Where Your Data Goes When You Hire an AI Agent (and How to Delete It)

When you use an AI agent, your data typically moves through the vendor's systems, the AI model provider and any connected tools, and it's kept according to the vendor's retention rules. Before you connect anything, ask for the list of sub-processors, how long each kind of data is kept, how deletion works and whether your data trains models.

An AI agent is useful because it can see your orders, customers and conversations, which is also why privacy matters. This guide traces where that data tends to go, shows how Chargezen handles each part, and gives you questions for any vendor. It's general information, not legal advice. For GDPR, CCPA or other rules that apply to your business, talk to your lawyer.

A sub-processor is an outside company that handles data on a vendor's behalf. Any serious vendor should publish a list. Chargezen's sub-processor list names each company and what it does, including AWS and Google Cloud for hosting, Telnyx for SMS and voice, Resend for email, and OpenAI and Anthropic for AI.

If your company already has data terms with an AI provider, you may be able to use them. Concierge lets you run agents on your own account with providers such as OpenAI, Anthropic, Google Gemini or AWS Bedrock.

This is the first question many buyers ask, and you should get a direct answer in writing. Ask the agent vendor about their own use, and ask what terms apply to each model provider they send data to. API terms, consumer chat apps and enterprise agreements can all differ, so don't assume the answer for one applies to another. If you run on your own provider account, read your agreement with that provider.

Retention should be published by data type, not given as a single vague line. Here are some examples from Chargezen's data retention schedule:

Deletion means different things in different places, so ask about each one. Here's how it works in Concierge:

• Learned memories: Forget erases a memory's text for good, and no agent, page or export can use it again. Archive is the reversible choice.
• Your personal memories: Under Only me, you can export them as a file, or use Forget all my memories.
• Call recordings: They are deleted automatically after the period you set in your recording policy, unless placed on legal hold.
• Shopify customers: When Shopify sends a customer data erasure request for your store, Chargezen processes it.
• Any other request: Email privacy@chargezen.com.

Under the GDPR, people have a right to erasure in certain cases (Article 17), and a company using a processor needs a contract covering that processing (Article 28). Chargezen's data processing agreement and privacy policy set out how we handle this. Some records may be kept longer where the law requires it, such as tax records.

An agent that learns can accidentally keep personal details, so look for rules that prevent this. In Concierge, learned memories describe how your business works and never name a customer:

• Before a model reads a handoff, the customer, their orders and the teammates involved are replaced with placeholders.
• Anything learned automatically is refused if it contains an email address, phone number, order number, tracking number or a link with a token.
• Someone writing a memory by hand confirms it names no customer and holds no private information.
• Each memory has an audience: customer-facing, team or only you.

Customer history is kept separately, so customers don't have to repeat themselves. See memory vs playbooks.

• Give only the access the job needs: In Concierge, each agent has its own capabilities and channels. This is called least privilege.
• Verify customers: Before an agent shares private order details or changes an account, the customer confirms a sign-in code.
• Keep job data short-lived: A Concierge job computer gets a read-only copy of the data it needs, and that copy is deleted when the job ends.
• Keep passwords out of chat: For a site that needs a login, you sign in yourself on a separate page. Your password doesn't go into the conversation or the agent's memory.
• Keep web searches clean: Concierge's web tools never send customer personal details in a search.

Our security checklist covers access in more depth. Also tell your team what not to paste in. A policy on unapproved tools helps here too; see shadow AI.

• Update your privacy policy to cover AI processing, and list the kinds of providers involved.
• Say when customers are talking to an AI where the law requires it, and consider doing so regardless.
• Tell callers when calls are recorded.
• Make sure your team knows how to handle a deletion request.

• Which sub-processors handle my data, and where do they operate?
• Does my data train any model, yours or a provider's? Can I see that in writing?
• How long do you keep each kind of data, and can I change any of those periods?
• How do I delete one customer's data? How do I delete everything when I leave?
• What does the agent keep in memory, and can I see and erase it?
• Can I use my own model provider account?
• Which security certifications do you hold, and can I read the report?
• Will you sign a data processing agreement?
• Who at your company can read my conversations, and is that logged?

Chargezen is SOC 2 Type II certified and supports single sign-on with SAML 2.0. You can review the details in our Trust Center and on the security page.

• Write down which data each agent needs, and remove what it doesn't.
• Read the vendor's sub-processor list and retention schedule.
• Sign a data processing agreement if the rules that apply to you call for one.
• Set your call recording policy before you turn on phone answering.
• Test a deletion request from start to finish, so you know it works.
• Decide who on your team reviews what agents learn.

Concierge's trust and privacy controls include access per agent, a private computer for each job, memory you can see and erase, and recording rules for each number. The security docs show where each control lives in the admin.

Keep reading: is it safe to connect an AI agent to your store?, how to connect your tools and keeping a human in the loop. Glossary: AI governance, agent sandbox and least privilege access.

That depends on the vendor and your setup. Check their security certifications, sub-processors, retention periods and deletion process, and limit the agent's access to what its job needs. No vendor's general promise replaces reading those details.

It varies by vendor and by model provider, and the terms can differ between API, consumer and enterprise products. Ask for the answer in writing, covering both the vendor and each provider it sends data to.

Ask the vendor how deletion works for conversations, memory, files and recordings. With Chargezen, Shopify customer erasure requests are processed when Shopify sends them, and any other deletion request goes to privacy@chargezen.com.

Compliance depends on how you and the vendor handle the data, so no product makes you compliant on its own. Look for a data processing agreement, a sub-processor list, clear retention and a working deletion process, and ask your lawyer to review them.