Will your AI vendor train on client data? Reading the terms
By the ChamberOne team 5 minute read
The question comes up in nearly every firm that tries an AI tool: if we paste a client's document in here, does the company use it to train its models? The honest answer is that it depends on the product, the plan and the contract. A free app on a lawyer's phone and an enterprise agreement negotiated by the firm can sit on the same underlying model and treat your data entirely differently.
Key points
- Training and retention depend on the plan and the contract, so read the terms for the exact product.
- Consumer terms are often looser than business terms; keep client data out of personal accounts.
- Ask about retention, human access, sub-processors and deletion, and get the answers in writing.
- Check your own duties and client requirements before any upload, and keep an approved tool list.
What 'training' means, and why people worry
A model is trained on large amounts of text before it is released. Some providers also use what customers type in, along with the model's replies, to improve later versions. The worry for a law firm is that privileged or confidential material could end up shaping a model that other people use, and might in rare cases be repeated back in some form.
How real the risk is depends on the provider's practices, which you cannot see directly. What you can see is the contract. So the sensible approach is to treat the terms, not the marketing, as the source of truth.
Consumer terms versus business terms
The same company often publishes several sets of terms. Consumer or free-tier terms may allow review of conversations, use for improvement of the service, and longer retention. Business, enterprise or API terms commonly say that customer content is not used for training by default, with shorter and more clearly defined retention.
That is a pattern, not a rule. Read the actual wording for the plan you are buying. A lawyer using a personal account for client work is the single most common way client data ends up under terms the firm never reviewed. Your policy should say that client material goes only into tools the firm has approved.
- May allow review of conversations
- May allow use to improve the service
- May mean longer retention
- Commonly say content is not used for training by default
- Shorter, more clearly defined retention
- Read the actual wording for your plan
The questions to put to a vendor
When you evaluate a tool, whether it is a standalone assistant or a feature inside your practice software, ask these questions and ask for the answers in writing.
- 1
Is customer content used to train or improve any model, yours or a third party's? Is that off by default, and is it stated in the contract?
- 2
Which AI providers does the product use, and under what terms with each of them?
- 3
How long is customer content kept by you and by the provider? Is there any retention for abuse monitoring, and who can see it?
- 4
Where is content processed and stored, by country?
- 5
Can a human at the vendor or provider read our content, and in what circumstances?
- 6
What happens to our content when we leave?
- 7
Will you tell us before you change providers or sub-processors?
Retention windows and the abuse-monitoring carve-out
Even where a provider does not train on business data, it may keep inputs and outputs for a period, often to detect misuse. That retained text is still your client's information sitting with a third party. Ask how long the window is, whether it can be shortened or switched off for your account, and who has access during the window.
Some providers offer zero-retention arrangements for eligible customers. If a matter is sensitive enough, ask whether that is available, and whether it applies to the particular route through which your software vendor reaches the model.
Opt-outs and defaults
A setting that you can switch off is only as good as the default and the person who controls it. Find out whether the opt-out is applied at firm level or left to each user, whether a user can reverse it, and whether it covers all the vendor's products or just one. If it exists only as a toggle in a consumer app, it is not a control you can rely on for a firm.
Where possible, prefer products that let an administrator set the position for the whole firm and that record who changed it and when.
What to get in writing
- A contract or data processing agreement that states the no-training position in the contract itself, not only on a web page that can be edited.
- A list of sub-processors and the countries where data is handled, with a duty to notify you of changes.
- The retention period for inputs and outputs, and what is retained by the underlying model provider.
- A deletion commitment on termination, with a time limit.
- Breach notification terms, including how quickly you will be told.
- Audit or information rights, even if limited to reasonable requests for documents.
Your own duties come first
Vendor terms are only half of the question. Your duties to clients, your engagement letters, any client instructions about technology, and your regulator's guidance all affect what you may send to a third party. Some clients, particularly large corporate and government ones, restrict use of AI tools in their outside counsel guidelines. Check the guidance from your regulator or bar, and read the client's own requirements before you upload anything.
Consider levels. Public information and your own templates are one category. Ordinary client documents are another. Highly sensitive matters, such as those involving national security, health, or restricted material, may be kept out of cloud AI tools altogether.
A short approval routine
Appoint one partner and one IT lead to approve tools. Have them complete the questions above for each product, keep the answers on file, and review them each year or when the vendor announces a change. Publish an approved list to everyone in the firm, with a simple instruction: if a tool is not on the list, client material does not go into it. That one rule closes most of the gap between what the firm has approved and what people actually do.
Keep the approval records simple enough that someone else can pick them up. A single sheet per tool, with the date, the answers, the name of the approver and the next review date, is far better than a long report nobody reads.
AI settings in ChamberOne
- Choose the model for each AI feature, GPT-5 or Claude, or switch AI features off for the firm.
- Every AI call is recorded in the audit log with the model and token counts.
- 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 (OpenAI or Anthropic), outside the country where ChamberOne is hosted. With AI features off, nothing is sent.
- It does not make a vendor's terms safe for you. A partner or IT lead still has to read them against the firm's own duties.
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
Australia's new statutory tort for serious invasions of privacy
The statutory tort for serious invasions of privacy commenced on 10 June 2025 under Schedule 2 of the Privacy Act 1988 (Cth). This article explains the five elements, defences, remedies, time limits and what the reform means for practitioners.
6 minute readClient confidentiality and AI tools: a checklist before you paste anything in
The questions to answer about retention, training use, access, hosting and sub-processors before client information goes into any AI tool.
5 minute readShadow AI in law firms: staff are already using it, so what next?
Why bans fail, how to find out what is being used, and how to offer an approved alternative with clear rules.
5 minute read