Sector playbooks
AI agent for UK law firms: what one actually does
Most of what is sold to firms as an AI agent is a chatbot with a nicer name. The distinction decides whether it saves anybody any time. Here is the job worth giving one first, and how to run it without creating a confidentiality problem.
The short version
- An agent is not a chatbot. A chatbot waits to be prompted; an agent runs a whole job on a trigger and hands you the result.
- The job worth doing first is file preparation: reading the bundle, pulling dates and parties, drafting the chronology, listing what is missing.
- It must not exercise judgement. A fee earner still owns the work, which is exactly what the SRA requires.
- Confidentiality is solved by where it runs, not by which model it uses.
An agent is not a chatbot
A chatbot waits for you to prompt it. You open a window, type a question, read an answer, and the value stops there. Everything depends on somebody remembering to go and use it, which is why so many firms pay for licences nobody opens.
An agent works the other way round. Something happens — a file lands, a deadline moves, a document arrives — and it does the whole job off the back of that without being asked. You do not interact with it. You find the work done.
That is the difference between a tool that gets used and one that does not. Nobody has to remember to use an agent.
The job worth giving one first
In most firms it is file preparation, because it repeats, it has a right answer, and it currently consumes the time of people who are expensive.
Overnight, on a new matter, the agent:
- Reads the bundle, including the scanned and badly named documents nobody wants to open.
- Pulls out the key dates and the parties and puts them in one place.
- Drafts the chronology in the firm’s own format rather than a generic one.
- Lists what is missing or inconsistent — usually the most valuable output, and the one a tired human is worst at.
By the time the fee earner opens the matter the sorting is done, and they start on the part that needs a lawyer. That is the whole proposition, and it is deliberately unglamorous.
What it must not do
It does not advise, it does not decide, and it does not produce anything that leaves the firm unsupervised.
That is not caution for its own sake. A solicitor remains personally responsible for the work whatever produced it, and the High Court has already sanctioned lawyers who cited authorities a generative tool invented. An agent that prepares material for a human to check sits comfortably inside that. One whose output nobody checks does not, and accuracy does not change the regulatory position.
The line that matters
Anything requiring judgement stays with the fee earner. The agent does the sorting, the extraction and the list-making that happens before judgement starts.
Confidentiality is about where it runs
“We can’t put client data into ChatGPT” is correct, and it is where most firms stop looking. The exposure comes from the destination, not the technology: a free public tool may retain your inputs and use them to improve future models, and that is a confidentiality problem whether or not anything visibly goes wrong.
A private deployment removes it. The agent runs on infrastructure you control, reads files that never leave your environment, and is contracted not to train on anything it sees. Same capability, none of the exposure — and something you can put in writing when a client, an insurer or a tender document asks where the data goes.
How to tell whether it is worth it
Three questions, answerable in an afternoon without buying anything.
- Is there a job with a right answer? If two competent people would produce the same output, an agent can do it. If the answer depends on judgement, it cannot.
- Does it happen often enough to matter? A weekly two-hour task beats a dramatic one that happens twice a year.
- Where does the data actually live? Not where it is supposed to live — where it is. This decides how hard the build is, and most firms have not asked it.
If the third answer is uncomfortable, that is worth knowing before you spend anything. It is often the real fix.
Common questions
What is an AI agent, in plain terms?
Software that completes a whole task on its own once something triggers it, rather than waiting for you to type a prompt. A chatbot answers a question. An agent notices a new matter has landed, reads the file, produces the chronology and puts it where you will find it.
Is it safe to use AI on client files?
It depends entirely on where the file goes. Pasting client material into a free public tool that may retain or train on it is the risk. An agent deployed privately, on infrastructure you control and contracted not to train on your data, does not create that exposure.
Does it replace a paralegal or a fee earner?
No, and anyone selling it that way is overselling. It removes the sorting, extraction and list-making that happens before judgement. The fee earner still reads the file and still owns the advice.
How long does one take to build?
For a single well-defined job, weeks rather than months, because the scope is one workflow rather than a platform. The longest part is usually agreeing what “finished” means and getting access to the systems the files already live in.
Related reading
Want to know which job in your firm is worth automating first?
Two weeks, £3,000 fixed. You leave with five written documents covering what AI should be doing, what it would cost and what it would save — including when the answer is that it should not.