Skip to content
be-human.ai
← Back to the journal

Human oversight

DPA Clauses for AI Vendors: What to Check Before You Sign

Which clauses to check in an AI vendor DPA: training use, retention of prompts and memory, subprocessors, deletion, audit rights and incident notice.

Be Human typographic cover with the title: DPA Clauses for AI Vendors: What to Check Before You Sign
Conceptual illustration created with AI · Be Human

A data processing agreement for an AI vendor should answer six questions in writing: whether your data is used to train models, how long prompts, memory and logs are kept, which subprocessors can touch them, how deletion is confirmed, what audit rights you hold, and how quickly incidents are reported to you. Read these clauses before any agent sees customer or employee information, because an AI agent copies and stores data in more places than a typical software tool does. Treat the contract as part of the operating design, not as a formality signed at the end.

Key takeaways

  • Get a written statement that your inputs, outputs and agent memory are excluded from model training by default, not only after someone changes a setting.
  • Map every place data can live: prompts, memory, tool call logs, vector stores, transcripts, evaluation records and backups.
  • Require a named list of subprocessors, with advance notice of changes and a right to object or leave if a change breaks your requirements.
  • Define deletion in operational terms, including derived memory and backups, and ask how the vendor verifies that deletion happened.
  • Agree on incident notice timing and cooperation before the pilot starts, not during the first incident.

What a DPA must cover when the vendor runs an AI agent

A standard software DPA was written for systems that store records and answer queries. An AI agent does more. It reads an inbox, decides what to pull from a customer system, writes a summary, calls a tool, and keeps notes for the next conversation. Each step creates a copy of personal or confidential data, and some of those copies sit outside the screens your staff normally check. The clauses below therefore need to describe the agent's data flows, not only the vendor's main application database.

The NIST AI Risk Management Framework frames AI risk as something to govern, map, measure and manage across the life of a system. That lifecycle view is a useful lens for contract review, because the risks change when a vendor updates a model, adds a tool or routes requests to a new provider. In Brazil, LGPD obligations for personal data apply to what the agent does with that data, whatever the vendor calls the feature. A review that starts from the agent's actual work, as described in the data processing agreement checklist for AI services, is more reliable than one that starts from the vendor's standard terms.

Training use: the clause that decides who benefits from your data

Start here, because it is the clause most often written in the vendor's favor. Ask whether prompts, uploaded files, outputs, feedback and memory are used to train or fine-tune any model, including models that serve other customers. Some vendors default to using service data for product improvement unless the customer opts out, and the opt-out may apply only to certain plans, interfaces or regions. Read the default, not the sales summary.

Ask these questions and get the answers in the contract or a referenced policy:

  • Is training excluded by default, or does it depend on an administrator changing a setting?
  • Does the exclusion also cover human review of conversations, and if review happens, who performs it and under what conditions?
  • Do feedback buttons, ratings and error reports fall outside the exclusion?
  • Does the same rule apply to every model the vendor routes your requests to, including third-party models?
  • Can the vendor show the current setting in writing, with a version reference, so you can file the evidence?

If the answer is that training is allowed unless you opt out, decide before signing whether your staff and your customers would accept that. Record the decision with an owner.

Retention: how long prompts, memory and logs survive

Retention is where most surprises appear. A sentence such as "we retain data as needed to provide the service" is not a retention period. The contract should separate categories (raw prompts, model outputs, agent memory, tool call logs, evaluation transcripts and backups) and give each one its own period and deletion route. Ask whether any period can be shortened through configuration, and whether that configuration is available on the plan you are buying.

Keep audit detail on purpose. You need enough logs to reconstruct what an agent read, decided and changed, but not so much that every personal detail is copied indefinitely. The trade-off is set out in how long to keep AI agent logs and transcripts, and the related question of removing stored memory is covered in deleting AI agent memory on request. Put both in the contract as concrete categories, because a retention rule that nobody can apply is not a control.

