Who's Responsible When Your AI Agent Makes a Move? The Supervision Gap Most Firms Are Ignoring

Who's Responsible When Your AI Agent Makes a Move? The Supervision Gap Most Firms Are Ignoring

A 12-person law firm in Chicago deployed CoCounsel for matter research in January. By March, they were using it for brief drafting and document synthesis. Neither the managing partner nor the associates had written down who was responsible for reviewing AI output before it went to clients. No policy, no checkpoint, no designated reviewer. Just "use it and check it."

That firm is not unusual. It is the median.

By May 2026, 43% of all law firms and legal departments have enterprise-wide generative AI — up from 14% in early 2024, according to the Thomson Reuters Institute's 2026 report. CoCounsel, Clio Work, Legora, and Lexis+ Protege all shipped agentic tools in early 2026. Agentic AI has arrived. Supervision frameworks have not.

This is not another governance policy article. We have a framework for that. This is a professional responsibility article: when your AI agent initiates a task, chains decisions, and produces an output that reaches a client — who supervised it? What does "supervision" actually require? And what does your malpractice insurer expect you to be able to show?

Agentic AI Has Arrived — But "Supervision" Hasn't Been Defined

The critical distinction that TR Institute flagged in its 2026 agentic AI report is the shift from AI that answers to AI that acts.

Generative AI tools — the ones most firm owners have been using for the last two years — produce text when you prompt them. You ask, you receive, you review, you use. The human checkpoint is built in by the nature of the workflow. You couldn't skip reviewing it even if you wanted to; you had to do something with it.

Agentic AI works differently. It initiates tasks. It chains decisions across steps. It produces outputs — sometimes delivered into downstream systems — without a natural pause for human review. What agentic AI looks like in practice is a multi-step workflow where the agent selects sources, synthesizes information, drafts language, and routes it forward, all without prompting at each step.

That's the value proposition. It's also the oversight gap.

The TR Institute found that less than 20% of law firms say they measure AI ROI — meaning most firms have no visibility into what their AI is actually doing on any given matter. When an agent is acting on your behalf and you're not watching, the professional responsibility question becomes urgent: who supervised the work?

The answer, under virtually every professional responsibility framework applicable to US attorneys and CPAs, is the same: the licensed professional who delivered or approved delivery of the work product. The AI's vendor did not supervise it. The software didn't supervise itself. You did — or you were supposed to.

The Professional Responsibility Problem Nobody Is Talking About

Here is the concrete problem: professional responsibility frameworks were built around human supervision of human work. They are now being applied to human supervision of AI-initiated work. The supervision standard doesn't change. The application of it is completely different.

Under ABA Model Rule 5.1, supervising attorneys are responsible for ensuring that subordinates comply with professional conduct rules. Under ABA Model Rule 5.3, attorneys must supervise non-attorney assistants to ensure their conduct is compatible with the attorney's professional obligations. The TR Institute's 2026 report interprets this as extending to AI agents: AI-generated work product is, for professional responsibility purposes, work product generated by a non-attorney assistant that requires attorney supervision.

This is not a fringe reading. It is the mainstream professional responsibility consensus as of 2026, and it is the interpretation that malpractice insurers are increasingly referencing.

The problem is that "supervision" means something different when the assistant is an agent running autonomously. With a paralegal, supervision includes the ability to observe the work in progress, ask questions mid-task, and redirect. With an agentic AI, you receive an output and must assess it retroactively. That's not supervision in the traditional sense — it's quality review after the fact.

Which is exactly why the checkpoint structure matters: you can't build traditional supervision into an agentic workflow, but you can build structured human decision points that constitute adequate oversight under professional responsibility standards.

Three Actions That Require a Human Checkpoint

This is the framework that satisfies the supervision standard without adding hours to your workflow. Three action types require a human checkpoint before proceeding:

Action 1: Any AI output going to a client.

Every document, memo, letter, or analysis that was AI-generated or AI-assisted — and is going outside the firm to a client, opposing counsel, or a court — requires attorney or CPA review before delivery. This is not optional. "The AI drafted it and I spot-checked it" is a supervision statement. "The AI drafted it and we sent it" is not.

