Eric Kuvykin

Reviewed October 17, 2026. This guide replaces an earlier general AI overview. The original publication date and URL are retained.

A vendor’s promise not to train on business data does not answer how long prompts, files or support logs remain accessible. Review the complete data flow before connecting an AI tool to merchant records.

Begin with the smallest useful input

For an AI assistant that drafts a payment-support reply, identify the information needed to explain the approved process. A sample order reference, status and redacted error message may suffice for a test; a complete customer history or bank statement may not. Keep payment credentials, passwords and unrelated personal information out of general-purpose prompts. Do not assume that a token or a removed name makes every remaining detail anonymous. Start evaluation with synthetic records and have the responsible security and privacy reviewers approve any production data flow.

Ask three different questions about the same file

Model training concerns whether the provider uses submitted material to develop or improve models. Retention concerns whether it continues to store material for another purpose, such as service delivery, security or support. Deletion concerns what happens when an authorized person requests removal or the relationship ends. Obtain the applicable terms for the specific product, account type and integration. A statement about one purpose should not be treated as a promise about all three.

Trace copies beyond the conversation window

Map where the input, generated output and diagnostic records can go: the merchant application, connector, AI service, saved workspace, support system and any relevant downstream service. Ask which people or services can access each copy, what is retained, and which exceptions or backup schedules affect removal. Record the provider’s answer and its date rather than inventing a retention period. Deleting a visible conversation is not sufficient evidence that every other copy has been removed.

Align vendor terms with the business’s own commitments

The FTC’s discussion of AI privacy commitments warns companies to honor representations about customer information, including its use for training. Its business data-security guide recommends limiting collected information, controlling access and documenting retention. Have qualified reviewers assess whether a proposed use fits existing customer promises and applicable obligations. A vendor setting alone does not establish that the merchant has permission to upload a particular record.

Make exit behavior part of the adoption decision

Before deployment, identify who can revoke connector access, request deletion, export necessary business records and confirm that obsolete copies are handled under the approved policy. Use a synthetic test record to observe the documented user-facing deletion process; ask the provider to explain backend handling that the merchant cannot independently inspect. Preserve the resulting evidence without claiming the test proves complete erasure from every system. Revisit the approval when the product, connector or terms materially change. If the data flow cannot be explained, keep that production use paused while the gap is resolved.

Related reading

Data handling and malicious document instructions are separate risks. Continue with the AI invoice prompt-injection guide for the latter, or follow the AMS New Jersey merchant reading path for related operating decisions.