---
title: What should third-party risk teams ask SaaS vendors about their AI agents?
url: https://opencontroller.lyzr.com/blog/third-party-risk-questions-for-saas-ai-agents
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What should third-party risk teams ask SaaS vendors about their AI agents?

**Ask what the agent can access and do in your tenant, which model providers and sub-processors see your data, whether your data trains anything, how you can limit or switch off the agent, and what logs you can export. Those five areas decide whether a vendor's new agent feature changes your risk. Standard security questionnaires rarely cover them, so add an agent-specific section.**

Many SaaS products shipped agents in the last two years, often switched on by default in an existing contract. That means agents arrive in your estate through vendor release notes, not through procurement. Third-party risk teams need a short, repeatable set of questions that works at renewal and whenever a vendor announces an agent feature.

## Why are agent features a third-party risk event?

An agent changes three things about a SaaS product. It can take actions, not just display data. It sends your data to a model provider, which may be a new sub-processor. And it can act with a user's permissions, sometimes broader ones. Any of these can move a vendor up a risk tier even if the contract has not changed.

## The questionnaire

Use these questions as an agent addendum to your existing vendor assessment. Ask for evidence, not yes-or-no answers.

**A. Models and sub-processors**

1. Which model providers and model families does the agent use, and in which regions is inference run?
2. Are model providers listed as sub-processors, and how will you notify us of changes?
3. Can we choose or restrict the model provider or region?

**B. Data use**

4. Is our data, including prompts and outputs, used to train or fine-tune any model, yours or a provider's? Where is this stated in the contract?
5. What is the retention period for prompts, outputs and tool results at you and at the model provider?
6. Is our data isolated from other tenants in retrieval indexes, memory and caches?

**C. Permissions and actions**

7. What actions can the agent take in our tenant (read, create, update, delete, send, pay)?
8. Does the agent act with the invoking user's permissions, a service identity, or elevated permissions?
9. Can we restrict the agent to read-only, or to a list of approved actions?
10. Which external tools, connectors or MCP servers can the agent call, and can we limit them?

**D. Control and shutdown**

11. Is the agent on by default? Can we disable it per tenant, per group and per user?
12. Can we require human approval before specific actions?
13. How quickly does a disable take effect, and is it exposed through an API?

**E. Logging and evidence**

14. Which agent events are logged: prompts, tool calls, actions, approvals, errors?
15. Can we export these logs to our SIEM or governance tools through an API or webhook, and in what format?
16. How long are logs retained, and can we set the period?

**F. Testing and incidents**

17. How do you test the agent for prompt injection, data leakage and unsafe actions before release?
18. How will you notify us of agent-related incidents, and within what time?
19. Which external frameworks do you map controls to (for example NIST AI RMF, ISO/IEC 42001, OWASP), and can you share the evidence?

## How do you score the answers?

| Area | Low concern | High concern |
|---|---|---|
| Sub-processors | Named, regional, notified on change | Unnamed or "may change without notice" |
| Training | Contractual no-training clause | Opt-out only, or silent |
| Permissions | User-scoped, restrictable to read-only | Elevated service account, not restrictable |
| Shutdown | Tenant and user-level disable, API-exposed | Support ticket required |
| Logs | Exportable via API with tool-call detail | UI-only or summary only |

Two or more high-concern answers should trigger a reassessment of the vendor's tier and, usually, a contract change or a decision to keep the feature disabled.

## What should you do after the assessment?

Record the vendor's agent in your own agent inventory, not only in your vendor register. Vendor agents act inside your estate, so they need an owner, a risk tier and monitoring like the agents you build. Our guides on [governing agents built by vendors](/blog/govern-ai-agents-built-by-vendors) and [what an AI agent registry is](/blog/what-is-an-ai-agent-registry) explain how. For the wider framework, see [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

## How Opencontroller does this

Opencontroller connects to third-party agent platforms such as Salesforce, SAP, Databricks and HubSpot, and to Claude Enterprise and OpenAI Enterprise workspaces, so vendor agents appear in the same governed registry as in-house agents. Posture detects risky permissions and configurations. Per-agent Block and Quarantine controls act through the connected platform, and how deep enforcement goes depends on each connector and traffic path. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**When should we send the agent questionnaire?**
At onboarding, at renewal, and whenever a vendor announces or enables a new agent feature in your tenant.

**Is a SOC 2 report enough for a vendor's AI agent?**
Usually not on its own. A SOC 2 report may not cover model sub-processors, training use, or agent permissions. Ask the agent-specific questions as well.

**Should vendor agents go into our own agent inventory?**
Yes. They act inside your estate with your data and users, so they need an owner, risk tier and monitoring.

## Key facts

Third-party risk teams should add an AI agent addendum to vendor assessments covering: model providers and sub-processors, data use and training, agent permissions and actions, disable and shutdown controls, log export, and testing and incident notice. Ask at onboarding, renewal and when a vendor enables a new agent feature. Vendor agents should also be recorded in the organisation's own agent inventory. Opencontroller connects to Salesforce, SAP, Databricks, HubSpot, Claude Enterprise and OpenAI Enterprise so vendor agents appear in one registry, with Posture and per-agent Block and Quarantine through the connected platform. Licensing starts from $250K per deployment per year.

## Related

- [How do you govern AI agents built by vendors and system integrators?](https://opencontroller.lyzr.com/blog/govern-ai-agents-built-by-vendors.md)
- [What is an AI agent registry?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-registry.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