The checkpoint here is simple: a designated reviewer signs off before external delivery. For a 10-person law firm, this is typically the responsible attorney on the matter. For a CPA firm, the partner-in-charge. The key is that it is designated in writing before you start using agentic AI — not decided case-by-case.

Action 2: Any multi-step agentic workflow — checkpoint at the decision branch.

When you set up a workflow where an agent is taking multiple sequential actions — research, then synthesis, then draft — you need at least one human decision point at the critical branch. Specifically: at the step where the AI is making a judgment call that affects the output, not just processing information.

For a brief drafting workflow, that decision branch is the AI's selection of which legal arguments to emphasize. For a contract review workflow, it's the AI's risk flagging — what did it identify as material versus immaterial? A human needs to confirm that judgment before the workflow proceeds to draft.

This doesn't require reviewing every step. It requires building one review gate at the step where the AI is reasoning, not just retrieving.

Action 3: Any AI-initiated communication — prior approval, not retroactive review.

If your agentic setup includes AI-initiated outbound communications — client status updates, appointment confirmations, request letters — those require prior approval of the template and scope before the AI sends them. Retroactive review of what the AI sent last week is not supervision. Prior authorization of what the AI is permitted to send is.

This third checkpoint trips up firms that move quickly into agentic workflow automation. The moment your AI is sending communications on behalf of your firm, you have moved from a drafting tool to an agent acting in your name. The professional responsibility bar shifts accordingly.

What Minimum Viable AI Supervision Looks Like for a 5–20 Person Firm

The goal is not to build a compliance function. The goal is to have enough documented process that you can demonstrate — to your malpractice insurer, to your state bar, to a client who asks — that someone was in charge of what the AI did.

For a firm with 5 to 20 professionals, minimum viable AI supervision has three elements:

1. A written eligibility list. Which matter types are eligible for agentic AI handling? This doesn't need to be comprehensive. It can be simple: "AI-assisted research and drafting is approved for [matter type A] and [matter type B]. [Matter type C] requires partner sign-off before AI use." This list should be in writing and reviewed quarterly.

2. A named responsible professional. For every matter where AI is used, the responsible professional is the person who reviews AI output before external delivery. This is typically already defined by your engagement structure — the partner-in-charge or the responsible CPA. You're not creating a new role; you're explicitly mapping the existing responsibility to AI oversight.

3. A matter-level log. A simple record of what AI tools were used on a matter file. This does not need to capture every prompt. It needs to capture: tool used, task type, date, and who reviewed the output before delivery. Most practice management systems have a note field. That's sufficient.

These three elements are what your professional liability insurer will request if a claim arises from AI-assisted work. See our full guide on professional liability insurance and AI tools for how insurers are approaching this in 2026.

For the specific deployment setup, our Claude Cowork guide for small law firms walks through how to configure a firm workspace that makes this supervision structure operational from day one.

For Accounting Firms: The AICPA Independence Layer

Law firms and accounting firms face the same base-level agentic AI supervision gap. Accounting firms carry an additional compliance layer that requires separate attention.

The AICPA's 2026 guidance update clarified two points:

First, AI-assisted work product does not by itself impair independence. Using CAS Sage AI, Intuit Accountant Suite, or CCH Axcess AI to assist with a client's tax preparation or financial statement work is permissible — provided the accountant makes all significant judgments and reviews all work product before delivery. The AI is treated as a sophisticated tool, not an independence-impairing relationship.

Second, the independence analysis changes if the AI tool creates a financial relationship between the firm and client outcomes. AI tools that suggest investment products, generate performance-linked outputs, or create a financial benefit to the firm tied to the AI's recommendations fall into a different independence analysis. If you are uncertain whether a specific agentic tool triggers this analysis, review AICPA ET §1.200 and your state CPA licensing board's current AI guidance before deployment.

