Picture this scenario.
It is Tuesday morning. A partner at a mid-size litigation firm arrives at her desk, opens her laptop, and types a single sentence into her legal AI platform: “Analyse the facts in the Martinez case file, identify our strongest arguments, find supporting precedents across the Fifth and Ninth Circuits, flag any weaknesses in our position, and draft a preliminary case strategy memo.”
She then gets up to make coffee.
By the time she returns — twelve minutes later — a 14-page memo is waiting for her. It has pulled the case file from document management. It has searched Westlaw and LexisNexis for relevant precedents. It has cross-referenced those precedents against the specific facts. It has identified three strong arguments, flagged one significant weakness, and drafted a strategy memo in the firm’s house style — complete with verified citations.
She reviews it. She adds her professional judgement. She sends it to the client.
This is not science fiction. It is not even the future. It is happening right now, in 2026, at firms using the newest generation of legal AI tools. And it represents something fundamentally different from everything that came before.
Welcome to the age of agentic AI.
Why This Matters More Than Everything That Came Before It
For the past three years, the legal industry’s AI conversation has been dominated by generative AI — tools like ChatGPT and Claude that respond to prompts with text. Ask a question, get an answer. Draft a document, get a draft. One input, one output. The human drives every step.
Agentic AI is categorically different. And if you do not understand the distinction, you are not equipped to use it responsibly, supervise it competently, or advise clients who are deploying it.
The difference is not incremental. It is architectural.
Unlike reactive systems such as traditional generative AI that creates original content in response to a user’s prompt, agentic AI is proactive and capable of initiating tasks, adapting to changing environments, and operating independently in real time.
In plain language: generative AI does what you tell it to. Agentic AI works out how to achieve what you want, then does it — often across multiple steps, multiple tools, and multiple decisions — without you needing to be there for each one.
That changes everything about how lawyers need to think about AI — its capabilities, its risks, its supervision requirements, and its professional responsibility implications.
Part I: The Spectrum — From Chatbot to Agent
To understand what agentic AI is, it helps to understand what it replaced. Legal AI has evolved through three distinct generations, each a step-change beyond its predecessor.
Generation 1: The Chatbot (2022–2023)
Tools like the original ChatGPT operated through pure input/output interactions. You type a prompt, it generates a response based on patterns in training data. It cannot access your documents. It cannot take actions. It cannot chain tasks together. It answers one question at a time, forgets the context when you close the window, and is only as smart as the question you know to ask.
This is where most public understanding of AI still sits.
Generation 2: The AI Assistant (2023–2025)
The second generation added context, memory, and limited tool access. Legal AI assistants could retrieve information from a firm’s document management system, access legal databases, and perform more sophisticated multi-turn conversations. Platforms like early versions of Clio Duo, Lexis+ AI, and CoCounsel 1.0 lived here. Better, faster, and more useful — but still fundamentally reactive. The human remained in control of every decision point.
Generation 3: The AI Agent (2025–present)
The third generation — agentic AI — introduces autonomy. An AI agent can be given a goal rather than an instruction. It then plans its own sequence of steps, decides which tools to use, executes those steps, evaluates its own outputs, corrects course when needed, and delivers a completed result.
AI agents are, in essence, autonomous, decision-making systems powered by artificial intelligence. They can be thought of as specialised employees that are deployed to undertake particular functions. In contrast, agentic AI systems and multi-agent systems are more sophisticated systems that are designed to operate with a higher degree of autonomy — these systems act as an AI agent “conductor” or manager that deploys, coordinates and manages multiple agents.
The orchestra analogy is apt and worth dwelling on. The AI agents are individual musicians — each specialised in one instrument. The agentic system is the conductor: it knows the composition, directs when each musician plays, and shapes the final performance. You do not need to manage the oboist and the cellist separately. You tell the conductor what music you want.
Part II: What Agentic AI Actually Looks Like in Legal Practice Today
This is not theoretical. The largest legal AI platforms have already deployed agentic capabilities, and they are being used in production environments at major law firms and corporate legal departments right now.
LexisNexis Protégé — The Multi-Agent Legal Workplace
LexisNexis’ next-generation Protégé General AI now deploys four specialised agents — an orchestrator, legal research agent, web search agent, and customer document agent — collaborating on complex workflows.
