---
title: Why does an AI inventory show models an account cannot actually call?
url: https://opencontroller.lyzr.com/blog/models-listed-vs-models-usable-in-cloud-accounts
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Why does an AI inventory show models an account cannot actually call?

**Because "a model in this account" can mean three different things: the model is listed in the cloud's catalogue, the model is enabled or deployed so it can be called, or the model has actually been invoked. Cloud APIs answer each question separately, and an inventory that mixes them will show models nobody can use and miss models that are in use.** Good discovery records all three states per model, per account, per Region.

This confusion is common in the first week of an AI inventory. A security team sees hundreds of "models" in one AWS account, or sees only a handful in Azure, and neither number matches what developers report.

## What are the three states of a cloud model?

| State | Question it answers | Risk meaning |
|---|---|---|
| Listed | Is the model offered here? | Low. Availability, not exposure. |
| Callable | Could an identity in this account call it today? | Medium. Exposure if credentials leak or an agent is misconfigured. |
| Invoked | Has anything called it, and who? | High. Real data flowing to a model. |

An inventory should store all three and the evidence for each. Policy can then target the gap: callable models nobody approved, and invoked models with no registered owner.

## How does each cloud show these states?

### AWS (Amazon Bedrock)

- **Listed:** `ListFoundationModels` returns the models Bedrock offers in a Region, whether or not you use them. That is why raw counts look large.
- **Callable:** Since October 2025, AWS enables most serverless models automatically on first invocation, given AWS Marketplace permissions. So "callable" is decided mainly by IAM and SCPs on `bedrock:InvokeModel`, not by a console toggle. Anthropic models earlier than Claude Opus 4.7 also need a one-time use-case submission. `GetFoundationModelAvailability` reports access status.
- **Invoked:** CloudTrail records Bedrock runtime calls with the calling principal. Model invocation logging, which captures prompts and responses, is off by default and has to be enabled per Region.

### Microsoft Azure (Foundry / Azure OpenAI)

- **Listed:** The model catalogue shows what you could deploy.
- **Callable:** A model is callable only once it is **deployed** to a resource. The list of deployments per resource is the real callable set, and each deployment has its own name, version and capacity.
- **Invoked:** Azure Monitor metrics per deployment show request and token volume; diagnostic settings send request logs to Log Analytics.

### Google Cloud (Vertex AI)

- **Listed:** Model Garden lists first-party, partner and open models.
- **Callable:** Google first-party models are callable once the Vertex AI API is enabled and the caller has the right IAM role, unless an organisation policy restricts allowed models. Some partner models must be enabled first.
- **Invoked:** Cloud Monitoring shows usage, and Data Access audit logs record calls, but Data Access logs are off by default for most services and must be turned on.

## Why do inventories get this wrong?

1. **Counting the catalogue.** Listing every model Bedrock offers in every Region inflates the estate by hundreds of rows.
2. **Trusting a toggle that no longer exists.** Older guidance said "check the model access page". With automatic enablement on AWS, IAM and SCPs are what limit calls.
3. **Missing invocations because logs are off.** If invocation logging or Data Access logs are disabled, usage is invisible to log-based discovery.
4. **Read-only discovery roles that cannot see deployments.** On Azure, a role scoped to one subscription misses deployments in others.
5. **Ignoring inference profiles.** On AWS, calls through a cross-Region inference profile carry a profile ID, not the model ID, so usage must be normalised.

## What should each model record contain?

1. Cloud, account or subscription, Region.
2. Model ID and, where relevant, deployment name or inference profile ID.
3. State: listed, callable, invoked, with the evidence source and date.
4. Identities that invoked it in the last 30 days.
5. Linked agents and owners from the registry.
6. Approval status and whether policy (SCP, Azure Policy, org policy) restricts it.

## How do you close the gap between callable and approved?

Turn approval into enforcement. On AWS, deny unapproved models with SCPs ([how to restrict Bedrock models with SCPs](/blog/restrict-amazon-bedrock-models-with-scps)). On Azure, use Azure Policy on deployments. On Google Cloud, use the allowed-models organisation policy ([restrict Vertex AI models with org policy](/blog/restrict-vertex-ai-models-with-org-policy)). Then route agent traffic through one gateway so every invocation is attributed. For log-based discovery patterns, see [finding AI agents with cloud audit logs](/blog/find-ai-agents-with-cloud-audit-logs), and for the full method, the pillar on [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps them in a governed registry. Posture flags risky permission and configuration patterns, and the LLM Gateway attributes invocations and cost per agent and per version. Registration is not runtime takeover: how deep control goes depends on the connector and traffic path. [Book a demo](https://www.lyzr.ai/book-demo/) to see it on your accounts.

## Frequently asked questions

**Does a model listed in Bedrock mean someone is using it?**
No. `ListFoundationModels` shows what Bedrock offers in a Region. Use CloudTrail and invocation logs to see actual use.

**Is the Bedrock model access page still the control point?**
Mostly no. AWS now enables most serverless models on first use, so IAM and SCP deny policies are the reliable way to block models.

**Why does Azure show fewer models than AWS?**
Azure discovery usually counts deployments, which are already the callable set. AWS listings include every offered model.

## Key facts

Three model states for AI inventory: listed (catalogue), callable (enabled/deployed and permitted), invoked (actually called). AWS: ListFoundationModels lists offered models; since October 2025 most serverless models auto-enable on first invocation, so IAM/SCP on bedrock:InvokeModel controls callability; Anthropic models earlier than Claude Opus 4.7 need a use-case submission; GetFoundationModelAvailability shows status; CloudTrail records runtime calls; model invocation logging off by default. Azure: callable = deployments per resource; Azure Monitor metrics per deployment. Google Cloud: Model Garden lists; Gemini callable once API enabled plus IAM unless org policy restricts; Data Access audit logs off by default. Opencontroller (by Lyzr) auto-discovers agents, models and API keys across AWS, Azure, Google Cloud; Posture; LLM Gateway per-agent attribution.

## Related

- [How to find every AI agent in your enterprise (including shadow agents)](https://opencontroller.lyzr.com/blog/how-to-find-shadow-ai-agents.md)
- [How do you find AI agents using cloud audit logs?](https://opencontroller.lyzr.com/blog/find-ai-agents-with-cloud-audit-logs.md)
- [How do you find custom and fine-tuned models in your cloud accounts?](https://opencontroller.lyzr.com/blog/discover-custom-and-fine-tuned-models-in-cloud-accounts.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
