Skip to content

Insight

Client data and AI: what a law firm should ask before using it

2 November 2026 · 7 min read

Most conversations about AI in law firms start with what the tool can do. Can it summarise a bundle, draft a first letter, pull the key dates out of a lease. Those are fair questions, and the answers are often yes. We think they belong last. Two questions come before them, and the person selling the tool has little reason to raise either: where does client data go once someone presses the button, and what will this cost compared with what you already pay for? A firm that can’t answer the first in writing shouldn’t be using the tool on client matters yet, and a firm that can’t answer the second shouldn’t be paying for it.

Your firm owes its clients confidentiality under the SRA’s Code of Conduct, and under UK GDPR it is the controller of the personal data in its files. Neither duty pauses because the processing happens inside someone else’s product. When an AI provider handles that data for you, the provider is usually a processor, and you remain responsible for choosing it and instructing it.

The problem usually starts with a free tool

The risk is less likely to come through a formal procurement than through a helpful person with a deadline. As an illustration: a trainee has a 40-page lease and a supervising partner who wants the break clauses by lunchtime. A free chatbot is open in another tab. The lease, complete with the client’s name and the property address, goes in. The summary comes back and it is good.

Nobody in that example meant any harm, and the summary may well be accurate. What the firm can’t say is where the lease now sits, whether it was kept, whether it can be used to train the provider’s models, or who at the provider could read it. The terms of a free tool are written for members of the public. They were never drafted with a regulated firm in mind, holding someone else’s confidential papers.

So the first step costs nothing. Decide, and tell everyone, that client data stays out of any AI tool the firm hasn’t approved. Then give people a sanctioned route for the work they were trying to do, because someone with a deadline and no alternative may paste it anyway and not mention it.

What to ask about the data

Get these answers in the contract or the provider’s written terms, since what is said on a sales call isn’t part of either unless someone writes it in.

Where is the data processed and stored? Name the country and the provider’s own suppliers. If it leaves the UK, ask what transfer mechanism covers it. Some clients and some matters will need UK or EU hosting.

Is anything we send used to train or improve a model? The answer you want is no, by default, for prompts, uploaded documents and outputs. If it is only “no” after someone switches a setting off, find out who controls that setting.

What is kept, and for how long? Many services keep prompts and outputs in logs for abuse monitoring or debugging. Check that deleting a document also deletes any prompt that quoted it.

Who at the provider can see our data? Ask about engineers and support staff as well as any human reviewers, and in what circumstances they can look.

Is there a data processing agreement? UK GDPR requires a contract between a controller and its processor, covering instructions, security, sub-processors, breaches and what happens at the end. If a provider has no agreement to offer, look elsewhere.

Does the tool respect our own access rules? If a tool can search across every matter in the firm, it can show a fee earner documents from a matter they are walled off from.

Can we get our data out, and can we have it deleted? Ask what format, how quickly, and whether deletion covers backups and logs.

You may also need a data protection impact assessment, depending on what the tool does and whose data it touches. That is a judgement for your firm and its data protection lead.

What to ask about the cost

A tool can pass every question above and still not be worth paying for. Ask for the cost over the length of the contract and a renewal or two after it, because the first invoice tells you little.

Start with the total: the per-user fee, any charge by document or by usage, set-up, training, and what can change at renewal. If the price moves with usage, ask what a busy month would come to and who gets told when a limit is reached.

Then ask who needs a licence. A tool priced for the whole firm and used by two teams costs far more per person than the quote suggests, so find out whether you can start with the people who’d use it.

Put a figure on checking too. Every output needs a person to check it, and that time comes off whatever the tool saves. If a summary takes nearly as long to check as the document took to read, the tool has saved you little.

Last, find out how you stop: the minimum term, the notice period, and what it takes to move away if the tool doesn’t pay for itself.

Does something you already pay for do it?

Ask this before you buy anything. Practice software vendors and office suites have been adding AI features, some included in what you pay and some sold as an add-on. One of them may do the job for less, with no new supplier holding your data.

Being familiar doesn’t let it off the data questions. An AI feature can come with its own terms and its own sub-processors, even from a supplier you’ve used for years, so ask the same things of it. Then count how often the task comes up. If it’s a handful of times a month, the sensible decision may be to keep doing it by hand and buy nothing.

Then ask what happens to the output

An AI summary of a contract is a draft, and it can be wrong while reading well. It can miss a clause, merge two dates or invent a reference that looks real. Each output has to be checked by a person at the firm before it is relied on or sent anywhere, and the system should make that the normal path through it.

If the output appears in a fee earner’s queue marked as a draft, with the source document beside it and an approve step before it can go anywhere, checking happens because the workflow asks for it. If it appears as finished text in an email window, it will sometimes be sent unread on a busy Friday.

The firm also needs a record of which matter the tool was used on, what went in and who approved the result. When a client or an insurer asks how a piece of work was produced, “someone used an AI tool at some point” is not an answer you want to give.

What we hold ourselves to

We have no AI product to sell. We do build AI into a firm’s own workflow when that’s the right answer, which means every question here applies to us, and a recommendation to build is one you should test. Our own rules on client data, model training and who checks the output are on our security and data page, and compliance stays with your firm.

Ask any supplier these questions, us included. Once the answers come back clear and in writing, the conversation about what the tool can do is worth having.

If this is your firm

If you’re being asked to approve an AI tool and can’t yet say where the data goes, what it would cost over a few years or whether you already own something that does the job, an independent AI and technology review works through those questions for your firm, with the figures written down.

Talk to us

Tell us what your firm pays for and what's taking up your fee earners' time. We'll arrange a short call and tell you whether a review is worth doing.