Does AI use client data to train models? The answer can depend on the specific product, account type, settings, and contract. For a financial services firm, a general statement on a provider’s website is not enough to determine how a particular deployment handles confidential information.
Start by separating model training from other data practices. A provider may say it does not train on prompts while retaining them for a period, using them for service improvement, or allowing limited review for security purposes. These are different questions, and each deserves a specific answer.
The U.S. Treasury has identified privacy and third-party-provider risks in financial-sector AI use, so review both the tool and the wider data path before sharing sensitive information. Read the Treasury’s summary of AI opportunities and risks in financial services.
Ask what “training” means for your account
Ask the provider to address each type of information you might submit: prompts, uploaded documents, generated outputs, and feedback such as ratings or corrections. Then ask whether any of that information is used to train or fine-tune a model, or to improve a product or service.
- Does the answer apply to prompts, files, outputs, and feedback, or only some of them?
- Can any of this information be used to train or fine-tune models operated by the provider or another company?
- Does the answer apply to this exact product, account tier, and configuration?
- Can an administrator change the data-use setting, and does it apply to every user and feature in the account?
- What happens if the firm changes its settings, subscription, or product plan?
Ask for a written answer tied to the product and account the firm plans to use. If a provider says that training is off by default or that customers can opt out, confirm who controls that setting, how it is applied, and whether it covers all relevant data types.

Check retention and access separately
Turning off model training does not necessarily answer whether prompts or files are retained. Ask how long each data type is stored, what purpose the retention serves, and whether the firm can set a shorter period or request deletion. Ask separately about backups, termination of the service, and any limits on deleting information from those backups.
Also ask who can access submitted content. The provider should explain whether access by staff is possible, under what circumstances, and what controls apply. Ask whether information may be reviewed for support, troubleshooting, security, or abuse monitoring, and whether those processes differ from model training.
- How long are prompts, uploaded files, and outputs retained?
- Can the provider delete them at the firm’s request, including after the account ends?
- Are backups handled differently, and what deletion timeline applies to them?
- Can provider personnel or support teams view client information? If so, when and with what restrictions?
- Are data practices different for security monitoring or product support than for model improvement?
Do not treat “not used for training” as a complete privacy commitment. Record the provider’s answers about training, retention, human access, and deletion as separate controls.
Map the full data path
An AI feature may rely on a model provider, subprocessors, connected business applications, or other service providers. Ask the vendor to identify which parties may receive or access the firm’s information, what each party does, and which data-use and retention commitments apply to them.
For tools embedded in existing business software, review the terms for the specific AI feature as well as the main product. Confirm whether connected applications send content to the AI feature and whether the feature’s settings match the protections the firm expects. A privacy notice for the broader platform may not resolve every question about a particular feature or configuration.
When evaluating an embedded Microsoft 365 feature, for example, use a product-specific review alongside the firm’s own policies. This Microsoft 365 Copilot safe-use guide can help teams consider how employees should handle information in that environment.
Compare written terms, settings, and contract language
Check the provider’s privacy notice, account-level settings, data-processing terms, and signed contract. These sources can describe different parts of the arrangement, so ask the vendor to confirm which terms apply to the firm’s product and deployment. If documents appear to conflict, request a written explanation of which document controls.
Before approval, have the firm’s appropriate compliance, legal, security, and procurement reviewers assess the commitments. The goal is not simply to collect a reassuring answer. It is to confirm that the answer is specific, documented, and consistent with the firm’s policies and obligations.
For the contract or other written terms, confirm that they clearly address:
- Whether prompts, files, outputs, or feedback may be used for training or service improvement.
- Retention periods, deletion requests, backups, and what happens when service ends.
- Subprocessors and model providers that may handle or access the information.
- Human access, support, and security-monitoring processes.
- Notice and approval processes if the provider changes its data practices or relevant subprocessors.
- How the firm can raise concerns or verify that the agreed controls are in place.
These are review points, not substitute contract language. Ask qualified internal reviewers to determine what terms are appropriate for the firm and its use case.
Use a cautious go/no-go process
Until the firm has clear, written answers about use, retention, deletion, and third parties, do not submit identifiable client information. Use synthetic or otherwise non-sensitive test material while evaluating a tool, and follow the firm’s own policies for deciding what information is appropriate to share.
Keep a record of the approved product, account settings, vendor responses, and contract terms. Revisit that record if the provider changes a feature, account configuration, or data-use terms. A structured AI vendor comparison can also help teams assess providers against consistent criteria rather than relying on broad assurances.
Frequently asked questions
Does turning off model training stop an AI provider from retaining prompts?
Not necessarily. Training and retention are separate data-handling questions. Ask how long prompts are retained, why they are retained, who may access them, and whether deletion is available.
Are data-use settings different across business, enterprise, and free accounts?
They may differ by product and account configuration. Ask the provider to confirm the rules and settings that apply to the exact account the firm will use, and get the answer in writing.
Can a vendor’s subcontractors or model providers access client information?
Possibly, depending on the service design. Ask the vendor to identify relevant subprocessors and model providers, explain their role, and confirm which protections and retention commitments apply.
What if the provider cannot give a specific written answer?
Do not submit identifiable client data until the firm’s compliance and legal reviewers have evaluated the uncertainty and approved a path forward.
Make data-use review part of AI approval
For financial services firms, checking model training is only one part of assessing an AI tool. Confirm retention, access, deletion, embedded features, and third-party handling as well, then compare the written commitments with the settings and contract that govern the firm’s account.
AI Services Near Me serves businesses throughout Florida and Georgia and offers a free AI consultation. If your firm is assessing an AI deployment, contact the team at (904) 658-0777 to discuss your review process.