Data Processing Agreement Clauses for Superintelligence Vendors
Which data processing agreement clauses to check before an AI vendor stores your company memory: scope, subprocessors, deletion, logs, audit and exit.

A data processing agreement with an AI vendor should state what the vendor may do with your company data, where that data is stored and processed, which subcontractors handle it, how long memory and logs are kept, how deletion is proven, and how you audit and leave the arrangement. When the vendor will hold company memory, these clauses matter more than the feature list, because memory persists across conversations and outlives any single task. The sections below name the clauses to check and the order in which to check them.
Key takeaways
- Start with scope: name the purposes the vendor may process data for, and forbid everything else, including model training unless you approve it in writing.
- Map every place memory, summaries, embeddings, prompts and logs live, then require the agreement to cover each one.
- Require a current subprocessor list, advance notice of changes and a real right to object.
- Write deletion terms that reach backups, logs and derived stores, not only the primary database.
- Treat audit evidence and exit assistance as contract clauses, not courtesy promises.
What a data processing agreement must cover when an agent holds memory
A standard data processing agreement was written for software that stores records. An agent does more than store: it reads records, writes summaries, keeps notes across sessions and calls tools that change other systems. The agreement therefore has to describe the memory itself, not only the database underneath it.
Ask the vendor to list the data categories it expects to hold, such as customer contact details, staff messages, quotes, support history and internal documents. Each category should appear in the agreement or in an annex, with the purpose it serves. If a category is missing from the list, the vendor may still receive it in practice, which is the situation the contract exists to prevent.
For companies that process personal data in Brazil, the same discipline applies under the national data protection law. The practical questions are covered in LGPD and AI Agents: What Brazilian Companies Should Check.
Scope and training: what the vendor may do with your data
The first clause to read is the purpose limitation. The vendor should process company data only to deliver the contracted service and only on your documented instructions. General wording such as "to improve our services" leaves room for uses you never intended, so the purposes should be named.
Model training deserves its own clause. State whether prompts, outputs, stored memory or logs may be used to train or improve any model, including models owned by the vendor's own providers. The safe default is no, with any exception approved in writing for a named purpose and a named data set.
The agreement should also address derived data. Usage statistics, anonymized samples and aggregated quality metrics are often kept after a contract ends. Ask what is derived, whether it can contain personal details, who can access it and how long it stays.
Finally, check human access. Identify which vendor staff can read stored memory or transcripts, under what conditions, and whether each access is logged. Support access should be exceptional, time-limited and traceable to a ticket.
Subprocessors, model providers and where data lives
An agent platform usually depends on a chain of providers: hosting, a model interface, vector storage, a messaging channel, error tracking and analytics. Each of them is a subprocessor that may touch your data. A DPA that names only the vendor describes a smaller system than the one you are buying.
The clauses to require are these:
- A named list of subprocessors, each with the function it performs and the data it receives.
- Advance notice before a new subprocessor is added, with enough time to review it.
- A right to object, and a defined outcome if you object, such as an alternative configuration or termination without penalty.
- Flow-down terms, so that each subprocessor is bound by obligations no weaker than the ones in your agreement.
- Data location for storage, backups and processing, and the conditions under which data may move across borders.
Illustrative scenario: a regional services firm asks an agent to draft replies to client emails. The vendor hosts the agent in one region, the model call goes to a separate provider in another, and an error-tracking tool sits in a third. The first draft of the DPA named only the vendor. Reviewing the subprocessor list exposed the other two parties, and the team added them to the annex, with the transfer basis for each, before the agent received any client mail.
For a deeper look at this layer, see AI Vendor Subprocessors: Who Else Handles Your Company Data?. Where data is physically kept is covered in Data Residency for AI Agents: Where Memory, Logs and Model Calls Live.
Memory, embeddings, logs and deletion
Deletion is the clause most often weakened in practice. Agent memory is not one table. It can include raw transcripts, extracted facts, summaries, vector embeddings, cached context, tool call records and audit logs. Each of these is a store, and each needs its own retention rule.
Embeddings deserve particular attention. An embedding is not a readable copy of a message, but it is derived from one and can still carry personal details. A deletion request that removes the source message while leaving its embedding in place is incomplete. The agreement should say that derived stores are deleted with their sources.
Backups and replicas must be covered too. State how deleted data leaves backups, and within what window, so that the contract does not promise a deletion the infrastructure cannot deliver. A written description of the backup cycle is a reasonable thing to request.
