Superintelligence Vendor DPA Checklist: What to Verify Before Signing
Clauses to check in an AI vendor DPA: training, subprocessors, retention, deletion, breach notice and audit rights for agent memory and records.

Before signing with any AI vendor that stores your agent's memory or processes your customer or staff records, verify that the data processing agreement answers six questions: what the data is used for, whether it trains models, who else receives it, how long copies live, how deletion is proven, and how incidents and audits are handled. A DPA that is silent on any of these leaves your company guessing at its obligations later. This checklist sets out what to verify, in what order, and what to do when a vendor pushes back.
Key takeaways
- Put the processing purpose in writing and limit it to the services you are buying.
- Get a contractual statement on whether your memory, prompts, documents and logs can train or improve any model, with the default visible.
- Require a current subprocessor list, advance notice of changes, and a right to object with a real exit path.
- Match retention and deletion terms to every copy of the data, including indexes, logs, caches and backups.
- Ask for evidence such as assurance reports, deletion confirmations and incident timelines, not verbal assurances.
What a DPA must cover when a vendor holds your memory
An AI agent is not a single feature. It keeps conversation history, summaries it writes for itself, documents it retrieves to answer questions, tool outputs and often an index of past records. Each of these is personal data when it describes a person, and each can outlive the session that created it. A DPA written for a generic software tool often describes storage of customer data in one place. That wording misses memory that persists, gets reused across tasks and gets copied into logs.
Ask the vendor to describe the data flow in an annex to the contract rather than in a sales deck. The annex should name the categories of personal data, the people concerned (customers, staff, suppliers), the purposes of processing and its duration. It should also state who acts as controller and who acts as processor or operator at each step. Under Brazil's LGPD, and under similar frameworks elsewhere, those roles determine who owes what to the individuals whose data is involved. The National Data Protection Authority publishes guidance on these roles, and your counsel should compare the vendor's definitions against it.
Does the vendor train on your data, and can you turn it off?
This is the clause that most often differs from what the sales conversation implied. Verify three things separately. First, whether customer content such as prompts, memory entries, uploaded documents and transcripts is used to train or fine-tune any model by default. Second, whether the vendor uses content for service improvement, safety review or human evaluation, and under what conditions. Third, whether usage metrics such as request counts, error rates and latency are treated differently from the content itself. Operational telemetry is normal. Reading your records for improvement is a separate decision that needs its own clause.
The contract should answer each question in plain terms. A setting in an admin console is weaker than a named contractual exclusion, because settings change with product updates and the DPA is the document you can enforce.
Illustrative scenario: a regional logistics company deploys a support agent that summarizes delivery disputes and keeps a memory of each carrier's history. The vendor's standard terms allow improvement use where the contract permits it. The operations lead asks for a named exclusion covering customer content, memory stores and logs, plus written confirmation that the exclusion applies to data already uploaded during a trial. The vendor accepts the exclusion but offers no retroactive confirmation, so the company deletes the trial data before go-live and records that step in its contract file.
Who else touches the data? Subprocessors and transfers
Your AI vendor probably relies on other companies: cloud hosts, model providers, speech services, email delivery and monitoring tools. Each one that receives personal data is a subprocessor. Ask for the current list, with the purpose each one serves, the location where data is processed, and whether each of them uses further providers. The list should be a document you can file and compare later, not a web page that changes without notice.
The DPA should promise advance notice before a subprocessor is added or replaced, a reasonable period to object, and a stated consequence if you object, such as the right to terminate with a fair return of any prepaid amounts. Check transfer terms when data leaves your country. Brazilian companies should confirm how international transfers are covered before any memory store is placed outside Brazil. For a deeper look at this layer, read AI Vendor Subprocessors: Who Else Handles Your Company Data?.
How long is memory kept, and how is deletion proven?
Retention is where agent systems differ most from ordinary software. One user request can create a copy in the primary memory store, an entry in a vector index, a conversation log, a support ticket, an analytics event and a backup. A DPA that says data is retained for the term of the agreement and then deleted tells you nothing about most of those copies.
