Blog · AI for small and midsize businesses
What to ask an AI vendor about retention, training and data access
Vendor assurances become useful when they describe the exact product and data route your business will use. Build an evidence checklist, test observable settings with synthetic records and record unresolved questions before approving a particular workflow or category of information.

In this article
“We do not train on your data” answers only one question about how your data is used. A business also needs to know what is retained, who can access it, which services process it and whether deletion reaches all relevant copies. Evaluate the exact product, plan and configuration you will use.
Draw the route of one business record
Consider a hypothetical support team summarizing customer complaints. The record may pass through a connector, application server, model provider, logging service and staff dashboard. A promise about the model provider alone does not describe the complete route.
Ask the vendor to identify the relevant services and responsibilities. Separate the company’s settings from provider defaults, optional features and contractual commitments. Record the evidence date because products and terms change. An unanswered question should remain open rather than be converted into an assumption.
| Question | Evidence to request |
|---|---|
| Is content used for training? | Terms for the exact service and plan, plus applicable settings |
| What is retained and why? | Input, output, diagnostic and backup retention details |
| Who can access records? | Administrative, support and user-access controls |
| Who else processes the data? | Relevant subprocessor and integration information |
| How does deletion work? | Procedure, scope, timing and documented exceptions |
Match evidence to the intended use
A public-content drafting task and confidential employee-record analysis require different assessments. The data owner defines the proposed information categories and purpose. The system owner reviews technical access and configuration. Contractual and privacy questions go to the company’s responsible advisers.
Avoid comparing broad vendor reputations instead of service-specific evidence. A consumer interface, business workspace and developer service may operate under different arrangements. Do not transfer an assurance from one to another without verification.
Verify the controls you can observe
Use synthetic records to check sharing defaults, export access, deletion behavior visible to the customer and whether sensitive content appears in logs. Record the limits of the test: deleting a visible conversation does not independently prove what happened inside a provider’s backups.
Ask how the company will be notified of material changes and how it can export or stop using the service. Identify who reviews updates after purchase. Vendor due diligence is incomplete if the evidence is filed once and nobody owns the continuing decision.
Conclude with a short decision record: accepted use, excluded data, required settings, unresolved questions and review triggers. This is operational evidence for an informed choice, not a generic certificate of privacy or legal compliance.
Assign a status to each vendor question: documented, configuration-dependent, awaiting clarification or unacceptable for the proposed use. Keep the answer and its supporting evidence together. If the intended workflow changes from public drafting to customer-record analysis, reopen the assessment instead of reusing the previous approval merely because the vendor is the same.
Bring one representative workflow and its current exceptions to discuss a practical AI implementation.