AI agents are forcing law firms to answer a question chatbots allowed them to avoid: who owns the action?

Until recently, the dominant legal-AI model was relatively simple: the lawyer selected the tool, directed the task, and remained responsible for the result. The lawyer opened the research platform, entered the prompt, and reviewed the answer.

In the traditional AI-assistance model, however, the lawyer remained the decision point between the system’s output and the next consequential action. Court decisions (such as Cooter & Gell v. Hartmarx Corp., 496 U.S. 384, 393 (1990), cited in Park v. Kim, 91 F.4th 610 (2d Cir. 2024)) expressly indicate that “Rule 11 imposes a duty on attorneys to certify that they have conducted a reasonable inquiry and have determined that any papers filed with the court are well grounded in fact, [and] legally tenable”.

That model is changing. An AI agent (see AI RMF 1.0) can retrieve information, invoke tools or other models, execute a task, and pass the resulting output into a subsequent step in a workflow.

So the important question is no longer which AI tool should a lawyer use?

It is: Once the lawyer clicks “Go,” who controls what happens next?

That question moves AI from application selection into architecture.

The Lawyer Should Own Judgment. Not Routing.

Insofar the ABA Formal Opinion 512 itself discusses electronic discovery and technology-assisted review as an established use of AI in legal practice, consider a litigation matter receiving a new document production.

The system may need to identify the matter, check permissions, determine the task, select the appropriate AI capability, apply the firm’s policies, decide whether human review is required, and determine where the result goes next.

AI may generate the analysis; the lawyer remains responsible for exercising legal judgment over whether that analysis can be relied upon. The lawyer may not have to decide, step by step, which model accesses which system, which capability performs document classification, or whether a routine extraction should consume a frontier reasoning model.

That is routing.

As AI becomes more agentic, routing becomes less an individual user’s preference and more a firm-level capability. Thomson Reuters’ 2026 research shows the shift is already underway: 15% of professionals say their organizations are using agentic AI, while another 53% say their organizations are planning or considering it. By 2030, 77% expect agentic AI to become central to their workflows.

The firms preparing for that environment therefore need to design not just how people use AI, but how AI moves through the firm.

Right Matter. Right Task. Right AI.

The mistake is to think of routing as simply choosing between AI vendors. It starts much earlier.

A matter carries context: client, jurisdiction, permissions, confidentiality requirements, practice area, workflow stage and the consequences of getting the task wrong. That context should determine what happens next. For instance,

  • a document-classification task may require a lightweight model.
  • a complex legal analysis may require a stronger reasoning capability.
  • a privileged document may require a different data boundary.
  • a client with specific AI restrictions may eliminate otherwise available tools.
  • a high-consequence output may require mandatory lawyer approval.

The architecture should therefore evaluate:

Matter → Context → Policy → Task → AI capability → Human review → Action

That is more than model selection. It is the firm’s decision about where intelligence is allowed to act.

The Biggest Model Is Not Always the Right Model

Routing also changes the economics of AI. Sending every task to the most capable model can also mean paying for capability the task does not require. Extraction, classification and summarization may require very different levels of reasoning capability from complex legal analysis.

McKinsey’s analysis shows that AI cost can vary significantly based on model choice, token consumption, routing decisions, agent behavior, orchestration patterns and user adoption

For a law firm, that makes routing an economic control.

The system should be able to match the task to the capability, rather than treating every request as if it deserves the most expensive available intelligence. Done well, that can reduce unnecessary model consumption while preserving stronger capabilities for work that genuinely requires deeper reasoning.

The principle is simple: use the right intelligence for the work, not the biggest intelligence available.

When AI Acts, Accountability Changes

The harder issue is not cost. It is responsibility.

ABA Formal Opinion 512 makes clear that using AI does not remove lawyers’ existing obligations concerning competence, confidentiality, supervision, communication, candor and reasonable fees.

Agentic AI adds a new layer to those obligations.

  • If an agent accesses the wrong matter, where was that decision made?

This question becomes material in the light of ABA Model Rule 1.6 which requires lawyers to protect information relating to representation.

  • If confidential information reaches an unapproved system, what prevented it?

This is another significant consideration within the ambit of ABA Model Rule 1.6 (supra).

  • If an agent decides that a task is routine and proceeds without human review, who established that threshold?

