---
title: How do you apply different AI agent policies to different business units?
url: https://opencontroller.lyzr.com/blog/different-ai-agent-policies-by-business-unit
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you apply different AI agent policies to different business units?

**Write one global baseline that every agent inherits, then add business-unit overlays that can only tighten the baseline or grant specific, approved exceptions. Scope overlays by attributes the agent already carries, such as its owning group in your identity provider, a cloud tag, or a registry field. Resolve conflicts with a fixed rule: deny wins unless an approved exception names the agent.**

Uniform rules fail quickly in a large enterprise. A sales research agent that cannot browse the web is useless. An underwriting agent that can browse the web is a data-leak and model-risk question. The answer is not two separate governance programmes. It is one policy model with layers.

## Why not just write separate policies per department?

Separate policy sets drift. Within a year, each business unit has its own definition of "sensitive data", its own approval steps and its own exceptions nobody remembers. Auditors then ask for evidence per department, and nobody can say what the company-wide floor is.

A layered model keeps one floor and makes differences explicit:

- **Baseline:** rules every agent follows, whatever the department. Example: no customer personal data sent to unapproved model providers; every agent has a named owner.
- **Overlay:** rules for a business unit, data class or risk tier. Example: underwriting agents cannot call external web tools.
- **Exception:** a named, time-bound waiver for one agent, with an approver and expiry date.

## How do you decide which overlay applies to an agent?

An overlay needs a reliable way to know which business unit an agent belongs to. Use attributes in this order of trust:

| Attribute | Where it comes from | Strength | Weakness |
|---|---|---|---|
| Registry field (business unit, risk tier) | Set at registration, approved by owner | Explicit and reviewable | Only as good as registration discipline |
| Owner's IdP group | Identity provider via SCIM or directory sync | Updates when people move | Owner may sit in a shared platform team |
| Cloud account, subscription or project | Cloud organisation structure | Hard to fake | Shared accounts blur boundaries |
| Resource tags or labels | Cloud tags, Kubernetes labels | Easy to apply at scale | Easy to omit or mistype |

Use the registry field as the source of truth. Use the others to check it. If an agent's registry says "Sales" but it runs in the underwriting cloud account, flag it for review rather than guessing.

Agents with no business unit should fall back to the strictest overlay, not the loosest. That gives owners a reason to register properly.

## Worked example: web access for sales, none for underwriting

Here is a simple layered policy for outbound web access.

**Baseline (all agents):**

1. Outbound web access is denied by default.
2. Every allowed destination is logged per agent.
3. No customer personal data in requests to external web tools.

**Sales overlay:**

1. Agents tagged business unit = Sales may use the approved web search and fetch tools.
2. Allowed domains: the public web, minus a blocklist of file-sharing and paste sites.
3. Requests pass through prompt redaction at the gateway before any web tool call.

**Underwriting overlay:**

1. Agents tagged business unit = Underwriting may not use web search or fetch tools.
2. Allowed external destinations: only the approved data providers on a named list.
3. Any attempt to call a web tool raises a high-severity alert to the agent owner.

**Exception:**

- One underwriting research agent may fetch from two named regulator websites until 31 March, approved by the head of model risk.

With this structure, the question "can this agent reach the internet?" has one answer per agent, and the reason is traceable to a layer.

## How should conflicts between overlays resolve?

An agent can match more than one overlay. A sales agent might also be tagged high-risk because it handles payment data. You need a conflict rule that everyone can predict.

A rule set that works for most enterprises:

1. **Deny beats allow** across overlays at the same level.
2. **The more specific scope wins** only when it was explicitly approved as an exception for that agent.
3. **Risk-tier overlays sit above business-unit overlays.** A high-risk tier rule overrides a business-unit allowance.
4. **Exceptions expire.** An expired exception falls back to the overlay automatically.
5. **Every decision records which layer decided.** This is the evidence trail auditors want.

These are the same patterns used in mature policy engines and cloud organisation policies, where an organisation-level deny cannot be overridden by an account-level allow.

## How do you test business-unit policies before enforcing them?

Run new overlays in observe mode first. Log what would have been blocked, review it with the business unit for one to two weeks, then switch to enforce. Treat policies as code: keep them in version control, review changes, and run tests against sample agent records. See [what is policy as code for AI agents](/blog/what-is-policy-as-code-for-ai-agents) for the workflow and [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents) for the tier model the overlays depend on. For the network side of the example above, see [how to control AI agent network egress](/blog/control-ai-agent-network-egress).

## How Opencontroller does this

Opencontroller supports policy as code that is extensible and applies across the agent estate, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. The governed registry holds each agent's owner, versions, tools and MCP servers, and OIDC SSO ties people to groups. Where traffic passes through the Lyzr Gateway, policies can allow, route, redact or block in the path. Enforcement depth depends on the connector and traffic path. Read more in [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should business units be allowed to loosen the global AI agent baseline?**
Only through a named, time-bound exception with an approver. Overlays should tighten the baseline or grant specific, reviewed allowances, never silently remove baseline rules.

**What happens to an agent with no business unit assigned?**
Apply the strictest overlay until the owner registers it properly. Defaulting to the loosest rules rewards incomplete registration.

**How do you prove to auditors which policy applied to an agent?**
Log every policy decision with the agent, version, rule, the layer that decided (baseline, overlay or exception) and a timestamp.

## Key facts

To apply different AI agent policies per business unit, use one global baseline every agent inherits, business-unit overlays that tighten it or grant approved allowances, and named time-bound exceptions. Scope overlays from a registry business-unit field, checked against the owner's IdP group, cloud account and tags; unassigned agents get the strictest overlay. Resolve conflicts predictably: deny beats allow, risk-tier overlays sit above business-unit overlays, exceptions expire, and every decision records which layer decided. Opencontroller by Lyzr provides extensible policy as code across the agent estate, a governed registry and in-path controls (allow, route, redact, block) via its gateway.

## Related

- [What is policy as code for AI agents?](https://opencontroller.lyzr.com/blog/what-is-policy-as-code-for-ai-agents.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [How do you control which websites and APIs an AI agent can reach?](https://opencontroller.lyzr.com/blog/control-ai-agent-network-egress.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