The practical addition for accounting firms: your engagement letters need one sentence that addresses AI-assisted work. Something like: "We may use AI-assisted tools to support research, data synthesis, and document preparation. All work product is reviewed and approved by a credentialed professional before delivery." Several malpractice insurers for CPA firms are now recommending this language at renewal. Adding it proactively costs nothing.

For a full review of building a 30-minute law firm AI governance policy that covers the supervision framework in a format you can hand to your team, see that guide — the accounting equivalent applies the same structure with AICPA-specific compliance references substituted for ABA.

What the Supervision Gap Costs When It Matters

A single professional responsibility complaint — even one that resolves without discipline — costs the average small firm 30 to 80 hours in response, documentation gathering, and attorney correspondence. That's before any remediation or coverage analysis.

The malpractice scenario is more expensive. The issue isn't whether an AI tool made an error — tools make errors. The issue is whether your supervision process was adequate when the error occurred. "We had a process" is a defensible position. "We used the tool but didn't have a formal review step" is not.

The firms running agentic AI without a supervision framework are not taking a calculated risk. They're taking a risk they haven't calculated.

The supervision framework described here — eligible matter list, named responsible professional, matter-level log — takes less than a day to build. That's the trade. One day of setup against an undefined liability exposure that grows with every AI-assisted matter file.

What Small Law Firms Are Asking About AI Supervision

Who is professionally responsible if an AI agent makes a mistake on client work?

The licensed professional who supervised the work — the attorney or CPA who delivered or approved delivery of the work product. Agentic AI does not change the accountability chain; it changes the production chain. If CoCounsel prepares a document and you deliver it to a client without review, the professional responsibility risk falls on you, not on Thomson Reuters. The supervision standard applies regardless of whether the first draft came from a junior associate or an AI agent. This is the central finding of the TR Institute's 2026 agentic AI oversight report.

What is the difference between generative AI and agentic AI from a professional responsibility standpoint?

Generative AI answers. Agentic AI acts. When you ask a generative AI tool a question, you review the answer before using it — there's a natural human checkpoint. When an agentic AI tool initiates a task, chains decisions, and produces an output, the workflow often completes without a natural pause for review. TR Institute identified this as the core oversight gap in 2026: lawyers and accountants are extending the same casual review habits they developed with generative AI tools to agentic workflows that have materially different action footprints. "Review the draft" is not the same as "supervise the agent."

Do I need a formal AI governance policy before deploying agentic AI tools like Clio Work or CoCounsel?

You need at minimum three things before activating agentic workflows on client matters: (1) a written decision about which matters are eligible for agentic AI handling — not every matter type carries the same risk; (2) a named professional responsible for reviewing AI-initiated outputs before delivery; (3) a log that documents what the AI did on each matter file. This doesn't require a lawyer or policy consultant to create. It requires a checklist and a responsible party. If your malpractice insurer asks, these three elements are what they'll want to see.

What do professional liability insurers require for firms using agentic AI?

In 2026, most professional liability insurers for law and accounting firms do not yet have specific agentic AI exclusions — but this is changing. The near-term risk is not exclusion but coverage dispute: if a claim arises from AI-assisted work, the insurer will investigate whether you had reasonable supervision processes in place. "We used CoCounsel and reviewed it" is a supervising statement. "CoCounsel ran overnight and we delivered the output in the morning" is not. The distinction matters in a claim. Several underwriters are now asking firms to disclose agentic AI use at renewal. Document your supervision process before they ask.

Is there a difference in oversight requirements between AI tools used for law and AI tools used for accounting?

Yes, with meaningful differences in two areas. First, attorney-client privilege: any AI tool that processes client communications, case strategy notes, or privileged materials creates an additional confidentiality obligation. Second, CPA independence: AICPA ethics guidance treats AI tools that access or analyze client financial data differently from tools that assist with firm operations. The 2026 AICPA update clarified that AI-assisted work product does not by itself impair independence, but using AI tools that give the firm a financial stake in the client's outcomes does. Law firms need privilege protocols; accounting firms need independence protocols. Both are distinct from generic AI governance policies.

---

The Crossing Report covers agentic AI deployment decisions for professional services firm owners — every Monday morning. Subscribe free to get next week's issue.