When the AI Files the Brief: Who's Liable When Autonomous Legal Systems Get It Wrong?

Agentic AI has moved past the chatbot phase inside law firms, and the professional responsibility framework hasn't caught up. Thomson Reuters' CoCounsel Legal, launched in August 2025 and built directly on the Claude Agent SDK, executes "one conversation" workflows that autonomously chain research, analysis, and drafting without per-step human approval. LexisNexis Protégé runs a similar architecture, deploying a Planner Agent, Interactive Agent, and Self-Reflection Agent to orchestrate complex workflows across proprietary legal databases. These are not autocomplete tools. They are autonomous execution systems operating inside the same firms that have already racked up over 700 documented AI hallucination incidents in court filings. The question of who owns the error when one of these agents makes a consequential mistake is genuinely unsettled — and the answer matters for every partner, GC, and ethics counsel deploying these platforms right now.

What "Agentic" Actually Means — and Why It's Categorically Different From a Chatbot

The distinction between a generative AI assistant and an agentic system is not one of degree; it is structural. A chatbot responds to a discrete prompt and awaits the next human instruction. An agentic system receives a goal, decomposes it into subtasks, executes each subtask using available tools (document retrieval, drafting interfaces, filing systems), evaluates its own output, and proceeds to the next step — all without requiring a human to review and approve each action in sequence.

This architecture is precisely what CoCounsel Legal and Protégé now offer. LexisNexis stated at launch that its goal is to automate 15–20% of lawyer tasks by 2028. Thomson Reuters beta-launched its rearchitected CoCounsel at Legalweek 2026, where Eric Dodson Greenberg, General Counsel of Cox Media Group, described witnessing "a fundamental unbundling of the legal profession." The conference itself was dominated by competing visions of which platform would become the primary operating system for legal work. In that race, every vendor's competitive differentiator is precisely the reduction of human friction — fewer interruptions, fewer approval checkpoints, more tasks completed per attorney hour.

For professional responsibility purposes, that frictionlessness is the entire problem.

The Platforms Already Inside Law Firms: CoCounsel, Protégé, and What They Can Do Unsupervised

Protégé's current capability set includes autonomously drafting deposition questions, generating discovery documents, constructing transactional timelines, and analyzing complex deal documents. CoCounsel Legal can execute research-to-draft workflows without a lawyer reviewing the research output before the draft begins. A legal workflow analysis published in early 2026 estimated that agentic systems can analyze 175,000 discovery pages in minutes and complete medical timeline construction in one to four hours — tasks previously requiring weeks of junior associate work.

These capabilities are not hypothetical. Law firm adoption of AI has exceeded 80% access rates, with corporate legal team adoption more than doubling in a single year. The ABA Task Force confirmed in December 2025 that AI has moved from experiment to infrastructure for the legal profession. The deployment is real. The governance to match it is not.

The Accountability Gap: When No Human Signs Off at Every Step, Who Owns the Error?

Under current doctrine, the answer is unambiguous on paper: the supervising attorney owns it. ABA Formal Opinion 512, issued in July 2024, applies Rules 1.1 (competence), 5.1 (supervisory responsibilities for other lawyers), and 5.3 (supervision of non-lawyer assistants) to AI use. The Opinion establishes that lawyers must verify all AI-generated output and that partners with managerial authority must establish clear policies for AI use and ensure compliance. Courts sanctioning lawyers for hallucinated citations have reinforced this: the attorney is accountable regardless of which vendor supplied the tool or how sophisticated the vendor's claims were.

The practical gap opens the moment the agentic workflow spans multiple steps without interruption. If a Protégé Orchestrator Agent conducts research, synthesizes findings, and populates a draft brief autonomously, and the attorney reviews only the final output, the attorney has not supervised the process — only the product. When a hallucination or analytical error is embedded in step two of a six-step chain, it may not be visible in the final document. A MIT-led study cited by Above the Law found that most deployed agent systems offer only minimal logging of actions and timing, often lack effective mechanisms to disclose when operating autonomously versus under human direction, and have no reliable kill switch once execution begins. Lawyers cannot supervise processes they cannot reconstruct.

