News · · 3 min read
Agentic AI in law firms: what supervision has to cover
As AI agents move from answering questions to completing tasks, firms must decide what to delegate, where a person approves, and how the work is reviewed.
What happened
Writing on the Legal Futures blog, Matthew Stringer, founder and CEO of Stridon, argues that the conversation about AI in law firms has moved on again. Legal Futures publishes his piece, which describes a shift from conversational tools to agentic AI. Stringer notes that for firms still working out how to get consistent value from the tools they already have, giving people access to AI is very different from embedding it successfully across a firm.
His distinction is simple. With conversational AI you ask for a summary of a matter file, review the response and ask for changes, so you remain central to the work. An agent, he writes, is an AI system that can plan steps, use tools and work across connected systems to deliver something you have asked for. The unit of work changes from a task to an outcome.
He gives the example of being asked to prepare tomorrow's client briefing. You specify the client, matter, deadline and template, and the agent gathers information from the sources it is allowed to access. At agreed checkpoints it returns for review or approval, and before an email is sent or a record changed, you decide whether it should proceed.
Stringer then goes a step further, to what he calls a standing mandate: asking an agent to keep the client briefed on a matter, so that it can act when defined events occur without a fresh instruction each time. That raises questions about what should trigger action, when the agent must stop, which decisions need approval and how someone will review what it has done. He also says Stridon used an example in which an agent could prepare a client briefing for around £2-5, although the actual cost will vary with how much information it needs to review and how complicated the work is.
What it means for the reader
The article does not discuss SRA rules, and it should not be read as doing so. Its value is the set of questions it puts, which are questions about supervision. Stringer says human supervision needs to be part of the process from the start, and that whichever platform a firm uses, the underlying questions about delegation, boundaries and supervision remain.
He proposes five tests for any proposed use: repeatable, checkable, supervised, bounded and worth it. All five need to hold, he says, otherwise the time unlocked at the beginning can disappear into checking and correcting the output. On cost, he adds that if a solicitor spends longer checking an AI-generated briefing than they would have spent writing it, there is little benefit.
He also cautions against handing an entire process to one agent. His suggestion is to use clear rules where consistency matters, AI where summarising or drafting helps, checks to identify exceptions, and a person to approve the outcome. Documenting what good looks like, he adds, makes supervision easier. Stringer suggests the first opportunities may sit in business services, such as HR starter preparation, finance reporting or marketing event administration, alongside carefully selected legal workflows.
His practical advice is that every agent should have someone responsible for its work, with clear rules about what information it can access, what it can do on its own and when it must ask for approval. He also says to keep a record of its actions, review its work regularly and set a spending limit.
Where Counsel AI fits
Counsel AI is software, not a law firm, and this article is general information. Nothing here describes Counsel AI as an agent that acts on a standing mandate.
One point bears on the "checkable" test. Counsel AI's drafts and first reads cite the section of legislation they rely on, so a reviewing solicitor has a specific provision to open and read, rather than a conclusion to take on trust.
A cited section does not make an output correct, and it does not remove the need for review.
Questions
What are the five tests for a proposed use?
Stridon's tests are repeatable, checkable, supervised, bounded and worth it. According to Stringer, all five need to hold, otherwise the time unlocked at the beginning can disappear into checking and correcting the output.
Does Counsel AI act on its own?
The article does not say so. The only product fact used here is that Counsel AI's drafts and first reads cite the section of legislation they rely on, which gives a reviewer a specific provision to check.
Sources
- From chat to Cowork: preparing your law firm for agentic AI · legalfutures.co.uk, 12 Nov 2026
- Counsel AI: /, read 12 Nov 2026
Run a free check, or try every workflow for 14 days
Start