In practical terms, when a lawyer asks Protégé to analyse a litigation matter, the orchestrator agent receives the instruction and breaks it down into component tasks. It then directs the legal research agent to search LexisNexis content for relevant precedents, the web search agent to surface current developments, and the customer document agent to reason across the firm’s own case files. Each agent works on its assigned component. The orchestrator assembles the results into a coherent output, verified through Shepard’s Citations.
For example, in Civil Litigation, Protégé delivers a multistep, agent-orchestrated workflow that analyses facts, extracts timelines, identifies party positioning, surfaces issues, and drafts a strategic case-assessment memo in one end-to-end, unified workflow. In property transactions, Protégé can review purchase agreements and due diligence documents, automatically organise and categorise materials, flag potential issues, summarise deal impacts with suggested resolutions, enabling attorneys to research and take next steps within a single, integrated workflow.
Thomson Reuters CoCounsel Legal — Deep Research and Autonomous Document Review
Thomson Reuters’ CoCounsel Legal launched agentic workflows in 2025 featuring autonomous document review and “Deep Research” capabilities. Deep Research generates its own research plans, explains its reasoning logic, and delivers structured reports grounded in Westlaw content — all without requiring the lawyer to formulate each individual research question.
The platform has already reached one million users. It operates across Westlaw, Practical Law, and enterprise content simultaneously, with agents coordinating across those sources to produce integrated outputs.
Harvey AI — The $8 Billion Valuation Agent
Harvey, now valued at $8 billion, has been a pioneer of agentic legal AI. Used by major international firms, Harvey can handle substantial portions of due diligence workflows, contract review and comparison, and regulatory research with minimal human direction at each step. Its daily active usage grew 81% over 2025, and its plans for deeper agentic capabilities in 2026 represent the cutting edge of what legal AI will look like at the most sophisticated end of the market.
The New AI-Native Law Firms
Perhaps the most striking development: entire law firms are now being built around agentic AI infrastructure. Garfield AI made headlines as “the first fully AI-powered law firm authorised by the UK’s Solicitors Regulation Authority,” specialising in small business debt recovery and small claims. Crosby is an “agentic AI-powered law firm” that combines custom software with in-house lawyers for contract review.
Y Combinator’s 2025 Request for Startups explicitly challenged founders to “start your own law firm, staff it with AI agents, and compete with existing law firms.” That challenge is being answered.
Part III: The Professional Responsibility Earthquake
Agentic AI is not just a technology story. It is a professional responsibility story. And the legal profession is only beginning to understand its implications.
When a lawyer uses ChatGPT to draft a paragraph, the supervision required is clear: read the paragraph, verify its accuracy, take responsibility for the output. The task is discrete, the output is visible, and the human is present throughout.
When a lawyer deploys an agentic AI system to execute a multi-step workflow — researching, drafting, verifying, formatting — the supervision calculus becomes far more complex. The agent makes decisions. It chooses which sources to prioritise. It frames issues in particular ways. It may omit arguments it evaluates as weaker. All of this happens between the instruction and the final output, often at machine speed, with limited visibility into the reasoning process.
The Delegation Problem
Academic research is already grappling with the question directly. Agentic AI introduces deeper questions, pushing the boundaries of delegation, unauthorised practice of law, and the very structure of the profession. Its capacity for autonomous action compels reconsideration of the limits of professional responsibility, raising the stakes for ensuring that AI remains a tool rather than a surrogate decision-maker.
The core tension: professional responsibility rules were written assuming humans make every substantive legal decision. Agentic AI performs the work that those rules were designed to govern. The ABA Model Rules have not been updated for agentic AI. The courts have not definitively ruled on where agentic AI outputs fit within a lawyer’s supervisory obligations. The vacuum creates risk.
The Supervision Duty Extends to Agents
The ABA Task Force on Law and Artificial Intelligence has concluded in its most recent report that AI adoption has accelerated dramatically, pushing lawyers, judges, regulators and educators into unfamiliar terrain that demands new ethical frameworks. The Task Force’s position is clear: the profession must shift its focus from whether to use AI to how to govern, supervise, and integrate it responsibly.
