Your AI Vendors Are Data Processors: The DPA Homework Behind Platform App Reviews

Meta's App Review includes a set of Data Handling Questions, and one of them is blunt: name every outside company that can access the Platform Data your app pulls from Meta. Behind that question sits a contractual requirement. Meta's Platform Terms expect each of those companies to be bound by written data-processing terms with you. If your product routes anything through a large language model, your AI vendors are on that list, and the review is where you find out whether the paperwork exists.

We went through this exercise for Savra this week. The audit removed one vendor from our draft list, confirmed one DPA we already had without knowing it, and surfaced one we had to execute on the spot. All three outcomes came from the same place: reading our own code instead of our architecture diagram.

What do the Data Handling Questions ask?

Three things, in escalating order of paperwork. First, whether outside processors or service providers can access your Platform Data, and if so, who they are, what category of service they provide, and in which countries they process it. Second, which legal entity is the data controller responsible for all of it. Third, whether you have received public-authority requests for user data, and whether you have a written policy for handling them.

The questionnaire has teeth: three insufficient answers in a row lock the questions for three days. That deadline pressure is exactly why the answers should be assembled before anyone types into the dashboard, and why each answer needs to survive a reviewer comparing it against your privacy policy and your app's observed behavior.

Which vendors actually belong on the processor list?

The honest list comes from tracing where Platform Data flows in code, and both directions of error are live risks. Miss a real processor and your filing is wrong. List a vendor that never touches the data and you have volunteered a false statement a reviewer can trip over.

We nearly made the second mistake. Our draft list included a hosted memory service, because the vendor's SDK sits in our dependency tree and a client for its cloud API exists in the codebase. The audit told a different story: our memory layer runs self-hosted on our own server, and the one code path that would call the vendor's cloud reads a configuration key that has never been set. Dead code, zero executions, zero data flow. The vendor came off the list, and the correction mattered, because the difference between "self-hosted library" and "hosted processor" is the difference between a truthful filing and a wrong one.

The same tracing settled our exclusions. An auth provider and a billing provider see account emails and invoices, never Platform Data. A DNS service passes traffic without terminating it. None of them belong on the list. What remained was three names: the hosting company whose server stores everything, and two AI vendors, because our publish flows place connected-account names, obtained through Meta's API and therefore Platform Data by definition, into LLM prompts.

That last category is the one to check hardest in an AI product. Data does not need to be sensitive to be Platform Data. An account display name inside a prompt is enough to put the model vendor on the list.

Do OpenAI and Anthropic have DPAs?

Both do, through different doors.

Anthropic folds its data-processing addendum into its commercial terms. If you are an API customer, the DPA already binds both parties, and there is nothing to sign. Finding that out converted one of our list entries from "paperwork owed" to "paperwork done" without lifting a pen.

OpenAI requires a one-time execution. Its policies site hosts a self-serve DPA form asking for four things: your organization ID from the platform dashboard, whether your company is based in the EEA or Switzerland (this decides whether you contract with the US or the Irish entity), the name and title of the person signing for your company, and that signer's email. Submitting generates a contract document, delivered by email for electronic signature. Ours arrived, got signed, and came back countersigned by OpenAI within the same day. Total cost: nothing. Total effort: minutes, once we knew which account to sign under.

Which account actually signs?

That last clause carried the real work, because we hold several OpenAI accounts, accumulated the way every team accumulates them: experiments, tools, editors, side projects. The DPA binds a single organization ID. Signing under the wrong one produces a legally executed agreement that covers none of your actual data flow.

The dashboard makes matching possible without exposing secrets. Its API-keys list shows each key's trailing four characters, and comparing those against the key your production environment holds identifies the owning account with certainty. One wrinkle to expect: your production key and the key on developer laptops may not be the same key, or even the same account. Ours were not, and the production key is the one that defines the processor relationship, so the production account is the one whose organization ID went on the form.

The audit paid a dividend on the way through. Checking key ownership revealed that our development key belonged to an account with an exhausted credit balance, quietly failing every call. Ten minutes of compliance homework doubled as an infrastructure find.

FAQ

Do OpenAI and Anthropic count as data processors for Meta App Review?
If any data obtained from Meta reaches their APIs, yes. An account name inside a prompt qualifies. The questions ask for every outside company in that position, the service category, and the processing countries.

Does Anthropic require a separate DPA signature?
No. Its DPA lives inside its commercial terms and binds API customers automatically. OpenAI's requires a one-time execution through the form on its policies site.

How long does the OpenAI DPA take?
Minutes to file, an e-signature when the document arrives, and in our case a same-day countersignature. There is no fee.

Which OpenAI account should sign if we have several?
The account owning your production API key. Match the key suffix in the dashboard against your production environment before signing, and expect development and production to differ.