Short answer: For an agency, secure AI is a tool you could defend to a regulator as a vendor choice: it does not train on client information, you know what it keeps and for how long, you know who at the vendor can read it, and the contract says so. Most business plans from OpenAI, Anthropic, Microsoft and Google now meet the first test by default. They differ on the other three, which is where an agency's vendor due diligence has to focus.
Why "secure" means something specific for an agency
An agency's AI questions come from the same place as its other vendor questions. Under the NAIC Insurance Data Security Model Law (#668), a vendor that can access your clients' nonpublic information is a third-party service provider, and your agency has to use due diligence in choosing it and require it to protect that information. Many states have adopted a version of the model, and the NAIC keeps a chart of which ones. Agencies with fewer than 10 employees are exempt from that program section in the model, though not from breach notice; see the small-agency page. The details, and which states, are on the rules page and the vendor due diligence page.
So the practical test is whether you can answer four questions about a vendor, in writing, from its own terms.
The four questions
1. Does it train on what we put in? A personal account may. OpenAI says that on ChatGPT Free, Plus and Pro, "data sharing is enabled by default" (OpenAI). A producer pasting a claims history into a personal account is the case to stop first.
2. What does it keep, and for how long? Business plans keep chats so people can return to them. That is useful, and it also means the vendor holds your clients' information until someone deletes it.
3. Who at the vendor can see it? Some vendors spell this out. OpenAI names "authorized employees" and "specialized third-party contractors" for ChatGPT Business (OpenAI). If a vendor's pages are silent, ask.
4. Will the contract say it? A web page can change. Your vendor file should hold the data processing terms you actually signed.
How the common options answer them
Each vendor fact below comes from that vendor's own page, read on November 13, 2026. No ratings.
| Tool | Trains on your data? | What it keeps | Who sets retention |
|---|---|---|---|
| ChatGPT Business | "We don't train our models on your organization's data by default." (OpenAI) | Chats; "deleted or unsaved conversations are removed from our systems within 30 days", with exceptions | "Your workspace admins can control how long your data is retained." |
| ChatGPT Enterprise | Same default | Chats; deleted ones "removed from our systems within 30 days, unless we are legally required to retain them" | "Your workspace admins control how long your data is retained." |
| Claude Team | "No model training on your content by default" | "we retain your chats and coding sessions in the product" (Anthropic) | Users delete chats; custom retention is a Claude Enterprise control |
| Claude Enterprise | Same default | Chats and projects; "by default, projects are retained indefinitely" | Admins, with "The minimum retention period is 30 days" |
| Microsoft 365 Copilot | "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs" | "The stored data includes the user's prompt and Copilot's response" | Admins, through Microsoft Purview |
| Gemini in Google Workspace | "Your content is not human reviewed or otherwise used for Generative AI model training outside your domain without permission." | Prompts and responses: "90 days to indefinite, as determined by admins" | Admins |
| Faradex | Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them | Files and conversations are held in memory and are gone when the Workspace is reset | Content is gone when the Workspace is reset |
Prices for most of these plans are on What a small agency can buy today.
What to put in your vendor file
Whichever tool you choose, keep a short record: the plan name, the data terms page and date you read it, the retention setting your admin chose, who at your agency may use it and for what kinds of client information. If you are in New York, the third-party policy in 23 NYCRR 500.11 applies even to agencies with the limited small-business exemption. The questionnaire on the due diligence page gives you the questions to send.
Health information raises one more question, the business associate agreement. That is covered on HIPAA and AI for benefits brokers.
General information for agencies, not legal or compliance advice. Vendor terms change, so read them again before you sign.