DPA Check for Superintelligence Vendors: Verify Before Signing
A practical DPA check for AI vendors: what the agreement must cover, including subprocessors, retention and deletion, before you connect company records.

A DPA check for an AI vendor means reading the data processing agreement against the product you plan to use and confirming that it states, in writing, what the vendor may do with your company data, which subprocessors may touch it, how long it is kept, and how deletion works. Do this before any customer record, staff file or internal document reaches an agent. A signed agreement that only promises general compliance does not answer those questions, and those are the answers your team will need when something goes wrong.
Key takeaways
- A DPA should name the specific processing the vendor performs for your account, and it should exclude training on your data unless you have agreed to it in writing.
- Subprocessors need a current list, advance notice before changes, and a practical right to object or exit.
- Retention and deletion need defined periods, a clear deletion scope that covers memory, indexes, logs and backups, and a written confirmation.
- Check the contract against the product tier, connectors and regions you will actually use, not the vendor's general terms page.
- Treat the DPA as a gate: no records connect until a named person has checked each item and recorded the answer.
What must a DPA with an AI vendor define?
Start with the elements every data processing agreement should contain: the parties and their roles, the subject matter, the purpose of the processing, the categories of personal data, the categories of people whose data appears, and the duration. In Brazil, the LGPD uses the terms controller and operator, so the contract should make clear which party holds which role at each stage, from prompt handling to log storage. The goal is not legal polish. The goal is to make sure the vendor's obligations match the way your agent will really use data. If you operate in Brazil, the guide on LGPD and AI agents explains which of these points local teams usually need to document.
A useful test is to read the purpose clause and ask whether it describes your planned use. A phrase such as providing and improving our services is too broad to govern an agent that reads your contracts folder. Ask the vendor to describe the processing for your account and to attach that description as an annex. The NIST AI Risk Management Framework is a helpful reference for turning this review into a routine rather than a one-time signature.
Which subprocessors will touch your data?
A subprocessor is any third party the vendor uses that handles your data on its behalf. For an AI agent, that can include model hosting, cloud storage for logs, a vector database for memory, speech services, email delivery or monitoring tools. Each one widens the group of parties who can see your records, so each one must be visible in the agreement.
When you review the subprocessor section, ask for the following:
- A current list with the legal name of each subprocessor, its function, and the country or region where it processes data.
- A commitment to give advance notice before adding or replacing a subprocessor, with enough time to object.
- Confirmation that each subprocessor is bound by obligations no weaker than the ones in your DPA.
- A clear statement of what happens to your data and your exports if you object and decide to leave.
Pay attention to connectors. A model provider may be clear about its own role while the browser tool, email relay or ticketing integration you add later brings in another party. Repeat the subprocessor check each time you add a tool. The article on AI vendor subprocessors walks through how to map this chain for company data.
How long is data kept, and what does deletion cover?
Retention is where many agreements stay vague. An agent creates several stores of data: prompts and outputs, application logs, any memory or summaries kept between sessions, embeddings or indexes built from your documents, cached intermediate results, and backups. The DPA should say what is retained by default, for how long, and whether you can shorten those periods through configuration rather than by sending a request each time.
Ask for written answers to these points:
- The default retention period for prompts, outputs and logs, and whether you can set it shorter.
- Which stores are included when you request deletion: primary storage, indexes, memory, logs and backups.
- How deletion is carried out, how long backups take to roll off, and what written confirmation you receive.
- Whether deletion can be verified for your account without exposing the data of other customers.
Memory deserves its own clause. If an agent remembers a staff member's preferences or a customer's history, deletion must reach that memory, not only the visible chat transcript. The guide on deleting agent memory on request sets out a find, remove and verify routine that you can ask a vendor to support.
Does the vendor train on your data or reuse it?
Read the training clause slowly. Some agreements state that customer content is not used to train models by default. Others leave room for quality improvement, safety review, evaluation or analytics. These are different promises. Ask for the exact sentence, then decide whether your records can be used for any secondary purpose, including human review of sampled conversations, evaluation datasets or aggregated usage reports.
Get the answer for each product tier and interface you plan to use. An API account, an enterprise workspace and a consumer assistant can carry different terms. The contract you sign should name the product, not only the company, so that a later change of tier does not quietly change the rules.
How are access, security and breach notices handled?
A DPA should describe security measures in terms you can check, not as a general promise to follow industry standards. Ask who at the vendor can read customer content, whether support access is restricted and logged, how data is encrypted in storage and in transit, and how the credentials for your connected tools are protected. The OWASP Authorization Cheat Sheet is a practical reference for the principle that each identity, human or agent, should hold only the access its task requires.
Breach notification should have a defined notice period and a named contact on the vendor side. The agreement should also say who cooperates in an investigation, who bears which costs, and what evidence you will receive afterwards. Confirm that the vendor can provide the logs you need to reconstruct an incident without handing you more personal data than the investigation requires.
How does a managed approach keep the DPA check current?
Illustrative scenario: a regional logistics company wants an agent to answer dispatch questions from a folder of supplier contracts. The vendor's DPA says logs are kept for a period the vendor sets and that the subprocessor list is available on request. The operations lead does not connect the folder yet. She asks for the subprocessor list, a shorter log retention setting and a deletion path that covers the document index. The folder is connected only after those answers are written into the agreement or an addendum.
Be Human designs, implements and operates personalized AI infrastructure for companies, and the DPA check is part of how a first engagement begins. The first engagement defines one useful role, the memory and tools it needs and its approval boundaries, and then maps each data store and subprocessor against the vendor terms before any company record is connected. Managed operations can keep that map current through permission review, integration maintenance and incident response, so that the contract and the running system do not drift apart over time.
Before you sign and before you connect
Use this sequence before any record moves:
- Name the product and tier the DPA covers, and confirm it matches the one you will use.
- Read the purpose, training and secondary use clauses against the tasks you have described.
- Obtain the current subprocessor list and the notice terms for changes.
- Write down the retention periods, the deletion scope and the confirmation you will receive.
- Record who on your side approves the connection and which evidence they reviewed.
Keep the signed set with your risk register, so that a later audit or incident review starts from the contract rather than from memory. Once you know which records a first task would need, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Is a DPA enough to make an AI vendor safe to use? A DPA sets obligations and makes responsibilities enforceable, but it does not configure your system. Pair the contract with settings that shorten retention, scoped access for each agent and a scheduled review of permissions. The contract tells you what the vendor must do, and your configuration decides what the agent can reach.
Do I need a separate DPA for each AI tool? Usually each vendor and each product tier needs its own agreement or annex, because terms differ by service. If a tool sends data to a model provider, that provider's terms also apply to the data passed through it. List each one as a separate line in your review so that none is missed.
What if a vendor refuses to list its subprocessors? Treat the refusal as a finding rather than a formality. Offer to receive the list under confidentiality if that is the concern, and decide whether the answer is sufficient for the records you plan to share. If it is not, keep those records out of the agent and choose a narrower first task.
Does the DPA check apply to small companies too? Yes, although the depth can match the scale. A small team with one agent and a limited document set can often complete the check with the vendor's standard terms, a short subprocessor list and a written retention setting. What matters is that these answers exist before any records move.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