Malpractice Doctrine Meets Autonomous Execution: What Courts Haven't Decided Yet

The sanctions record to date involves supervised AI use, and it is already alarming. By late 2025, over 700 hallucination incidents had been documented, with cases accumulating at four to five new instances per day. In early 2026, the Sixth Circuit imposed $30,000 in sanctions against two attorneys for briefs containing fabricated citations. Gordon Rees, one of the largest firms in the United States, faced its third hallucination-related sanction in six months by March 2026.

Every one of those cases involved a lawyer who reviewed — however inadequately — the AI output before submission. None involved an agent that acted without review. Multiple experts contributing to the National Law Review's 2026 AI predictions identified the first major "agentic liability" crisis — an autonomous system taking a binding legal action such as filing a motion or accepting a settlement without human approval — as a near-term certainty. No court has yet issued a ruling allocating liability in that scenario, because the scenario hasn't fully materialized yet. When it does, existing malpractice doctrine provides only partial guidance: the supervising attorney is responsible for what the agent did, but the vendor's contractual limitations-of-liability provisions, the firm's engagement letter terms, and the client's own AI use policies will all be in play simultaneously.

The EU AI Act complicates this further for firms with European operations. Its high-risk obligations become fully applicable on August 2, 2026, and the EU Product Liability Directive explicitly includes software and AI systems as products. Firms deploying agentic workflows in EU-adjacent matters face a regulatory layer that American malpractice doctrine has never needed to account for.

What a Real Governance Framework Looks Like — and How Few Firms Have One

The Harvard Law School Forum on Corporate Governance identified four layers required for responsible agentic AI deployment: secure engagement interfaces, compliance-first AI behavior with human governance points, enterprise-grade integrated work platforms, and unified client intelligence that protects privilege. Critically, the framework calls for "auditable agent behavior" with "traceable decision paths" — meaning every step in a multi-step workflow must be logged, attributable, and reconstructible for post-hoc review.

Almost no firm has this. The MIT-led research found that minimal logging is the norm, not the exception, among currently deployed agent systems. Clients are noticing: trust in AI governance has become, according to the Harvard analysis, a deal-maker or deal-breaker in major engagements. General counsels contributing to the 2026 AI predictions survey emphasized that "workflows become the primary control surface for legal AI governance" — but the firms moving fastest on deployment are, by design, minimizing workflow interruptions.

A real governance framework requires designating an executive AI sponsor with actual authority, implementing action logs at the agent level (not just the output level), establishing human checkpoints before any action that could constitute a legal commitment, and maintaining a reliable mechanism to halt an agent mid-execution. Few firms have any of these; almost none have all four.

The Duty of Competence in 2026: Deploying Fast vs. Supervising Responsibly

The competitive pressure is real. Clients expect AI efficiency gains passed through to them directly, and Harvard's analysis is blunt that firms without AI integration will lose business to "leaner, more agile competitors." The era of the pilot program is over. But the duty of competence under Rule 1.1 does not bend to competitive pressure, and neither do courts handing out five-figure sanctions.

The firms that will survive the first major agentic liability event — and there will be one — are those that have treated governance as a competitive differentiator rather than a deployment obstacle. That means auditable workflows, explicit human checkpoints at legally consequential steps, and client-facing AI governance disclosures built into engagement letters before the matter opens. The firms that have deployed CoCounsel or Protégé as plug-and-play productivity tools, without workflow-level controls, have transferred malpractice exposure from the vendor to themselves with no corresponding infrastructure to manage it. As the MIT research put it: "Capability is advancing faster than control frameworks. That imbalance creates exposure." In 2026, that exposure has a name: it's called your next bar complaint.