Before anyone writes an AI strategy, find out where the AI already is. Agents built on someone's credit card or in a sandbox subscription are easy to start and hard to find later. Microsoft's Cloud Adoption Framework (CAF) puts it plainly: "Untracked or "shadow" deployments pose security and cost risks." Flexera's 2026 State of the Cloud found 45% of organisations using generative AI extensively in production, up from 36% a year earlier.
If your organisation runs Azure properly, it already has a landing zone: the shared foundations (network, security, identity and policy) every new project plugs into, plus a standard request, called subscription vending, that gives each project its own space with those guardrails already switched on. Agents fit straight into it.
Four questions, and what already answers them
CAF says "Leaders must know which agents exist, who owns them, what they can access, and how to intervene when behavior falls outside policy." It also says agents "don't require a specialized workload landing zone architecture", and that they belong to "the same leaders responsible for cloud governance, security, and compliance. Avoid creating parallel governance models."
| The question | What answers it | Who owns it |
|---|---|---|
| Which agents exist? | Every agent comes through the standard request and gets its own identity | Platform team |
| Who owns each one? | The request names a business owner and a cost centre | Business owner |
| What can it reach? | The central firewall, private connections and access rights | Platform team |
| How do we stop one? | A spending cap at the AI gateway, a policy block, or switching off its identity | Platform team |
For an exec team that makes the decision simple: agents go through the same front door as everything else. Flexera also found 47% of large enterprises setting up dedicated AI governance teams or leaders. That person can write the rules, but the rules should still be enforced through the same management groups and policies as the rest of the estate.
What the landing zone gains
Most of it stays as it is. This is the standard Azure landing zone layout, with what agents add marked in green:
An identity for each agent
Microsoft Entra Agent ID gives every agent its own login, much like a staff badge: you can see what it opened and switch it off on its own. CAF says to "require each agent to operate under a distinct agent identity". Applying Conditional Access to agents needs extra licensing (Microsoft 365 E7, or Agent 365 alongside Entra P1 or Microsoft 365 E3), so check that before anyone designs around it.
A gateway that caps spend
AI services bill by tokens, roughly the words going in and out. Azure budgets can warn you about overspend, but Microsoft's Foundry cost guidance says they can't stop Azure OpenAI spending ("Azure OpenAI doesn't currently provide this functionality"). An AI gateway (Azure API Management sitting in front of the models) can: its token limit policy gives each agent a monthly allowance and refuses requests once it's used.
A short set of AI policies
Policies are rules Azure enforces automatically. Three built-ins cover most of it: AI services with key access turned off, AI services with network access restricted, and Foundry using approved models only. All three can report first and block later. They cover the AI account itself. Agents under online need more, because online doesn't block public endpoints on the data services around them; part 2 covers that.
What to decide, and when
For your platform team: the inventory query and the request form
This Resource Graph query lists every Azure AI services account in the tenant and the management group it sits under. Anything under a sandbox or the tenant root was built outside the landing zone, and so was anything still taking keys over a public endpoint.
resources
| where type =~ 'microsoft.cognitiveservices/accounts'
| join kind=leftouter (
resourcecontainers
| where type =~ 'microsoft.resources/subscriptions'
| project subscriptionId, subName = name,
mg = tostring(properties.managementGroupAncestorsChain[0].displayName)
) on subscriptionId
| project name, kind, subName, mg,
keyAuthOff = properties.disableLocalAuth,
publicAccess = properties.publicNetworkAccess
| order by mg asc
Microsoft's baseline architecture for Foundry in a landing zone says the platform team "should follow a subscription vending process that uses a structured set of questions". For an agent, the request needs five sections:
AGENT LANDING REQUEST, [agent name], [date]
1. Placement
Users: internal only -> corp | internet-facing -> online
Business owner: [name, role] Cost centre: [code]
2. Network
Spoke size: [/24] Agent egress subnet: [/27]
Private endpoints: Foundry, Storage, AI Search, Cosmos DB, Key Vault
3. Egress (what the agent may reach through the hub firewall)
Grounding sources: [FQDNs]
MCP servers, tools, other agents: [FQDNs]
4. Models
Own Foundry resource | shared models through the AI gateway
Approved models: [list] Monthly token quota: [n]
5. Identity
Agent identity: [Entra Agent ID] Key access: disabled
Placement is the platform team's call: internal agents in corp, public ones in online. Egress needs the most detail, because each grounding source, MCP server or tool outside the estate becomes a firewall rule the platform team will own. Part 2 covers the network side.
The code is in blog-examples/ai-agent-guardrails: the inventory query, plus Terraform for the key-access and network policies and the gateway quota. If the AI strategy itself hasn't been written down yet, strategy before the statement of work has a one-page brief to start from.