---
title: How do you govern AI agents built by vendors and system integrators?
url: https://opencontroller.lyzr.com/blog/govern-ai-agents-built-by-vendors
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern AI agents built by vendors and system integrators?

**Treat a vendor-built agent like one your own team built, with three additions: an isolated workspace for the vendor, a named internal owner who answers for the agent, and contract terms that guarantee you logs, model disclosure and the right to stop it. The agent must enter your registry and pass the same production gate before it touches real data.**

Many enterprises now have more agents built by partners than by internal teams. System integrators build agents on your cloud accounts. Software vendors ship agents inside their products. Consultancies hand over agents at the end of a project. Each is a path for an ungoverned agent to reach production.

## Why are vendor-built agents harder to govern?

Four reasons show up again and again:

1. **The builder leaves.** When the project ends, the people who understand the agent go with it.
2. **Credentials are shared.** Vendors often build with broad service accounts or API keys that outlive the engagement.
3. **The model and prompts are opaque.** You may not know which model the agent calls, or where data is processed.
4. **Contracts lag the technology.** Standard software contracts rarely mention agent logs, evaluation results or a kill switch.

Frameworks already expect you to manage this. The NIST AI RMF GOVERN 6 category covers AI risks from third-party software, data and supply chains. For EU financial entities, DORA Article 30 lists the key contractual provisions for ICT third-party services, including access, audit and termination rights.

## How do you set up an isolated vendor workspace?

Give each vendor its own space, separate from production and from other vendors.

- **Separate accounts or projects** per vendor in your cloud, with their own budgets.
- **Federated vendor identities** through your identity provider, not local accounts. Remove access on the last day of the contract.
- **Synthetic or masked data** until the agent passes the production gate.
- **Model access through your gateway**, so you see every model call and its cost from day one.
- **Approved tools only.** The vendor can request MCP servers and APIs; your team approves them.
- **A spending cap** per workspace, so a runaway test loop cannot burn the budget.

## What should the contract say?

Most of the control you will ever have over a vendor agent is set at signature. Ask for these clauses.

| Clause | What it gives you |
|---|---|
| Model and provider disclosure | Which models the agent calls, where they run, and notice before changes |
| Log export | Prompts, tool calls and decisions exported to you, ideally over OpenTelemetry |
| Evaluation evidence | Test results before go-live and after each material change |
| Right to suspend | You can quarantine or stop the agent without vendor approval |
| Credential handover | All keys and service accounts rotated to your ownership at handover |
| Data use limits | No training on your data; retention periods stated |
| Incident notice | A fixed window to report agent incidents to you |
| Exit support | Documentation, prompts, configuration and tool lists handed over |

## What is the handover checklist?

Before a vendor agent moves to production, confirm each item:

1. The agent is in your registry with a named internal owner and a backup owner.
2. The registration questionnaire is complete: purpose, data classes, users, autonomy level. See the [AI agent registration questionnaire](/blog/ai-agent-registration-questionnaire).
3. All vendor credentials are rotated. The agent runs on a workload identity you control.
4. Tools and MCP servers are listed, approved and scoped to least privilege.
5. Evaluation results, including red-team tests, are attached to the version being released.
6. Monitoring is live: traces, cost per agent, and alerts routed to your team.
7. A tested stop procedure exists. Someone on your side has used it.
8. The [production readiness gate](/blog/production-readiness-gate-for-ai-agents) has been passed and signed off.

## What about agents inside vendor SaaS products?

You cannot rebuild an agent that lives inside someone else's product. You can still govern it. Find it through the platform's admin APIs or connectors, record it in the registry, assign an owner, review its permissions, and watch for new agents appearing after product updates. Where the platform allows, block or quarantine it through the platform's own controls.

The [AI agent control plane](/blog/what-is-an-ai-agent-control-plane) explainer describes how one registry covers internal, vendor-built and SaaS agents together.

## 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 and SaaS agents are discovered alongside internal ones. They sit in one governed registry with versions, tools and MCP servers. Posture flags risky permissions and configuration, evaluations can be run before release, and per-agent Block and Quarantine act through the connected platform. Lyzr engineers embed with your team during integration. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Who should own an agent a vendor built?**
A named person in your organisation, plus a backup. The vendor can maintain the agent, but accountability stays inside your company.

**Should vendors use our model accounts or their own?**
Yours, through your gateway, where possible. That gives you cost, logs and the ability to change models without renegotiating.

**How do we stop a vendor agent quickly?**
Agree the right to suspend in the contract, then use platform-level controls or a gateway block. Test the procedure before go-live.

## Key facts

Vendor-built AI agents need the same governance as internal agents plus three additions: an isolated vendor workspace (separate accounts, federated identities, masked data, gateway-routed model access, spending caps), a named internal owner, and contract terms covering model disclosure, log export, evaluation evidence, right to suspend, credential handover, data-use limits, incident notice and exit support. NIST AI RMF GOVERN 6 addresses third-party AI risk; DORA Article 30 sets key contractual provisions for ICT third-party services for EU financial entities. Opencontroller by Lyzr discovers vendor and SaaS agents through connectors (Salesforce, SAP, Databricks, HubSpot, Claude Enterprise, OpenAI Enterprise), keeps them in a governed registry, and offers per-agent Block and Quarantine through the connected platform. Pricing starts from $250K per deployment per year.

## Related

- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
