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

Human oversight

AI DPA Compliance Check: Vendor Terms to Verify Before Agents Touch Data

The DPA items to verify with AI vendors before agents touch customer or employee data, with a note on when counsel must review final terms.

Be Human typographic cover with the title: AI DPA Compliance Check: Vendor Terms to Verify Before Agents Touch Data
Conceptual illustration created with AI · Be Human

An AI DPA compliance check is a structured review of a data processing agreement and the related vendor terms before an AI agent receives customer or employee personal data. It confirms what data the vendor may process, for which purposes, where it is stored, who else handles it, how long it is kept, whether it trains models, and how deletion and breach notice work. Use the result as a working checklist for your business team, and have counsel review the final terms before anything is signed.

Key takeaways

  • Verify purpose and scope before the agent gets access, not after the pilot has already run on live records.
  • Get subprocessors, data locations and model training use in writing, because agent tools usually add more parties than the chat product did.
  • Check that retention and deletion reach agent memory, retrieval stores, tool outputs and logs, not only the main account data.
  • Match breach notice, audit rights and access controls to the actions the agent can actually take.
  • Treat this article as an operational checklist. Counsel should review final terms, liability and jurisdiction clauses.

Why vendor terms decide what an agent can do

Many companies look at an AI tool's features first and read the contract later. For a tool that only drafts text, that order is tolerable. For an agent that reads a support inbox, looks up an order, updates a customer record or summarizes employee cases, the contract defines the edges of what is allowed. A data processing agreement sets the vendor's duties when it processes data on your behalf. The rest of the paperwork, including terms of service, security annexes, acceptable use rules and order forms, can change those duties in practice, so read the documents together.

Agents also widen the surface. A single agent may call a model provider, a retrieval index, a messaging channel, a CRM and a logging service, and each link can hold personal data. The useful question is not only whether a vendor signs a DPA. It is whether the DPA covers every component the agent uses, in the way you plan to use it.

Scope of processing: what data, for which purposes, through which actions

Start with a written description of the data categories the agent will touch: names and contact details, order history, support transcripts, scheduling data for staff, and any health or financial details. Then state the purposes in the plain terms the DPA uses. A vendor may be authorized to process data "to provide the service," which is broad. Your internal purpose should be narrower, for example "draft replies to delivery questions and open a case for a person to review." The narrower the stated purpose, the easier it is to hold both the vendor and your own team to it.

Check that the DPA's instructions clause lets you direct processing through your configuration, including which actions the agent may take. If the agent can send messages or change records, that is processing too, and the vendor should know about it.

Illustrative scenario: a retail company wants an agent to answer delivery questions. The first configuration pulled full order histories, including partial payment details, into every prompt. During the review, the team changed the agent to request only the order number and delivery status. The vendor then processed less personal data, and the logs held less of it as well. A DPA cannot fix over-sharing that is built into the workflow, so the scope check and the data design have to happen together. Our article on data minimization for AI agents covers this in more detail.

Subprocessors, model providers and data location

Ask for a current list of subprocessors. For an agent, this usually includes the model provider, the hosting cloud, any vector or document store, the email or messaging gateway, and any monitoring tool that receives prompts or outputs. Verify how the vendor gives notice when it adds or replaces a subprocessor, and whether you have a right to object and what happens if you do. Ask where data is stored and processed, and whether backups or support staff access it from other countries. If the model provider's own API terms apply to your traffic, read what that provider retains and for how long.

For companies handling Brazilian personal data, check the processor roles and any international transfer conditions against the LGPD. The [ANPD guidance portal](https://www.gov.br/anpd/pt-br) is the official starting point for the Brazilian rules. Our article on subprocessors for AI vendors explains how to trace the chain, and the piece on LGPD and AI agents lists the questions Brazilian operators usually need to answer.

Training, retention and deletion for memory and logs

Make two questions explicit. First, will your data, prompts, outputs or logs be used to train or improve models, including under default settings? Second, how long does the vendor keep each category: prompts, tool outputs, conversation memory, embeddings, audit logs and backups, and how is deletion confirmed?

