Can AI vendors use our client data to train their models?

Can AI vendors use client data to train their models?

For law firms, accountancy practices, consultancies and regulated teams, the safe answer is no unless the vendor can prove the opposite in writing. The issue is not whether an AI tool sounds useful. It is whether client data, matter context, board papers, audit files or regulated advice can be retained, reviewed or used to improve a model outside your control.

The source question asks whether customer data is used to train, fine-tune or improve AI models. That is exactly the right question to ask before a firm lets an assistant, search tool, knowledge base or workflow product touch professional data.

The safest answer is contractual, technical and evidenced

You should only permit client data in an AI vendor system when three things line up: the contract prohibits model training without explicit permission, the product settings enforce that position, and the vendor can provide evidence through security, privacy and audit documentation. A sales answer on its own is not enough.

If the vendor says data is not used for training, ask what counts as training, fine-tuning, model evaluation, abuse monitoring, human review, product analytics and service improvement. Those categories often sit in different parts of a data processing addendum, privacy notice and support policy.

Check AI vendor risk before client data enters the tool

What to ask the vendor before entering client data

Use plain questions that map to real controls:

• Will any customer, client or end-user data be used to train, fine-tune or improve foundation models?
• Is data used for evaluation, quality review, abuse monitoring or product analytics?
• Can model-training use be disabled by default at tenant, workspace and user level?
• Can the opt-out be written into the contract, not only selected in an admin screen?
• How long are prompts, files, outputs, logs and embeddings retained?
• Can the firm delete data and prove deletion has propagated to backups or derived stores where relevant?
• Which subprocessors can access the data and in which jurisdictions?
• Will support staff or human reviewers see client content?
• What audit logs can the firm export for regulator, client or insurer questions?

Why professional services firms need a higher bar

Professional services firms are not testing AI with harmless sample text. They may hold privileged material, regulated advice, special category data, transaction documents, claims files, audit evidence or commercially sensitive strategy. A weak vendor answer can create a confidentiality, data protection, insurance and client-trust problem.

For legal teams, connect the review to confidentiality, privilege and client disclosure obligations. For financial services, connect it to Consumer Duty, SM&CR accountability, outsourcing governance and record keeping. For insurance and accountancy, connect it to claims sensitivity, audit trail, quality review and professional standards.

How to turn the answer into an internal rule

Once the vendor response is clear, translate it into an internal rule that staff can actually follow. Do not leave people to interpret a 30-page data processing addendum each time they use a tool.

• Approved for public or synthetic data: suitable for demos, templates and non-client examples.
• Approved for client data with controls: allowed only in named workflows, with training disabled, retention understood and audit logs enabled.
• Not approved for client data: useful for general drafting or research, but no confidential, personal, privileged or regulated data may be entered.
• Blocked: vendor cannot answer training, retention, security or subprocessor questions to the required standard.

Build an AI governance operating model that staff can use

Red flags in vendor answers

Be cautious when the vendor says data is not used for training but reserves broad rights for product improvement, evaluation or monitoring without explaining what data is involved. Also watch for opt-out controls that are available only on enterprise plans, undocumented admin settings, vague deletion commitments, missing subprocessor lists, or support policies that allow human review of content without a tight process.

A strong answer is specific. It names the data categories, retention period, training position, opt-out mechanism, subprocessors, security controls, deletion route and audit evidence. It also confirms whether the same rules apply to prompts, uploaded files, generated outputs, logs, embeddings and support tickets.

Conclusion

Client data should only enter an AI vendor system after the firm has a written training position, a technical opt-out, retention limits and evidence it can keep. If any of those are missing, treat the tool as suitable for low-risk content only until the gap is closed.

Implement AI tools with the right data controls from the start