ChamberOne

Client confidentiality and AI tools: a checklist before you paste anything in

By the ChamberOne team 5 minute read

A paralegal pastes a witness statement into a free chatbot to tidy the language. It takes ten seconds and saves twenty minutes. Where did the statement go? Is it stored? Can anyone at the provider read it? Might it appear, in some form, in someone else's answer later? Most people cannot say. This checklist gives you the questions to put to any AI tool, and the answers to look for, before client material goes near it.

Key points

  • Write down the supplier's answers on retention, training, access, hosting and sub-processors before use.
  • If input can be used for training and that cannot be switched off, keep client material out.
  • Check that restricted matters stay restricted once content is inside an AI tool.
  • Ask for the whole chain of processors and locations, not just the first supplier.
  • Publish a short approved-tools list so staff are not left to guess.

Start with the duty, not the tool

Every legal system has some duty to keep client information confidential, and professional bodies are increasingly publishing views on AI. What that duty requires in your jurisdiction is a question for your regulator's guidance, not for this article. What this checklist does is turn the duty into practical questions you can ask a supplier and record the answers to.

Treat the checklist as a gate. Nothing confidential goes into a tool until the answers are written down and a partner or a designated person has accepted them.

Suppose a six-lawyer firm wants to try a drafting assistant. The practice manager sends the supplier the six questions below, files the replies in a shared folder and asks a partner to sign them off. It takes an afternoon. Six months later, when a corporate client asks how the firm handles AI, the answers are already on file.

Question one: what happens to what you type?

Find out, in writing, whether the provider keeps prompts and uploaded documents, and for how long. Some services delete after processing, some keep for a period for abuse monitoring, and some keep indefinitely unless you opt out. The difference between consumer and business versions of the same product can be large.

Do not assume that paying for a tool changes the terms. Read what applies to your account type, and check again when the supplier announces changes, because terms are updated and defaults can shift.

  • Is input stored, and if so where and for how long?
  • Can the firm delete it, and can that be confirmed?
  • Are conversation histories visible to other users in the same account?

Question two: will it be used to train a model?

Ask directly whether your inputs or outputs may be used to improve the provider's models, and whether that is on by default. Do not rely on a marketing page. Look for the contractual terms, and for the data processing terms that apply to your account tier.

If training use cannot be turned off, the tool is not suitable for client material.

If it can be turned off, check who has authority to turn it off and verify that the setting has actually taken effect for every user, not just the administrator.

Question three: who can see it?

Access applies on both sides. On the supplier's side, ask which staff can read customer content, under what conditions, and whether that access is logged. On your side, ask who in the firm can see what has been entered, and whether restricted matters stay restricted once their content is in the tool.

A conflict wall that works in your document system but disappears in an AI tool is a leak. Check that access controls follow the material.

The same applies to exports and integrations. If a tool can read from your document store, find out whether it respects the permissions on each folder or reads everything the connecting account can see.

Question four: where is it processed and stored?

Find out which countries the data passes through and is held in, for the tool itself and for any underlying model provider. Some clients, and some regulators, have views on cross-border transfer. Some engagement terms or sector rules restrict where data may go.

Be aware that a tool hosted in one country may send text to a model provider in another. Ask for the full chain, not just the first hop.

Where the supplier's answer is vague, ask a follow-up in writing. A clear answer is easy to give. A vague one usually means the supplier either does not know or would prefer not to say.

Question five: who are the sub-processors?

A supplier's product often depends on others: a model provider, a cloud host, a logging service. Ask for the list of sub-processors, how you are told when it changes, and whether you can object. A supplier who cannot produce the list has not done the work you need them to have done.

Also ask how long the supplier gives you before a new sub-processor starts work, and what you can do if you object. A notice period you can act on is worth more than a promise to keep you informed.

Question six: what can you prove afterwards?

If a client asks what happened to their information, or something goes wrong, you will want a record. Check whether the tool logs who used it, when, and on what. Check what the supplier will tell you in an incident, and how quickly.

Add your own records. Note on the file when a tool was used for a matter, which tool, and who reviewed the output.

Keep the supplier's answers in a dated register as well. If a supplier changes its terms or its sub-processors, you want to see what you accepted before and what has altered, so the firm can decide whether the tool remains acceptable.

Make it usable for the rest of the firm

A checklist that lives with the compliance officer will not stop the paralegal at ten past five. Turn it into a short list of approved tools, what each may be used for, and what must never go in. Make the approved tool easy to find and quick to use, because people choose the easy option.

Tell clients, where appropriate, how their information may be processed. Some engagement letters now say so. Check the guidance from your regulator or bar on what to tell clients and when to ask for consent.

Finally, give staff a simple way to report a mistake. Someone will eventually paste the wrong thing into the wrong tool. A firm that makes it easy and safe to say so can respond quickly, while a firm that punishes it will hear about the problem much later.

Do
  • Turn the checklist into a short approved-tools list.
  • Say what each tool may be used for and what must never go in.
  • Give staff a simple way to report a mistake.
  • Tell clients how their information may be processed, where appropriate.
Don't
  • Leave the checklist with the compliance officer alone.
  • Punish people who report a mistake.
Making the checklist work for everyone

How ChamberOne handles this

  • Every read, write and AI call is recorded in the audit log with who, when and the model used, and kept for 7 years by default.
  • A matter or client can be set to Restricted, so only its team, the Owners and people you name can see it.
  • The firm can choose the model for each AI feature, GPT-5 or Claude, or switch AI features off for the firm.
  • The Security and AI page says what is sent to the AI provider and on what terms.

What it does not do

  • With AI features on, the text needed for a request is sent to the AI provider for that feature, outside the country where ChamberOne is hosted. With AI features off, nothing is sent.
  • ChamberOne does not decide whether your use meets your professional duties. That remains for your firm and your regulator's guidance.
  • AI output can be wrong or incomplete and does not replace a lawyer's review.

This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.

Read next

Client confidentiality and AI tools: a checklist