Agents hold more than a chat application does. Memory stores, retrieval indexes and action logs can each keep personal data after the task is finished. The DPA should name each store and its retention rule, not just refer to "customer data." Ask for deletion to reach backups on a stated schedule, and ask for written confirmation when a deletion request is completed. Logs are needed for audit, so the aim is to keep enough detail to reconstruct decisions without storing more personal data than necessary.

Security, breach notice and audit rights

The breach clause should state when the vendor must notify you, measured from discovery rather than from the end of its internal investigation. It should say what information you will receive and who at the vendor is the contact. Check whether vendor staff access is restricted and logged, and whether you can request evidence after an incident, such as an independent security report, a completed questionnaire or access logs covering your tenant.

The DPA describes the vendor's security measures, but your agent's permissions are your responsibility. Each tool the agent uses should carry only the scopes the task needs, and high-impact actions should require a human decision. The [OWASP Authorization Cheat Sheet](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a practical reference for thinking about this layer. Agents that read email or documents also face prompt injection, where hostile text tries to make the agent reveal or send data. A vendor's security annex does not remove that risk, so review it alongside your own controls.

Exit, liability and who signs off

Read the termination clause as carefully as the processing clause. Confirm how you get your data back, in what format, within what period, and how the vendor certifies deletion after exit. Check what happens if the vendor changes its terms or retires a model you depend on, and whether you can leave without penalty.

Liability caps, indemnities and governing law are commercial decisions. A DPA allocates roles under the data protection rules, but it does not decide who is accountable when an agent makes a mistake inside your business. That question needs an internal answer and, where relevant, an insurance review. Counsel should review these clauses before signature.

How to run the check before go-live

The following sequence keeps the review short enough to finish and specific enough to act on:

  • Write the agent's role, the data categories, the purposes and the actions on one page.
  • Collect the DPA, security annex, subprocessor list and product terms for every component the agent uses.
  • Mark each required item as present, missing or unclear, with the clause reference beside it.
  • Send written questions for every unclear item, and file the answers with the review record.
  • Confirm training use, retention and deletion for memory, retrieval stores, logs and prompts.
  • Pass the marked list and the open items to counsel and to the business owner who is accountable for the workflow.
  • Repeat the check whenever a subprocessor, model, purpose or data source changes.

In Be Human's managed model, the agent's role, memory and tool scope are defined first, and each data flow is mapped against the vendor documents before connection. Permissions remain under your control, changes to tools or data sources are recorded, and a change in subprocessors or terms triggers a new review as part of ongoing permission checks. The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) offers a useful structure for recording these risks. Be Human does not replace your legal counsel, and the final contract terms remain a decision for your company.

If you want to find the first task worth handing to an agent, the free Be Human AI diagnostic is a practical place to start.

Frequently asked questions

Does a signed DPA make an AI vendor compliant? No. A DPA sets the vendor's duties as a processor, but compliance depends on how you configure the agent, which data you send, how long you keep outputs and whether your own internal controls work. Use the DPA as one input to your review, not as the whole review.

Do we need a DPA if the agent only works with public information? Often the DPA still matters, because agents tend to combine public data with customer messages, account records or employee details during a task. Check what the agent actually receives at runtime, not only what the source documents contain. If personal data can enter the workflow at any point, the processing terms should cover it.

Can we accept the vendor's standard DPA without negotiating? You can accept a standard DPA if its scope, subprocessor list, retention rules, training use and breach notice fit your intended use. If any of those items conflict with your obligations or your customers' expectations, raise it in writing and involve counsel. A standard form is a starting point, and the gaps are usually found by comparing it with your specific workflow.

Who should own this check inside a small company? Assign one business owner who is accountable for the workflow and its outcomes, and involve whoever handles privacy or legal questions, even if that person is an outside advisor. The owner should keep the marked checklist, the open questions and the approval record, and should repeat the check when the agent or its vendors change.

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 →