Agentic AI in Legal Practice
What it is, what it risks, and what every practising lawyer needs to ask before using it. A practitioner guide for common law jurisdictions.
What it is, what it risks, and what every practising lawyer needs to ask before using it
Common law jurisdictions: England & Wales | India | Singapore
Why This Guide, and Why Now
Most lawyers have heard the phrase ‘agentic AI’ in the last six months. Few have a clear understanding of what it actually means, or why it is materially different from the AI tools they are already using.
That distinction matters. A lot.
Generative AI, the kind that powers ChatGPT, Claude, and Gemini, produces output when you prompt it. You ask. It answers. You are in control of every step.
Agentic AI is different. It does not wait to be asked. It sets goals, plans actions, and executes them across platforms with minimal human instruction. It can send communications, retrieve documents, make decisions, and complete multi-step workflows, autonomously.
The Law Society’s April 2026 report on agentic AI in legal practice puts the core problem plainly: solicitors remain responsible for outputs produced by agentic AI tools, despite not being able to fully audit those outputs.
That gap, responsibility without auditability, is the defining legal risk of agentic AI for practitioners. This guide explains what that means in practice, and what you need to know before any agentic tool enters your workflow.
Part 1: What Is Agentic AI, and How Is It Different?
The distinction that matters
There are three generations of AI tool that lawyers are encountering in 2026. Understanding the differences is not a technical exercise — it is a professional accountability exercise.
| Type | What it does | Who initiates each step? |
|---|---|---|
| Generative AI | Produces text, analysis, or drafts when prompted | You, every time |
| AI Agents | Executes a defined task within guardrails you set | You set the task; agent handles steps |
| Agentic AI | Sets its own goals, plans actions, executes across platforms, autonomously and over time | The system, with minimal human input |
What does this look like in a legal context?
An agentic AI system in legal practice might be instructed to: ‘Review the contracts in this folder, identify change of control clauses, flag those that do not contain a carve-out for internal restructuring, and draft a summary for each flagged contract.’
In completing that task, the system will make a series of autonomous decisions: what constitutes a change of control clause, what counts as a carve-out, how to interpret ambiguous drafting, and what to include in each summary. It will do this across potentially hundreds of documents, without checking in at each step.
That is not inherently wrong. But it creates a category of professional risk that generative AI does not.
Part 2: The Accountability Gap
Responsible, but cannot fully audit
The most important sentence in the Law Society’s agentic AI report is this one:
Solicitors remain responsible for outputs produced by agentic AI tools, despite not being able to fully audit those outputs.
This is not a hypothetical risk. It is a structural feature of how agentic systems work. The system makes decisions across multiple steps. Those decisions may not be logged in a form you can review. The output you receive is the end product of a chain of autonomous choices, and your professional accountability attaches to that output regardless.
Three accountability risks for practitioners
1. The invisible decision chain
When a generative AI tool produces a draft, you can read the draft and assess it. What you cannot see, with agentic systems, is the sequence of decisions the system made in producing it. Which documents did it consult? How did it interpret ambiguous language? What did it exclude, and why?
This matters because professional accountability in law requires not just a correct output, but a defensible process. If an agentic system made an analytical error at step three of a twelve-step workflow, and that error compounded through to the final output, you may not be able to identify it on review, let alone explain it to a client or regulator.
2. The authority question
When an agentic system takes an action in a matter (sends a communication, makes a privilege call, flags a document for disclosure), does that action carry your professional authority?
The answer under current professional conduct rules is yes. The SRA’s standards for solicitors in England and Wales, and the Bar Council of India’s rules for advocates, both operate on the principle that the professional is accountable for work done on their behalf or in their name. An autonomous system acting within your workflow is acting in your name.
This has not yet been tested in a disciplinary or malpractice context. It will be.
3. The privilege and confidentiality risk
Agentic tools often operate across platforms: they may pull from a document management system, interact with an email client, and log interactions to a cloud service. Each connection is a potential point of data exposure.
The Law Society’s report flags that when actions are initiated by agentic AI systems rather than people, established assumptions about what is privileged, what must be disclosed, and what constitutes a legal communication become complicated. A privileged communication generated autonomously by an AI system, without explicit instruction, raises questions about whether the dominant purpose test is satisfied in the same way as a document prepared by a lawyer.
Errors in agentic AI systems also risk being repeated rapidly across multiple matters. A misinterpretation that would affect one document in a manual review can affect hundreds in an agentic workflow, before anyone notices.
Part 3: Where Agentic AI Is Already Entering Legal Practice
Practitioners do not always know when they are using agentic AI. Vendors do not always make this clear. If a tool in your workflow does any of the following without you initiating each step, it is functioning agentically:
- Reviews a set of documents and produces a structured output without individual prompting
- Sends or drafts communications based on triggers rather than explicit instruction
- Makes classification decisions (privilege, relevance, risk level) and acts on them
- Connects to external platforms (email, DMS, calendars) and takes actions across them
- Monitors for changes in a document set or dataset and initiates a response
The Thomson Reuters Institute’s 2026 AI in Professional Services Report found that 15% of legal organisations have already adopted some form of agentic AI, and an additional 53% are either planning for it or actively considering it. Adoption is moving faster than understanding.
At Legalweek 2026, a vendor demonstrated an agent that took a single litigation hold notice and autonomously identified custodians, mapped data sources, drafted preservation letters, and scheduled collection, in under four minutes. The room went quiet. Not from awe. From unease.
