---
title: How do you discover AI usage when you cannot get admin API keys?
url: https://opencontroller.lyzr.com/blog/discover-ai-usage-without-admin-api-keys
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# How do you discover AI usage when you cannot get admin API keys?

**You can still find most of it. Combine cloud billing and read-only cloud APIs (for models bought through AWS, Azure or Google Cloud), identity provider logs (for SaaS AI apps), egress or DNS logs (for direct calls to model APIs), secret scanning (for keys in code) and expense data (for personal subscriptions). What you lose without the admin key is exact per-key usage for that provider's direct API, so plan to close that gap by routing new traffic through a gateway.**

Provider admin keys are powerful. They can usually read usage for the whole organisation and manage members and keys, so security teams are right to be careful about handing one to any tool. That is a common reason discovery stalls. It does not have to.

## What does a provider admin key give you, and why is it hard to get?

Direct model APIs typically reserve organisation-wide usage and cost reporting for admin credentials:

- **Anthropic.** The Usage and Cost API requires an Admin API key (prefix `sk-ant-admin01-`) or another admin credential. Ordinary workspace API keys are rejected.
- **OpenAI.** Organisation usage and cost endpoints require an admin key rather than a project key.

The same keys can often create users or keys, which is why owners hesitate. Ask whether a read-only or reporting-only admin credential exists for your plan; if not, use the signals below.

## Which other signals reveal AI usage?

Ranked roughly by how much they cover.

| # | Signal | What it shows | What it misses | Access needed |
|---|---|---|---|---|
| 1 | Cloud billing and cost exports | Spend on Amazon Bedrock, Azure AI services, Vertex AI, and marketplace AI subscriptions, by account or project | Which agent or person; usage paid outside the cloud bill | Billing read |
| 2 | Read-only cloud control-plane APIs | Deployed agents, model endpoints, inference profiles, keys stored in secret managers | SaaS AI and direct provider APIs | Read-only roles per account or subscription |
| 3 | Identity provider logs and OAuth grants | Sign-ins to SaaS AI apps; apps granted access to mail, files or calendars | Tools used without SSO; API traffic | IdP log and app-grant read |
| 4 | Egress proxy, firewall or DNS logs | Which hosts call model API domains, how often, and how much data | Content (unless TLS is inspected); which agent inside a host | Network log access |
| 5 | Secret scanning | Provider API keys committed to repositories | Keys held only in vaults or local files | Repository scanning |
| 6 | Expense and procurement data | Personal or team subscriptions to AI tools | Free tiers; shared accounts | Finance data export |

For the cloud permissions in row 2, see our list of [AI agent discovery permissions for AWS, Azure and Google Cloud](/blog/ai-agent-discovery-permissions-aws-azure-google-cloud). For browser-based tools, which mostly appear in rows 3 and 4, see [how to discover AI use in the browser](/blog/how-to-discover-ai-use-in-the-browser).

## How do you combine the signals into an inventory?

1. **Start with cloud billing.** It is the fastest way to size spend and find the accounts that matter.
2. **Run read-only discovery in those accounts** to turn spend into named agents, endpoints and keys.
3. **Join egress logs to your asset database** so a call to a model API maps to a host, a workload and an owner.
4. **Add IdP grants** to catch SaaS agents and AI apps with access to company data.
5. **Feed secret-scanning hits** into the same list, marking each key as "found in code" until its owner confirms it.
6. **Deduplicate by owner and purpose.** The same agent often appears in three signals.
7. **Repeat on a schedule.** A one-off sweep goes stale within weeks; see [how often AI agent discovery should run](/blog/how-often-should-ai-agent-discovery-run).

GitHub's secret scanning, for example, recognises token formats from many providers through its partner and supported-pattern programme, which makes row 5 cheap to switch on for code you host there.

## How do you close the remaining gap?

The one thing these signals cannot give you is precise per-key usage for a provider's direct API. Two approaches close it over time:

- **Route new traffic through a gateway.** Every call then carries the agent, version, model, tokens and cost, regardless of which provider serves it.
- **Revisit the admin key with a narrower ask.** Once the inventory shows real spend, owners are more willing to grant a reporting credential for that one purpose.

Our pillar guide on [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents) shows how these steps fit into a full discovery programme.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud using cloud APIs, and connects to Claude Enterprise and OpenAI Enterprise workspaces and to platforms such as Salesforce, SAP, Databricks and HubSpot. Discovery is read-first: registration does not mean runtime takeover, and the depth of control depends on the connector and traffic path. The LLM Gateway (Lyzr Gateway) exposes OpenAI-compatible endpoints and attributes cost per agent and per version, which closes the direct-API gap as traffic moves onto it. Opencontroller Edge extends discovery to coding agents, skills and local MCP servers on developer devices, installed by IT via MDM. [Book a demo](https://www.lyzr.ai/book-demo/) to see discovery on your estate.

## Frequently asked questions

**Can you see Anthropic or OpenAI API usage without an admin key?**
Not the provider's own organisation-wide report. You can see spend paid through clouds, calls in egress logs, and keys in code.

**What is the fastest first signal for AI spend?**
Cloud billing. It shows Bedrock, Azure AI and Vertex AI spend by account or project within a day.

**Do egress logs show what was sent to the model?**
Usually not. They show which host called which API and how much data moved, unless TLS inspection is in place.

## Key facts

Without a provider admin key (Anthropic's Usage and Cost API requires an Admin API key with prefix sk-ant-admin01-; OpenAI organisation usage endpoints require an admin key), AI usage can still be discovered through cloud billing exports, read-only cloud control-plane APIs, identity provider sign-in logs and OAuth grants, egress or DNS logs, secret scanning of repositories, and expense data. The remaining gap, exact per-key direct API usage, closes by routing traffic through a gateway. Opencontroller by Lyzr auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, connects to Claude Enterprise and OpenAI Enterprise workspaces, attributes cost per agent through its LLM Gateway, and extends to devices with Opencontroller Edge. Pricing starts from $250K per deployment per year.

## Related

- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [How do you discover employees' AI use in the browser?](https://opencontroller.lyzr.com/blog/how-to-discover-ai-use-in-the-browser.md)
- [How often should AI agent discovery scans run?](https://opencontroller.lyzr.com/blog/how-often-should-ai-agent-discovery-run.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