Model Rules 1.1 (Competence), 5.1 (Responsibilities of Partners, Managers, and Supervisory Lawyers), and 5.3 (Responsibilities Regarding Nonlawyer Assistance) impose obligations concerning lawyer competence and supervision, including the duty to make reasonable efforts to ensure that lawyers and others assisting the firm conform to applicable professional obligations.

The fact that a system can perform an action does not itself establish that the system should be permitted to perform that action without supervision.

  • If one agent’s output becomes another agent’s instruction, who controls that transition?

Reference to NIST is particularly useful here because it treats AI risk as a system-level and lifecycle issue, rather than merely a property of the individual model.

These are not questions an LLM can answer for the firm. They have to be answered before the action occurs. That is why governance cannot sit separately from the workflow. The controls have to travel with the work.

The Firm Should Control the Route

The AI market is becoming more capable and more fragmented at the same time.

Legal research, drafting, document analysis, knowledge management and general-purpose models are increasingly connected to agents and enterprise workflows. That creates an architectural choice for law firms.

Law firms can allow each application to determine how its own AI operates. Or they can establish a firm-level layer that determines how those capabilities work together. The second approach matters because today’s best model may not be tomorrow’s best model. If the firm’s workflows are built directly around one vendor, replacing that capability becomes an operational project.

If the firm’s architecture controls routing, models and applications can remain replaceable components. The firm owns the workflow. The AI capabilities serve the workflow.

That is the more durable position.

From AI Tools to an AI Operating Layer

This is the gap V Group is built to address. The objective is not to give lawyers another destination for AI. It is to provide the operating layer that connects the firm’s systems, data, AI capabilities and workflows.

That layer can determine:

  • What is this matter? So context and permissions travel with the work.
  • What is this task? So the system distinguishes extraction from reasoning and routine work from consequential judgment.
  • What is permitted? This inquiry ensures that policies, confidentiality requirements and client restrictions are applied before execution.
  • What should perform it? So the appropriate model, application or agent receives the work.
  • Where does human judgment enter? So lawyers remain accountable for decisions that actually require legal judgment rather than becoming manual workflow coordinators.
  • What happened? So the firm can understand how the work moved through its AI environment.

This is the difference between connecting AI tools and operating AI across the firm.

The most advanced law firm will not necessarily be the one that gives its AI agents the most freedom, a report from Thomson Reuters confirms. It will be the one that knows where autonomy creates value and where human judgment must remain in the loop. Thomson Reuters identifies this balance as central to the adoption of agentic AI in legal settings, emphasizing governance, accountability, and rigorous human oversight.

The Question Law-Firm Leaders Should Ask Now

Not: How many AI tools do we have?

Not: How many lawyers are using AI?

And not even: Which model is best?

Ask: When a matter enters our AI environment, how does the firm determine what should happen next?

If the answer is that every lawyer chooses manually, the firm has adopted AI applications. If the answer is that every vendor determines its own workflow, the firm has adopted platforms.

If the answer is that the firm’s architecture routes work according to matter context, policy, task, capability, risk and human judgment, the firm is building something more durable. An AI operating model.

That is where the next competitive advantage will emerge.

Not from having more AI. From making the AI the firm already has work together intelligently.

Join the Conversation

These are the questions we are taking beyond the conference floor through the V Group-sponsored ILTA Roadshows.

The conversation is not about adding another AI tool. It is about what happens when a firm has several AI capabilities, multiple models, connected systems and agents that can act, and needs all of them to operate within one coherent architecture.

If your firm has moved beyond “Should we use AI?” and is asking “How should AI actually work here?”, we would like you to join the conversation.

V Group ILTA Roadshows

AI adoption is the beginning. Architecture is what accelerates the matter.

Resources cited:

  1. https://www.thomsonreuters.com/en/institute/articles/agentic-ai-in-legal-settings
  2. https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10
  3. https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf
  4. https://www.thomsonreuters.com/en/reports/2026-ai-in-professional-services-report
  5. https://www.mckinsey.com/capabilities/quantumblack/our-insights/is-that-ai-agent-worth-it-agentic-economics-and-the-modern-operating-model
  6. https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_6_confidentiality_of_information/
  7. https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_1_competence/
  8. https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_1_responsibilities_of_a_partner_or_supervisory_lawyer/
  9. https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_3_responsibilities_regarding_nonlawyer_assistant/
  10. https://www.nist.gov/itl/ai-risk-management-framework