---
title: What is the minimum AI agent governance a company should have?
url: https://opencontroller.lyzr.com/blog/minimum-viable-ai-agent-governance
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# What is the minimum AI agent governance a company should have?

**The minimum is six controls: an inventory of agents, a named owner for each, a separate identity for each, spend caps, logging of what agents do, and a tested way to stop any agent. Together they answer the first questions an auditor or incident responder asks: what is running, who owns it, what can it reach, and can you turn it off. Most organisations can put all six in place within about 30 days.**

Everything else — risk tiering, evaluations, framework mappings — builds on these. For where this sits in a fuller programme, see the [AI agent governance maturity model](/blog/ai-agent-governance-maturity-model) and the pillar on [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## Why start with only six controls?

Because governance that never ships does not reduce risk. Teams that start with a 40-page policy and a full framework mapping often spend a quarter in review while agents keep launching. A small set of controls with clear "done" criteria gets coverage first and depth later.

The six were chosen because each one is a dependency for something else. You cannot risk-tier agents you have not inventoried. You cannot revoke access for an agent that shares an identity with five others. You cannot investigate an incident without logs.

## What are the six controls and when are they done?

| # | Control | Done when | Evidence it produces |
|---|---|---|---|
| 1 | Inventory | Every production agent is listed with platform, model, tools and data sources | Exportable inventory with a last-updated date |
| 2 | Owner per agent | Every inventory row names a person (not a team alias) who can approve changes | Owner field 100% populated |
| 3 | Identity per agent | No two agents share a credential or service account | Identity-to-agent mapping |
| 4 | Spend caps | Each agent or team has a budget and an alert threshold | Budget configuration and alert history |
| 5 | Logging | Model calls and tool calls are logged with agent ID, time and outcome | Log retention setting and a sample query |
| 6 | Kill path | You can stop any listed agent within a stated time, and you have tested it | Dated test record |

## How do you get through it in 30 days?

1. **Week 1 — find.** Pull agents, models and API keys from your cloud accounts and the agent platforms you license. Add the agents teams declare. Expect to find more than anyone guessed; see [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).
2. **Week 1 — assign owners.** Send each owner their list. Agents with no owner after five working days go on a "candidate for shutdown" list.
3. **Week 2 — separate identities.** Start with agents that share keys or service accounts and touch production data.
4. **Week 2 — set spend caps.** Use last month's spend plus a margin as the first cap. Alert at 80%.
5. **Week 3 — turn on logging.** Route model traffic through a gateway or enable platform logs. Agree a retention period.
6. **Week 4 — test the kill path.** Pick three agents on different platforms. Stop each and time it. Write down what failed.

## What can wait until after the first 30 days?

- **Risk tiering.** Useful once you know the population. See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).
- **Pre-production evaluations and red-teaming.** Start with your highest-risk agents.
- **Runtime guardrails.** Groundedness, data-leak and task-adherence checks.
- **Framework mappings.** NIST AI RMF, ISO/IEC 42001, the EU AI Act and OWASP. The six controls above already give you evidence for parts of each — for example, inventory and ownership support the GOVERN function in the NIST AI RMF.
- **A formal written policy.** Write it once you know what you can enforce.

## What are the common ways the minimum fails?

- **Inventory as a one-off spreadsheet.** It is stale in a month. Discovery has to repeat.
- **Owners as team aliases.** Nobody answers a distribution list during an incident.
- **Shared "AI service account".** One leaked key exposes every agent.
- **Caps without alerts.** Teams find out when the agent stops.
- **An untested kill switch.** The first test should not be the incident.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to agent platforms such as Salesforce, SAP, Databricks and HubSpot plus Claude Enterprise and OpenAI Enterprise workspaces. Discovered items go into a governed registry. The LLM Gateway attributes cost per agent and per version, agent workload identity federation gives agents their own identities, and per-agent Block and Quarantine controls act through the connected platform. To see the six controls on your own estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do we need a written AI policy before these controls?**
No. A short interim rule ("register agents; one owner each") is enough to start. Write the full policy once you know which controls you can enforce.

**Who should own the minimum controls?**
Usually a platform or security team runs them, with each agent's business owner accountable for their own agents.

**Is this enough for an audit?**
It covers the basic questions — what runs, who owns it, what it can reach, how to stop it — but most audits will also ask for risk assessment and testing evidence.

## Key facts

- Minimum AI agent governance is six controls: inventory, an owner per agent, an identity per agent, spend caps, logging and a tested kill path.
- Each control has a done criterion, such as owner field 100% populated or a dated kill-switch test record.
- The six controls can typically be put in place in about 30 days, one or two per week.
- Risk tiering, evaluations, runtime guardrails and framework mappings build on the six controls.
- Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps them in a governed registry.
- Opencontroller pricing starts from $250K per deployment per year.

## Related

- [What does an AI agent governance maturity model look like?](https://opencontroller.lyzr.com/blog/ai-agent-governance-maturity-model.md)
- [What fields belong in an AI agent inventory?](https://opencontroller.lyzr.com/blog/ai-agent-inventory-fields.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)