Logs need two rules at once: enough detail to investigate a mistake, and no more personal data than that investigation needs. Observability practice, as described in the [OpenTelemetry observability primer](https://opentelemetry.io/docs/concepts/observability-primer/), treats logs, traces and metrics as separate signals that can be collected at different levels of detail. Use that separation in the contract: specify which signals are retained, for how long, and what is redacted before storage.
The checklist for this section is short:
- A retention period for each store, set by your purpose rather than the vendor's default.
- Deletion on instruction, with written confirmation that you can keep.
- Embeddings and summaries deleted together with their source records.
- Backups and replicas included, with a stated expiry.
- Logs limited to the detail an audit needs, with personal data redacted where possible.
Security, access and incident terms
Ask for security measures written into the agreement or an attached schedule, not a link to a general page that can change without notice. The schedule should describe encryption in transit and at rest, how credentials for tools are stored and rotated, and how access to agent permissions is granted and removed.
Agent permissions deserve a specific clause. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a useful reference for checking that access decisions are made for each request, based on the role and the resource, rather than assumed from a session. Your agreement should state that the vendor supports least-privilege tool access and that permissions can be reviewed and revoked by you.
Prompt injection belongs in the security schedule as well. Content from emails, web pages or documents can instruct an agent to reveal memory or call a tool it should not use. The [OWASP Top 10 for LLM Applications](https://genai.owasp.org/llm-top-10/) lists this class of risk, and the vendor should describe the controls it applies rather than only asserting that it is handled.
The incident clause needs four elements: notification without undue delay after the vendor becomes aware of a breach, a defined content for that notice, cooperation in investigation, and a statement of who bears the cost of notifying affected people. Agree on a named contact on each side before go-live, so the first hours of an incident are not spent finding the right person.
Audit evidence, exit and return of data
Audit rights are only useful if you can act on what you receive. At a minimum, request a current security report or independent assessment, the ability to send a written questionnaire, and, for higher-risk deployments, a right to inspect or commission an inspection. Ask which systems the report covers. A report that excludes the agent platform does not evidence the platform.
Exit terms decide what happens when you stop. The agreement should require return of your data in a usable, documented format, assistance during migration to another system, a written confirmation of deletion after return, and a stated period for holding data after termination. The period should be as short as your purpose allows.
The [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) gives a shared vocabulary for mapping, measuring and managing AI risk. It is a practical way to organize the evidence you ask for, because each clause can be tied to a risk the vendor is expected to manage and a record you can inspect later.
For compliance teams, the evidence list is worth writing down before negotiation begins. The companion piece Superintelligence Readiness for Compliance Teams: Evidence to Keep Now covers the records that help during an audit.
Turning the clauses into a signing plan
Work through the agreement in this order, so that the commercial discussion does not settle the hard questions first:
- Step 1: write the purposes and data categories the agent needs, in plain language.
- Step 2: list every store and every subprocessor the vendor will use for that purpose.
- Step 3: set a retention rule and a deletion method for each store, including backups.
- Step 4: decide on training use, human access and the location of each store.
- Step 5: agree the security schedule, the incident contact and the notification content.
- Step 6: fix the audit evidence you will receive and the exit terms.
- Step 7: read the final text against the list from Steps 1 to 6, and reject any clause that leaves a store or party unaddressed.
Be Human designs, implements and operates personalized AI infrastructure, so the memory, channels, tools and permissions an agent uses are mapped during the first engagement and kept current in managed operations. That map is what makes a data processing agreement checkable: each store, subprocessor and tool is listed with its purpose, retention and owner, and permission reviews and incident response work from the same record. Contract terms and operating records are then kept together rather than in separate folders.
If you are not yet sure which part of your operations should be the first to hold agent memory, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Does a data processing agreement replace a security review? No. The agreement sets obligations, while a security review tests whether the vendor and your own configuration meet them. Read the DPA first, then review the systems it describes, so that the review covers what you have actually contracted for.
Can we sign with a vendor that uses our data to train its models? You can, but it should be a deliberate decision recorded in writing, with a named purpose and data set. The default position for company memory and logs should be no training use. If the vendor will not separate training use from service use, treat that as a material finding in your review.
What if the vendor refuses to name its subprocessors? A refusal to name subprocessors means you cannot assess where your data goes, and that is a reason to pause. Ask for the list in writing, the function of each party and the location of processing. If the vendor offers only a general statement, the agreement is not yet specific enough to sign.
Should we review the DPA again after a model or tool change? Yes. A new model provider, a new messaging channel or a new analytics tool changes the subprocessor list and sometimes the data location. Set a standing trigger in your operating process so that every change to the agent's tools or providers is checked against the agreement before it goes live.
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.


