AI provider on the vendor list
Should an AI or model provider go on your vendor list?
Put the provider on the list when you send it customer data. Write what it does, then check training, retention, region, and the provider's own data terms before you answer.
This is a practical note for a security review, not legal advice.
When a model provider counts
A subprocessor, sometimes written sub-processor, is a company that processes personal data for you. An AI or model provider counts when customer data is in what you send: prompts, files, tickets, or names.
A personal chat account you use with no customer data does not belong on the list. A feature in your product that calls the model does, even when the customer never sees that brand.
The company name is the row. If the reviewer also asks which model, name the one you call in the purpose, or in the notes you keep. Do not write AI as the company.
What reviewers usually ask
These answers live on the provider's own page for the plan you pay for. Read that page. A consumer chat plan and an API plan often differ, and a blog post can be out of date. Write what you read in your own notes.
- What data goes to itPrompts, files, customer names, support tickets, or only your own staff notes. If customer data never reaches the model, leave the provider off the list.
- TrainingWhether they use the data you send to train or improve their models. Check the current terms. Do not fill this in from memory.
- RetentionHow long they keep that data, and whether your plan can turn keeping it off. The answer is on their page.
- RegionWhether they name a country or region, or whether you selected one in their console. If the page does not name one place, leave location blank.
- Their data termsThe public data processing agreement, or the public terms page if that is what they publish. Link that page. A privacy policy is not the same thing. If a privacy page is all they publish, link it and say what it is.
When the page says global, names several regions, or names none, what to write is in the processing location guide.
How long you keep customer data, and what to write when a vendor will not name a period, is in the data retention guide.
Where to open their terms
Start at the provider's own page for the product you actually call. Open it, and read training, retention, and region there. This guide does not restate those policies.
- OpenAIData processing addendum https://openai.com/policies/data-processing-addendum/
- AnthropicData processing addendum https://www.anthropic.com/legal/data-processing-addendum
- Google Gemini APIGemini API terms https://ai.google.dev/gemini-api/terms
- Google CloudData processing addendum, when the calls go through Google Cloud https://cloud.google.com/terms/data-processing-addendum
Gemini in the Gemini API and Gemini through Google Cloud are different paths, so they can have different terms. Link the page that covers your plan. If they will only share the agreement after you upgrade, write that down and still link the public page they did publish.
An example row
This row is an example, not your list. Replace the company if you do not call OpenAI. Location is blank on purpose. Fill it only after their page names one place, or after you have selected a region. Change the date to the day you open the link.
The columns are the ones the list already uses: name, what they do, location, agreement link, and the date you last checked the row.
| Name | What they do | Location | Agreement link | Updated |
|---|---|---|---|---|
| OpenAI | Model API for customer prompts | https://openai.com/policies/data-processing-addendum/ | Oct 6, 2026 |
List each provider in the data path
- Add a row for every provider that receives customer data, including one you reach only through another tool. A help desk, a search box, or a note-taker that sends customer text to a model is still a path.
- Say which tool sends the data, in the purpose. Model calls made by the help desk is enough.
- If you cannot tell which company receives the data, ask the tool vendor. Leave the model row off until you have a name.
- OpenAI, Anthropic, and Google are three rows when all three see customer data. A provider you tried and then removed does not stay on the list.
A reviewer may open that provider's own list of companies and ask about a firm named there. Read that list on the provider's site, and keep what you learn in your review notes. Your list names the providers you chose. It does not copy theirs.
What the row can hold
The row has five fields. The review questions that do not fit stay in the notes you keep.
- NameThe provider the reviewer will look up. Use the company name on the agreement.
- What they doWhat you send and why. Model API for customer prompts is enough. Skip a slogan.
- LocationA country or region you have read on their page, or the region you selected. Blank is honest when they do not name one place.
- Agreement linkTheir public data terms. Open the link once before you send the list.
- UpdatedThe date you last checked the row. Change it when you check again.
There is no column for the kind of data, and no field for training or retention. You can save the row while the location or the link is still blank.
If you type OpenAI, the editor can suggest that public agreement link. It does not suggest a location. If you type Google Cloud, it can suggest the Google Cloud agreement link, and no location. Nothing is added until you use it. A suggestion is not a reading of their training or retention terms. Anthropic and Gemini are not suggested. You type those rows yourself.
Trust Prep does not watch a provider's own list of companies, tell you when that list changes, or write a transfer assessment. It does not decide whether sending the data is allowed.
If you want the longer notes
Why reviewers ask for the list is in the vendor list guide.
A table you can copy, with more example rows, is in the subprocessor list template.
How to tell a controller, a processor, and a subprocessor apart is in the controller and processor guide.
When the agreement uses a legal name and the list uses the brand, how to write the row is in the company names guide.
What to write when they ask if that provider encrypts customer data, and who manages the keys, is in the encryption guide.
What to write when they ask what data that provider can see, and what does not reach it, is in the data access guide.
How your product keeps one customer's data out of another's, including a shared model or cache, is a different row from naming the model company. What to write is in the customer data isolation guide.
Whether your own product uses customer data to train, fine-tune, or evaluate a model is a different sentence from naming the model company. What to write is in the customer data training guide.
What the agent is allowed to do, and which of those actions a person confirms, is a different sentence from naming the model company. What to write is in the agent permissions guide.
Related guides
Optional next step: a living page
The notes and the example row are enough for the review in front of you.
If you want one link for the list, Trust Prep publishes that vendor list as a page you edit when a company changes. You can name the model provider, say what it does, record a location when you know one, and link that provider's public data terms. Reviewers open a public link made from your company name, without an account. They can copy the table or download the spreadsheet there. You can fill in the list and preview the page before you pay to publish the link. Publishing turns on after we confirm your payment.
The page does not store a training answer or a retention answer, and it has no column for the kind of data. It does not watch a provider's own list of companies, notice you when that list changes, or write a transfer assessment.