Part 4: The Questions Every Practising Lawyer Must Ask
Before you use any tool that operates agentically, or before you approve its use in your team, these are the questions that matter professionally.
On oversight
Question 1 — Where does this system’s autonomy begin and my oversight end, and can I actually see that line?
Why it matters: If you cannot map the decision points in the workflow, you cannot review the output at a level that satisfies your professional obligations. ‘I reviewed the final output’ is not the same as ‘I reviewed the process.’
Question 2 — What checkpoints does this system have built in, and can I add more?
Why it matters: The best-designed agentic legal tools include human-in-the-loop review points at high-risk decisions. If the tool does not allow you to configure these, that is a risk signal.
On authority and accountability
Question 3 — If this system takes an action (sends a communication, flags a document, produces advice), am I in a position to stand behind that action professionally?
Why it matters: Your name is on the work. The SRA and BCI rules on supervision and accountability apply to agentic outputs just as they do to work done by a junior. The question is whether you have reviewed it with equivalent rigour.
Question 4 — Has the client authorised the use of this type of tool in their matter?
Why it matters: An agent acting on inferred instructions, rather than explicit client authority, may breach the fundamental principle that a lawyer acts on client instructions. Client engagement terms written for generative AI may not cover agentic systems. Check.
On data and confidentiality
Question 5 — What data is this system accessing, storing, and sharing, and have I read the vendor’s terms?
Why it matters: Agentic tools that connect across platforms create data flows that may expose client-confidential information. The UK ICO has published guidance confirming that organisations remain responsible for data protection compliance of agentic AI they deploy. The PDPB framework in India imposes equivalent obligations on data fiduciaries. Ignorance of the vendor’s data terms is not a defence.
Question 6 — Would legal professional privilege attach to communications or documents generated autonomously by this system?
Why it matters: The dominant purpose test requires that a document was created for the dominant purpose of litigation or legal advice. A document generated by an autonomous system, without explicit professional instruction at the point of creation, may not satisfy this test in the same way. This is untested in court, which is a reason for caution, not comfort.
On the vendor
Question 7 — What does this vendor actually mean by ‘agentic’?
Why it matters: The Law Society’s report notes that several in-house solicitors described being asked to ‘deliver agentic solutions’ despite the underlying technology lacking the right capabilities. The term is being used as a marketing label as often as a technical descriptor. Ask the vendor to explain specifically what autonomous decisions the system makes, and what it cannot do.
Part 5: A Practical Risk Framework
Before deploying any agentic tool in legal work, run this framework. It is not a compliance checklist — it is a professional judgement tool.
| Risk factor | Lower risk | Higher risk |
|---|---|---|
| Autonomy level | System pauses for human review at key steps | System completes full workflow without check-ins |
| Data sensitivity | Anonymised or non-client data | Live client files, privileged communications |
| Output type | Internal draft for human review | Client-facing communication or court document |
| Decision type | Classification or summarisation | Advice, privilege call, or legal assessment |
| Auditability | Full decision log available | Black-box output, process not visible |
| Client authority | Explicit client consent to agentic tool use | Implied or assumed consent only |
| Matter stage | Early research or admin | Active litigation or high-stakes transaction |
No single factor is determinative. The more higher-risk factors present in a deployment, the more robust your oversight framework needs to be, and the stronger the case for not deploying at all until the tool, and your understanding of it, is ready.
Part 6: What Good Practice Looks Like
The profession does not yet have settled standards for agentic AI in legal practice. Regulators are developing guidance; the Law Society has called for clearer standards on testing, auditing, and accreditation. In that gap, good practice falls on the individual practitioner.
Before deployment
- Read the vendor’s terms, specifically the sections on data processing, logging, and autonomous decision-making
- Map the decision points in the workflow. For each one: who or what makes the decision, and can you review it?
- Confirm client authority. Update engagement letters if necessary to cover agentic tool use explicitly
- Check your firm’s AI policy covers agentic systems, not just generative AI. Many policies written in 2024 and 2025 do not
During use
- Build in human review at every high-risk decision point: privilege calls, client-facing outputs, any document going to court
- Do not treat the final output as a proxy for reviewing the process. Where the process is not auditable, increase scrutiny of the output
- Keep records. If you cannot produce a log of how an agentic system reached an output, note what steps you took to verify it independently
For supervision
- If you are supervising others using agentic tools, the same supervisory standard applies as for any other work. ‘The AI did it’ is not an answer to a supervision failure
- Junior lawyers using agentic tools without understanding how they work is the same risk as junior lawyers working without supervision. Train before you deploy
The AI advantage in legal practice will concentrate in individuals and teams with strong systems thinking, not in firms that simply buy the best software. Two lawyers using the same agentic tool can get radically different results based on how they frame the task, set the guardrails, and review the output.
Final Note
Agentic AI is not a reason to avoid AI. It is a reason to understand it properly.
The lawyers who will use these tools well, and keep their clients and their licences safe, are not the ones who use them most enthusiastically. They are the ones who understand the difference between a tool that assists and a tool that acts; between supervised output and autonomous decision-making; between delegation with oversight and delegation without it.
That understanding is what The AI Bar exists to build.
Sources: Law Society — The Future of Agentic AI in Legal Practice (April 2026); Thomson Reuters Institute — AI in Professional Services Report 2026; National Law Review — 85 Predictions for AI and the Law 2026; Squire Patton Boggs — The Agentic AI Revolution: Managing Legal Risks; PlatinumIDS — Agentic AI in Legal Technology (April 2026).