Agentic AI in Legal Practice: The Governance Gap Nobody Is Talking About

Most AI governance conversations in law firms are about supervised AI tools — systems where a fee earner inputs a prompt, the AI generates an output, and a solicitor reviews it before use. The governance framework for this model, while still underdeveloped in most firms, is at least conceptually understood.

Agentic AI is different. And most law firms are not ready for it.

What is Agentic AI?

An AI agent is a system that can take sequences of actions autonomously — using tools, making decisions, and producing outputs — with limited or no human intervention at each step. Where a standard AI tool responds to a single prompt, an agent plans, executes multiple steps, and adapts to what it encounters. The human defines the goal. The agent determines how to achieve it.

Why Agentic AI is Already in Your Firm

Legal AI platforms are moving rapidly toward agentic capabilities. Tools already deployed in law firms — or actively being piloted — include agentic features such as:

Autonomous contract review — an agent reads an entire document set, identifies issues, cross-references defined terms, and produces a structured report without human instruction at each step
Autonomous legal research — an agent searches multiple databases, synthesises results, identifies relevant authorities, and drafts a research memorandum
Automated matter management — agents that schedule tasks, draft and send routine client communications, update matter management systems, and flag deadlines
Multi-step due diligence — agents that work through a due diligence checklist autonomously, gathering information from multiple sources and flagging gaps
Agentic document drafting — agents that draft entire agreements by reference to precedents, client instructions, and negotiation history

If your firm is using Harvey AI, Microsoft Copilot for legal workflows, or any platform marketed as performing multi-step legal tasks autonomously — you are already deploying agentic AI.

Why Existing Governance Frameworks Do Not Cover Agentic AI

Current law firm AI governance frameworks — where they exist at all — are built around a simple model: human prompts AI, AI responds, human reviews. The governance controls are designed for that model: prompt guidelines, output verification checklists, disclosure protocols.

Agentic AI breaks this model in four ways:

The agent takes multiple actions, not one. Each action may require a separate governance decision — but the agent makes those decisions autonomously. A verification checklist designed for a single output cannot govern a 50-step autonomous workflow.
The agent accesses tools and data independently. Agents may search the internet, access client files, query databases, and send communications — all within a single workflow. The data governance implications of each action require separate assessment.
Errors compound across the workflow. An error in step 3 of a 20-step agentic workflow may not be visible until step 18 — by which point the agent has taken 15 further actions on the basis of flawed information.
The supervising solicitor may not know what the agent did. Without mandatory audit logging, agentic actions are invisible. A supervising partner cannot take responsibility for AI outputs they cannot see.

SRA Supervision Applies to Agentic AI

SRA Code Paragraph 7.1 requires effective supervision of work. This applies to agentic AI workflows regardless of their complexity. The supervising solicitor on a matter is responsible for the outputs of any agentic AI used on that matter — whether or not they personally initiated or monitored the agent's actions. Ignorance of what the agent did is not a defence.

The Agentic AI Supervision Framework — Six Elements

An effective governance framework for agentic AI in legal practice must address six distinct elements that do not apply to supervised AI tools:

01 — Task Boundary Definition

Define precisely what the agent is authorised to do — and what it is not. Authorised tasks, data sources it may access, systems it may interact with, and communications it may generate. Boundaries must be set before deployment, not discovered after an incident.

02 — Human Checkpoint Architecture

Define the points in every agentic workflow where human review is mandatory before the agent proceeds. For legal work: before any client communication is sent, before any filing is made, before any commitment is entered into, and at any point where the agent encounters an unexpected situation.

03 — Mandatory Audit Logging

Every agentic action must be logged — what the agent did, what data it accessed, what outputs it produced, and what decisions it made. Logs must be retained and reviewable by the supervising solicitor. No audit log means no supervision and no defence in an SRA investigation.

04 — Escalation and Override Protocols

Define the conditions under which the agent must stop and escalate to a human — ambiguity, unexpected data, potential conflict, or any situation outside defined parameters. Escalation must be automatic, not reliant on the fee earner noticing something has gone wrong.

05 — Data Governance for Agentic Access

An agent that accesses multiple data sources in a single workflow requires data governance assessment for each source. Client confidentiality, data minimisation, and cross-matter conflict checks must be built into the agent's access permissions — not left to the agent's discretion.

06 — Liability Allocation

Establish clearly — in writing — the allocation of liability between the firm, the supervising solicitor, and the AI vendor when agentic AI causes an error. Review vendor contracts for liability caps and indemnities covering agentic AI failures. Many do not address agentic use at all.

EU AI Act Implications for Agentic Legal AI

For EU-facing law firms, agentic AI deployed in legal research, document analysis, or case assessment workflows may be classified as high-risk under EU AI Act Annex III. High-risk agentic systems require: mandatory human oversight mechanisms that allow intervention and override at any point; technical documentation of the system's decision-making logic; and logging of all high-risk system operations for at least six months.

The EU AI Act's human oversight requirements were specifically designed with agentic systems in mind — the regulation anticipates systems that act autonomously and requires that deployers retain meaningful control throughout.

Three Questions Every Managing Partner Should Ask This Week

Which AI tools currently deployed in our firm have agentic capabilities — and are fee earners using those capabilities on client matters?
If an agentic AI took an incorrect action on a client matter today, would we know what it did, when it did it, and what data it accessed?
Does our current AI governance framework — if we have one — address agentic workflows, or was it designed for single-prompt AI tools?

Sources: Law Society AI Practice Notes 2025 · SRA Code of Conduct for Solicitors 2019 (Para 3.5) · SRA Code of Conduct for Firms 2019 (Rule 4.4) · EU AI Act (Regulation 2024/1689) Articles 9, 14 · EU AI Act Annex III · SRA Technology and Innovation Guidance 2024 · SRA Effective Supervision Guidance 2024

This briefing is for informational purposes only and does not constitute legal advice. Ronke Jegede · Cardinal AI Systems · June 2026