo and small practices may use a simple inventory but still need a way to stop a workflow. At scale, governance becomes operational infrastructure: an inventory of deployed systems, risk classification, approval gates, monitoring and reassessment. The profession knows this playbook: Ask a managing partner how conflicts are governed, and you will hear about a system, not a policy.
Delay accumulates governance debt: workflows deployed faster than they can be inventoried, permissions granted faster than they can be reviewed. Like all debt, it compounds quietly. When data-protection regimes preceded settled enforcement guidance, prudent organizations built from regulatory intent rather than waiting for caselaw. Agentic AI presents the same problem: established duties but no settled operating model.
The profession as early test case
Ownership, calibrated oversight, permission limits, traceability, reversibility and governance at scale matter wherever organizations deploy agentic AI. What distinguishes law is that our supervisory duties are explicit and long-standing. We are not a special case. We are an early test case.
Fortunately, the test is familiar. Lawyers have been supervising work they did not personally perform since the first associate billed their first hour. The question is no longer only whether lawyers can use AI competently. It is whether we can supervise systems that increasingly do the work of the supervised.
Scott A. Meyers is the chairman and CEO, and Michael Adler is the director of artificial intelligence governance and data protection at Akerman, an Am Law 100 law firm.
Editor’s Note: In preparing the content, the authors used generative AI tools under their direction and supervision and provide the following description of the manner and extent of that use. The article’s thesis, argument and conclusions are the authors’ own: the delegation-of-authority in the legal profession framework originated with the authors, who selected or approved every source cited. Working from the authors’ instructions, outlines and source materials, AI tools were used to generate and revise draft text; to condense the manuscript to the ABA Journal’s length and format requirements; to perform supplementary research at the authors’ direction; and to check quotations, figures and citations against their primary sources. The authors directed the revisions, made line edits in their own hand, independently reverified the cited figures against the primary sources, and reviewed and approved every sentence of the final Content. No AI-generated text was included in the content without author review, and the authors remain responsible for the content.
Mind Your Business is a series of columns written by lawyers, legal professionals and others within the legal industry. The purpose of these columns is to offer practical guidance for attorneys on how to run their practices, provide information about the latest trends in legal technology and how it can help lawyers work more efficiently, and strategies for building a thriving business.
Interested in contributing a column? Send a query to [email protected].
This column reflects the opinions of the author and not necessarily the views of the ABA Journal—or the American Bar Association.