Ask the vendor to produce a table of data stores for your account, with a retention period and a deletion trigger for each. Then ask what happens when you request deletion of one person's records: which stores are purged, how quickly, whether backups are excluded until they rotate, and what confirmation you receive. Ask also about any caching layer that holds prompt content and how long its entries last.
A practical plan for this section:
- Map every store that can hold the data: primary memory, vector index, logs, backups, caches and support exports.
- Write the retention period and deletion trigger beside each store in the DPA annex.
- Test one deletion request from start to finish before go-live, and keep the confirmation in your file.
- Get the backup rotation timing in writing, so you know when a deleted record truly leaves the vendor's systems.
For the operational side of this step, see Deleting AI Agent Memory on Request: Find, Remove, Verify.
Breach notice, audit rights and the evidence to request
Incident clauses decide how quickly you learn about a problem. Check whether the vendor must notify you of a security incident affecting your data within a fixed period written in the contract, what the notice must include, and who bears the cost of notifying affected people when the vendor's failure caused the incident. The notice duty should cover incidents at subprocessors too, since a breach upstream is still a breach of your data.
Audit rights should be practical. Few vendors accept on-site audits by customers, but most can provide independent assurance reports. Ask for the most recent report, the systems it covers, the period it describes and any exceptions noted. Be cautious about a report whose scope leaves out the memory service or the agent runtime. Also ask for a written description of how staff access to customer content is limited, logged and reviewed. For a fuller list, see SOC 2 Evidence for AI Agent Vendors: What to Request Before Sharing Memory.
How Be Human handles this: in managed engagements, Be Human keeps a written inventory of each agent's memory stores, tools and vendor data flows, so the DPA review starts from what is actually stored rather than from the vendor's template. Permissions and vendor access are reviewed on a recurring schedule, and any change to a subprocessor, memory setting or retention rule is checked against the terms the company agreed to. Be Human does not replace legal counsel. The company keeps ownership of the contract and decides the final wording with its own advisers.
A signing checklist, and what to do when a vendor refuses
Use this sequence before anyone signs:
- Write down every system that will hold personal data, including memory, logs, indexes and backups.
- Ask for the data flow annex and compare it with what the product actually stores.
- Confirm training and improvement terms as named exclusions, not as default settings.
- Obtain the current subprocessor list and the notice terms for changes.
- Match retention periods to each store and test one deletion end to end.
- Confirm the incident notice duty, cooperation obligations and the assurance evidence you will receive.
- Check exit terms: the format for returning data, the deletion certificate and the timing.
- Have counsel review the package against your local law and the vendor's standard terms.
When a vendor refuses a clause, ask which part of its standard terms causes the refusal and whether a narrower scope is possible. Often the practical answer is to reduce what the agent stores. You can keep memory on your side, send only summaries to the vendor, or remove fields the agent does not need. You can also restrict the agent's permissions so it cannot write certain record types to the vendor at all. Record the decision and the reason in your file, because a later review will ask why a gap was accepted. If you are unsure which task should go to an agent first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Do we need a DPA if staff only use the vendor's chat interface? If staff enter customer or employee personal data into the tool, the vendor processes that data on your behalf, and the contract should say so. Free or consumer terms often differ from business terms on training and retention, so check which tier applies. Use the business terms for company data, confirm the training default in writing, and keep a list of tools approved for personal data.
Is a signed DPA enough, or do we also need a security review? A DPA sets obligations, but it does not show how the controls work in practice. Pair it with a security questionnaire and the vendor's assurance reports. The contract tells you what the vendor promised, and the evidence tells you whether the controls exist and cover the systems that hold your data.
Who is accountable if the vendor misuses our data? Your company usually remains accountable to the individuals and regulators concerned, even when a vendor causes the failure. That is why the contract should give you clear notice duties, cooperation obligations and a path to recover losses. Liability caps and indemnities are negotiated terms, and counsel should review them before signature.
What if the vendor will not change its standard DPA? Many vendors offer a standard addendum and resist edits to it. Ask for the specific clauses you need as a written addendum, and compare the standard text line by line against this checklist. If important gaps remain, reduce the data the agent handles, or choose a different vendor for that workload.
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.