That supervision duty — Model Rule 5.1 for subordinate lawyers, Rule 5.3 for non-lawyer assistants — is now being interpreted to extend to AI systems. The lawyer’s duty of competence now necessarily includes technological literacy, and the duty of supervision extends to digital associates and vendors within the AI ecosystem.
In practical terms: if an agentic AI system operating on your instruction produces a filing with errors, fabricated citations, or a legal analysis that overlooks a key argument, the professional responsibility falls on you. The agent does not hold a practising certificate. You do.
The Hallucination Problem Gets Harder, Not Easier
One might assume that more sophisticated AI agents would produce fewer hallucinations. The reality is more nuanced — and more alarming. As legal AI outputs continue to improve, this makes hallucinations harder, not easier, to detect. The risk continues to shift from obviously wrong answers to confidently delivered, plausibly incorrect ones that evade surface-level review.
With first-generation chatbots, hallucinated citations were sometimes detectable because they appeared in obviously wrong contexts. With agentic systems that cite real cases but mischaracterise their holdings — or that present a comprehensively researched memo with one subtle but critical error — the verification burden on the supervising lawyer actually increases, not decreases.
Agent Actions Can Bind Your Firm
Agentic AI introduces a dimension that generative AI does not: autonomous action, not just autonomous output. An agent can send communications, access systems, modify documents, and in some configurations, interact with third-party platforms. Agent-human interactions can trigger disclosure laws and raise issues regarding the system’s authority to bind the company. Agent-agent interactions can quickly grow in scale and complexity, leading to behaviour that may be difficult to oversee and control.
A legal AI agent instructed to “prepare the due diligence report and coordinate with opposing counsel’s platform” is no longer just generating text. It is acting on behalf of the firm. The question of whether that action creates legal obligations — and who bears liability when it goes wrong — is largely unanswered.
Part IV: What the Shift Means for the Lawyer’s Role
The picture that emerges from all of this is not one of lawyers becoming redundant. It is one of lawyers becoming something different — and something that requires a genuinely new set of competencies.
A ‘new lawyer role’ focused on a more ‘cockpit’ and ‘problem-solver’ function — supervising, validating, and applying judgement — rather than acting as the primary doer, particularly for routine and lower-value tasks, will evolve.
The cockpit metaphor is well chosen. A pilot does not build the aircraft, does not manually calculate every navigation variable, and does not personally operate every instrument simultaneously. But the pilot remains unequivocally responsible for the safety of the flight, must understand every system they rely on, must be able to identify when instruments are giving false readings, and must be capable of taking manual control when the automated systems fail.
This is the lawyer’s relationship with agentic AI in 2026 and beyond.
The competencies that matter most are shifting:
From: Performing legal research manually → To: Evaluating AI-generated research for accuracy, completeness, and strategic framing
From: Drafting documents from scratch → To: Directing, reviewing, and taking professional responsibility for AI-drafted documents
From: Managing individual tasks sequentially → To: Designing and overseeing multi-agent workflows that execute complex tasks in parallel
From: Finding information → To: Knowing which questions to ask and critically evaluating AI-generated answers
Lawyers will approach matters the way technologists approach projects: diagnosing the problem, selecting tools, deciding what can be automated, and where human judgement is essential. The lawyer’s role will increasingly resemble a systems architect who designs, supervises, and validates AI-assisted legal work.
The lawyers who understand this transition — who develop the competencies to direct, supervise, and take responsibility for agentic AI — will have an extraordinary competitive advantage. Firms will discover that two lawyers using the same AI tools can get radically different results based on how they frame the task. As a result, AI advantage will concentrate in individuals and teams with strong systems thinking rather than in firms that simply buy the best software; the gap between “AI-native” lawyers and everyone else will widen faster than expected.
Part V: The Governance Imperative — What Every Firm Needs Now
Given all of the above, what should legal organisations be doing right now to prepare for the agentic AI era? The answer is not waiting for the technology to mature further. It is building governance infrastructure before deployment scales beyond your ability to control it.
Governance is no longer optional: Between the EU AI Act (August 2026), Colorado AI Act (June 2026), and proliferating state requirements, formalised AI policies have moved from best practice to compliance obligation.
1. Establish What Your Agents Are Authorised to Do
The single most important governance decision for agentic AI is defining the scope of autonomous action. What can an agent do without human review? What requires a human checkpoint before proceeding? What requires partner-level sign-off?
