What Happens to Customer Data Inside an AI Workflow?

Kevin works in customer service at an apparel decorator. During a vendor demo, he asked a reasonable question. "When the agent reads a customer's purchase order, where does the PO go?" The sales rep smiled and said, "The cloud."

Linda, his manager, asked which cloud. The rep said, "The secure one." Linda wrote "secure cloud" on her notepad, underlined it twice, and drew a small cloud with a question mark inside. Kevin later admitted the cloud drawing was the most honest part of the meeting. Nobody was hiding anything, but nobody in the room could describe the path either.

Customer data in an AI workflow can pass through connectors, model services, storage, logs, and support tools before it reaches its destination. The real path depends on the vendor and the setup. Before approving a workflow, ask the vendor to map each stop and tell you what is kept, who can see it, and when it is deleted.

Start with one real transaction

Pick a single, ordinary transaction and follow it. For a promo distributor, that might be a customer PO arriving as a PDF attached to an email, a common starting point for order entry automation. A connector reads the message. Some extraction step pulls fields out of the PDF. A model helps interpret the request. A tool writes a draft order to the ERP, and a log records what happened.

That sequence is a common shape, not a description of any particular vendor. Ask the vendor to replace it with their actual components for your workflow. For each stop, you want four answers: what data arrives, where it is processed, whether a copy stays behind, and what causes that copy to be deleted.

The six stops to map

1Sources

The source is where data starts, such as a shared inbox, an ERP, a CRM, or a file share. Ask which mailboxes, folders, or record types the workflow can read. Ask whether it reads everything it has access to or only what the job needs.

2Connectors

Connectors are the credentials and integrations that let the workflow reach your systems. Ask what permissions each connector holds, whether it can read only or also write and send, and who on your side can revoke it. Ask whether revoking access also removes data already copied, or only stops future access.

3Models

The model step is where text and documents get interpreted. Ask which model providers are used, including any separate service for reading documents or images. Ask what exactly is sent to the model, under which terms, and whether the provider can use that data for training or keep it for abuse monitoring.

4Subprocessors

Most workflows rely on several outside companies for hosting, storage, email, monitoring, and support. Ask for the current subprocessor list, what each one does, where it processes data, and how you will hear about changes. The list for a vendor's website may differ from the list for its production workflows.

5Logs and indirect copies

Logs often hold more than people expect. Error traces, screenshots, cached tool responses, test sets, and support tickets can all contain business data. Ask the vendor to show you a successful run and a failed run, and to point out every place your data appears in each.

6Retention and deletion

Retention can differ for working data, source files, logs, backups, and support records. Ask for the period for each category. Ask what happens when you delete one workflow, when you revoke a connector, and when the contract ends, because those three events may have different effects.

Keep the questions separate

Buyers often bundle everything into one question, which is whether their data is private. That question has several parts, and a yes to one part does not answer the others. A promise not to train models on your data says nothing about how long logs are kept. A short retention period says nothing about which subprocessors see the data first.

Ask about training, retention, location, human access, and isolation as separate items. Get written answers for each, and check that they match the contract you will actually sign. Answers given for a different product tier or a consumer version of a tool do not carry over by assumption.

Buyer question list

Bring this list to the vendor. Record an answer, an owner, or "unknown" for each line.

• Which sources can the workflow read, and which fields does it actually need?
• What permissions does each connector hold, and who can revoke them?
• Does revoking a connector delete stored copies or only block future access?
• Which model providers receive data, and what exactly is sent to them?
• Can any provider use our data for training or model improvement?
• Does any provider keep inputs or outputs for monitoring or troubleshooting, and for how long?
• Who are the subprocessors for this workflow, and where do they process data?
• How are we told when the subprocessor list changes?
• What separates our data from other customers' data in storage, processing, and logs?
• Which vendor staff can see our production data, and how is that access approved and recorded?
• What appears in logs for a successful run and a failed run?
• What is the retention period for working data, source files, logs, backups, and support records?
• What happens to our data when a workflow is deleted, and when the contract ends?
• What evidence will we receive once deletion is complete?

Trelium builds AI agents for distributors and promo and apparel companies, handling work like order entry and invoice matching. If you are evaluating Trelium, run this question list with us and ask for the answers in writing for your specific workflow. Compare those answers with the agreement you would sign.

FAQ

1Where does customer data go in an AI workflow?

It typically moves from a source system through a connector, into an extraction or model step, and then into a destination system like an ERP. Along the way, copies may land in logs, storage, and support tools. The exact path depends on the vendor, so ask for a map of your specific workflow.

2Does a no training promise mean my data is not stored?

No. A commitment not to train models on your data covers one use of that data. It does not, on its own, tell you how long inputs, outputs, or logs are kept, or which subprocessors handle them. Ask about training and retention as separate questions.

3What is a subprocessor in an AI workflow?

A subprocessor is an outside company the vendor uses to process your data, such as a hosting provider, a model provider, or a support tool. Each one is another place your data may travel. Ask for the list, what each one does, where it operates, and how changes are announced.

4What happens to my data when I revoke a connector?

It depends on the vendor. Revoking a connector usually stops future access, but data already copied into storage, logs, or backups may follow a separate retention schedule. Ask the vendor to explain both effects in writing.

5How can I reduce the data an AI workflow needs?

Define the smallest set of records and fields the job requires, then limit connector permissions to match. For example, an order status workflow may need an order number and tracking data but not full purchase history. Less data in the workflow means fewer copies to track and delete.