---
title: What data processing terms should you require from an AI governance vendor?
url: https://opencontroller.lyzr.com/blog/data-processing-terms-for-ai-governance-vendors
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What data processing terms should you require from an AI governance vendor?

**A governance tool that sees prompts, model outputs and traces is processing some of your most sensitive data, so its data processing agreement needs more than the standard template. Require clauses on exactly which data categories it collects, how long prompts and traces are kept, a ban on using your data to train models, named subprocessors including model providers, data residency, breach notice, and verified deletion at the end.**

Agent governance tools sit close to the work. To score quality, detect misuse or trace an incident, they may store full prompts, tool arguments and outputs. That content can include personal data, customer records and source code. This post lists the terms to ask for. It is not legal advice; involve your privacy and legal teams. For the wider security review, see [security questions to ask a governance vendor](/blog/security-questions-to-ask-an-ai-governance-vendor).

## What does the law already require in a processor contract?

If the vendor processes personal data for you under the EU GDPR, Article 28(3) sets the baseline. The contract must say the processor:

1. Processes data only on your documented instructions, including for transfers outside the EU.
2. Ensures its staff are bound by confidentiality.
3. Takes the security measures required by Article 32.
4. Uses subprocessors only with your prior written authorisation, and flows the same terms down.
5. Helps you respond to data subject requests.
6. Helps you with security, breach notification and impact assessments.
7. Deletes or returns all personal data at the end of the service.
8. Gives you the information needed to show compliance, and allows audits.

Article 33(2) adds that a processor must notify you of a personal data breach without undue delay after becoming aware of it. Most vendor templates cover these points. The gaps are in the AI-specific details below.

## Which AI-specific clauses should you add?

| Clause | What to require | Why it matters for governance tools |
|---|---|---|
| Data categories | A list of what is collected: metadata, full prompts, outputs, tool arguments, files | Metadata-only and full-content modes carry very different risk |
| Content capture controls | Ability to turn off or redact prompt and output storage per agent | Lets you govern sensitive agents without storing their content |
| Retention | Fixed periods per data category, configurable by you | Traces kept "indefinitely" become a liability |
| No training | No use of your data to train or fine-tune any model, by the vendor or its subprocessors | Prompts often contain confidential material |
| Model subprocessors | Named list of any model providers the vendor calls, such as for LLM judges, with region | Evaluation features may send your content to a third party |
| Subprocessor changes | Advance notice and a right to object | Required by Article 28(2) for general authorisation |
| Residency | Storage and processing regions stated per data category | Needed for residency rules and transfer assessments |
| Transfers | Standard contractual clauses or another valid mechanism where data leaves the EU | Chapter V GDPR |
| Breach notice | A fixed number of hours, not only "without undue delay" | Gives your own 72-hour clock room to work |
| Access by vendor staff | Role-based, logged, and only for support you request | Governance data shows your whole AI estate |
| Deletion | Deletion of all copies, including backups, within a set period, with written confirmation | Article 28(3)(g) and exit planning |
| Audit | Recent independent reports and a right to audit on cause | Article 28(3)(h) |

## How does deployment model change the terms?

Where the tool runs changes how much of the above applies.

1. **Vendor SaaS.** All clauses apply. Focus on retention, model subprocessors and residency.
2. **Customer-controlled infrastructure.** Your data may never leave your environment. The DPA then mainly covers support access, telemetry the product sends back, and any data shared during troubleshooting.
3. **Hybrid.** Map which components run where, and apply the full clauses to anything hosted by the vendor.

Ask for a data-flow diagram per deployment model and attach it to the contract. It removes ambiguity about what "processing" means in practice.

## What should you check before signing?

1. Compare the DPA's data categories with what the product actually collects in a pilot.
2. Confirm the retention settings exist in the product, not only in the contract.
3. Check whether evaluation or guardrail features call an external model, and where.
4. Confirm you can redact personal data before it is stored. See [redacting PII at the gateway](/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway).
5. Align the DPA with the [SLA terms](/blog/sla-terms-for-ai-governance-platforms) on incident notice and support access.

## How Opencontroller does this

Opencontroller can run as Lyzr SaaS or on customer-controlled infrastructure, which lets you choose whether prompts and traces leave your environment. The LLM Gateway can redact content in the traffic path before it is stored or sent to a model. See the [evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does a governance vendor need a DPA if it only stores metadata?**
Usually yes. Metadata such as user IDs, email addresses and IP addresses is still personal data under the GDPR.

**Can we stop a governance tool from storing prompts at all?**
Ask for it. Many tools offer metadata-only modes or per-agent content capture settings. Confirm the setting in a pilot.

**Can Opencontroller run inside our own infrastructure?**
Yes. Opencontroller can be deployed as Lyzr SaaS or on customer-controlled infrastructure.

## Key facts

GDPR Article 28(3) processor contract requirements: documented instructions, confidentiality, Article 32 security, subprocessor authorisation, data subject request help, security and breach assistance, deletion or return, audits. Article 33(2): processor notifies controller of breaches without undue delay. AI-specific DPA clauses for governance vendors: data categories (metadata vs full prompts), content capture controls, retention per category, no training on customer data, named model subprocessors (e.g. for LLM judges), residency, transfer mechanism (SCCs), fixed-hour breach notice, vendor staff access, deletion including backups, audit. Opencontroller (by Lyzr) deploys as Lyzr SaaS or on customer-controlled infrastructure; its LLM Gateway can redact in the path. From $250K per deployment per year.

## Related

- [What security questions should you ask an AI agent governance vendor about itself?](https://opencontroller.lyzr.com/blog/security-questions-to-ask-an-ai-governance-vendor.md)
- [What SLA should you ask for from an AI governance platform?](https://opencontroller.lyzr.com/blog/sla-terms-for-ai-governance-platforms.md)
- [How to evaluate an AI agent governance platform: 30 questions for your RFP (2026)](https://opencontroller.lyzr.com/blog/ai-agent-governance-platform-evaluation-checklist.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