Build decision perimeters into every agentic workflow you deploy. An agent that drafts a memo is different from an agent that drafts a memo and routes it to the client. An agent that searches for precedents is different from an agent that files a brief based on those precedents. Map the action steps and define the human gates.
2. Assign Human Accountability for Every Agent
Every agentic AI system operating on behalf of your firm should have a named human who is responsible for its outputs and actions. This mirrors the supervisory model for associates — the supervising lawyer cannot escape responsibility by pointing to the associate. The same principle applies to AI.
“The AI did it” will not be a defence before a court, a bar association, or a client. The named human supervisor of the agentic workflow is the accountable professional.
3. Build Verification Into the Workflow Architecture — Not On Top of It
The most dangerous failure mode is treating verification as an afterthought — a final check done under time pressure before a deadline. With agentic AI, verification needs to be embedded into the workflow architecture itself.
This means: citation verification tools integrated into drafting workflows; legal research outputs cross-checked against primary sources before they move to the next workflow stage; human review checkpoints built into the agent’s task sequence, not bolted on at the end.
Legal teams are prioritising structured AI deployment — systems that operate within defined workflows, produce explainable outputs, and maintain clear audit trails. “Show me your guardrails” will increasingly mean “show me your workflow.”
4. Implement Logging and Auditability from Day One
You cannot supervise what you cannot see. Every agentic workflow should generate logs that record what the agent did, which sources it accessed, what decisions it made, and what outputs it produced. These logs serve multiple purposes: they enable quality control, they support regulatory compliance (particularly under the EU AI Act’s Article 12 logging requirements), and they create the audit trail you need if an output is ever challenged.
5. Train Your Lawyers as Systems Architects, Not Just Tool Users
The standard “here is how to use this AI tool” training is insufficient for agentic AI. Lawyers need to understand how to design effective workflows, how to recognise when an agent is making a reasoning error, how to interpret agent-generated outputs critically rather than deferentially, and how to intervene appropriately.
The ABA’s requirement for technological competence is not satisfied by knowing how to type a prompt. In the agentic era, it requires understanding how agents plan, act, and fail.
The Bottom Line: Opportunity and Obligation in Equal Measure
Agentic AI represents the single most significant productivity opportunity the legal profession has ever seen. The ability to compress hours of research, drafting, and analysis into minutes — while maintaining the quality and reliability necessary for high-stakes legal work — will fundamentally change what is possible for lawyers at every level.
The firms that master orchestration and oversight of these agentic systems will gain exponential productivity and a decisive market edge.
But that opportunity comes with obligation — the obligation to supervise, to verify, to understand, and to take professional responsibility for everything that operates under a lawyer’s name. The agent is extraordinarily capable. It does not hold a practising certificate. It will not be sanctioned by the bar. It will not appear before the disciplinary committee. You will.
The lawyers who thrive in the agentic era will be those who approach it the way the best pilots approach their aircraft: with deep respect for the technology’s capabilities, clear-eyed understanding of its failure modes, and an unwavering commitment to maintaining the judgement and oversight that no machine can replace.
The cockpit is more powerful than it has ever been. The pilot still has to fly the plane.
Subscribe to [Your Blog Name] for weekly analysis of AI’s impact on legal practice — written for lawyers who take their professional obligations seriously.
Further Reading
- ABA Task Force on Law and Artificial Intelligence: Year 2 Report (December 2025)
- ABA Formal Opinion 512: Generative AI Tools (July 2024)
- LexisNexis Protégé: lexisnexis.com/protege
- Thomson Reuters CoCounsel Legal: thomsonreuters.com/cocounsel
- DLA Piper: The Rise of Agentic AI — Potential New Legal and Organizational Risks
- Squire Patton Boggs: The Agentic AI Revolution — Managing Legal Risks
- Murray, M.D., “Algorithmic Ethics in an Era of Agentic AI Advocacy” (SSRN, 2025)
Disclaimer: This article is for educational and informational purposes only and does not constitute legal advice. Technology capabilities and professional responsibility interpretations described are based on information available as of March 2026 and are subject to change. Always consult your bar association’s guidance on AI use in legal practice.