Subprocessors and where the data travels

An AI vendor often relies on hosting providers, model providers, speech or email services, and monitoring tools. Each of these is a subprocessor that receives some of your data. Ask for the current list, the function of each entry, and the countries where processing takes place. The clause should also describe how new subprocessors are added, how much notice you receive, and what happens if you object.

  • A current subprocessor list with the function and processing location for each entry.
  • Advance written notice before any addition or replacement, with a defined window to object.
  • Flow-down terms: each subprocessor is bound to obligations no weaker than the ones in your DPA.
  • The vendor remains responsible to you for failures of its subprocessors.
  • A termination right without penalty if an objection cannot be resolved.

The practical questions are explored further in AI vendor subprocessors and who else handles company data. Ask also whether the model provider can see your prompts, and whether that provider has its own retention terms that differ from yours.

Deletion, return and audit rights

At the end of the relationship you need three things: your data returned in a usable format, deletion from active systems within a stated window, and written confirmation covering backups and derived material. Deleting a database record is not the same as deleting an embedding, a summary the agent wrote into its notes, or a cached response. Name each of these in the clause, and ask how the vendor proves completion.

Audit rights matter as much as deletion. Ask for a current independent security report if the vendor has one, and for a process to inspect how your data was handled without requesting the vendor's source code. Confirm who inside the vendor can read your logs and memory, and whether that access is limited and recorded. The OECD AI Principles emphasize transparency and accountability across the AI supply chain, and a DPA that makes those qualities checkable is more useful than one that promises them in general terms.

How to run the review before you sign

The review is faster when it follows a fixed sequence. Work from the agent's job outward, so that each clause is tested against a real data flow rather than an abstract category.

  • Write down the one role the agent will perform and the data categories it will read, write and store.
  • Ask the vendor for a data flow diagram and have them correct it where it is wrong.
  • Mark each category as personal, confidential or public, using your existing classification.
  • Review training, retention, subprocessor and deletion clauses against that map, one category at a time.
  • Record every exception as a decision with an owner, a date and a condition that would trigger a new review.
  • Ask legal and security to approve the exceptions list, not only the signed contract.

Illustrative scenario: a regional services company wants an agent to answer billing questions from customers. The data map shows that the agent reads invoices, stores a short summary in memory, and sends call transcripts to a speech vendor. The review finds that training on transcripts is allowed by default, that the subprocessor list omits the speech vendor, and that memory has no deletion route. The team asks for the training opt-out in writing, adds the speech vendor to the list, and limits the pilot's memory to order references. The agent starts only after those three changes appear in the contract.

How Be Human handles this: in a managed engagement, Be Human begins with the data map for the one role the agent will perform, then configures memory, tool access and log retention to match what the contract actually allows. Each vendor setting is recorded as evidence, and periodic permission reviews check that the configuration has not drifted after vendor updates or new subprocessors. The work is remote and available in English and Portuguese.

If you are not yet sure which task would be the safest first step for an agent, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.

Frequently asked questions

Does a strong DPA make an AI vendor safe to use? A DPA allocates responsibilities and sets the rules for handling data, but it does not change how the agent behaves in daily work. Pair the contract with configuration checks, permission limits and a review schedule, so the written terms match what actually happens in the system.

Can we accept a default that lets the vendor use our data to improve its products? You can, but only as a deliberate decision with an owner and a record. Confirm what "improve" covers, whether human review is included, and whether the opt-out is available on the plan you are buying. If the answer leaves customer or employee data exposed, ask for the exclusion as a condition of signing.

What should we do if the vendor refuses to change its standard DPA? Ask which specific clauses it will not change, and whether it offers a configuration, an addendum or a different plan that meets your requirement. If none of these work, treat the gap as a risk to record in your register and decide whether the agent's task should run with less data or not run at all. A refusal on training, deletion or subprocessor notice is a signal to reconsider the use case.

Should a Brazilian company check LGPD obligations in addition to the vendor's DPA? Yes. The vendor's DPA governs the relationship between you and the vendor, but your obligations toward the people whose data is processed come from the law and from your own privacy notices. Review the agent's purposes, legal bases and retention against those obligations, and keep the evidence with the contract file.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →