---
title: How do you estimate how many AI agents your company has?
url: https://opencontroller.lyzr.com/blog/estimate-how-many-ai-agents-you-have
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How do you estimate how many AI agents your company has?

**Count by source, not by survey. List every place an agent can run or be bought (cloud accounts, SaaS agent platforms, code repositories, model API keys, developer laptops and expense data), count what each source shows, then de-duplicate and report the result as a low–high range with a confidence level per source. A range you can defend beats a single number you cannot.**

Leadership often asks for "the number" before any tooling is approved. That is a fair request. You cannot size a governance programme, a licence or a team without it. The trap is answering with a survey result. Surveys count what people remember to declare, which is usually the agents that are already approved.

## What exactly are you counting?

Agree on the unit before you count anything. Otherwise two teams will report the same assistant three times.

A workable definition: an **agent** is a deployed piece of software that uses a model to decide on and take actions through tools, APIs or other systems, with its own configuration and an owner. Under that definition:

- A chatbot that only answers questions is an AI application, not an agent. Count it separately.
- Two versions of the same agent are one agent with two versions.
- The same agent deployed in three environments is one agent with three deployments.
- A seat licence for an assistant product is a user count, not an agent count, unless users can build their own agents in it.

Count models, API keys and MCP servers alongside agents. They are the evidence that agents exist, and you will need them in the inventory anyway. Our list of [AI asset types to inventory](/blog/ai-asset-types-to-inventory) covers the full set.

## Which sources should you count, and how reliable is each?

Use this worksheet. Fill one row per source, with a low and a high count.

| Source | What you count | How to get it | Typical confidence |
|---|---|---|---|
| Cloud AI services (AWS, Azure, Google Cloud) | Agents, custom models, endpoints, deployments | Service list APIs per account and region; [cloud audit logs](/blog/find-ai-agents-with-cloud-audit-logs) for invocations | High for what exists; medium for what is used |
| SaaS agent platforms | Agents and copilots built inside each platform | Admin consoles and admin APIs of each platform | High where you hold admin rights; low elsewhere |
| Code repositories | Services importing agent frameworks or model SDKs | Code search for SDK and framework imports ([method](/blog/find-ai-agents-in-source-code)) | Medium: finds code, not deployments |
| Model provider API keys | Keys, projects and workspaces with recent usage | Provider admin consoles and usage exports | Medium: one key can serve many agents |
| Developer devices | Coding agents, local MCP servers, skills | Device management and endpoint inventory | Low to medium without a device agent |
| Expense and card data | Paid AI subscriptions outside procurement | Card and expense system exports | Low for agents; good for shadow tools |
| Identity provider | OAuth apps and service principals with AI vendor names | App grants and enterprise app lists | Medium |

## How do you turn the worksheet into one number?

Follow these steps in order.

1. **Count each source independently.** Do not try to reconcile while you collect.
2. **Record a low and a high.** Low is what you have confirmed. High adds what is likely, for example repositories with agent code but no deployment you have found yet.
3. **De-duplicate across sources.** The same agent often appears as a code repository, a cloud deployment and an API key. Match on owner, repository URL, cloud resource ID and key ID.
4. **Mark coverage gaps.** If you could not read three cloud accounts or two SaaS tenants, say so. Do not extrapolate silently.
5. **Report by tier.** Split the total into production agents with tool access, internal pilots and personal or experimental agents. Governance effort differs sharply between them.
6. **Repeat in 30 days.** The second count tells you the growth rate, which is often more useful to leadership than the first number.

An illustrative report line looks like this: "Production agents: 40–55 (high confidence for cloud, medium for SaaS). Pilots: 120–200 (code-derived, medium). Unmanaged tools found in expenses: 60 subscriptions across 14 vendors." The numbers are examples only; the format is the point.

## Why do first estimates usually come in low?

Three patterns explain most of the gap.

- **Agents inside SaaS platforms are invisible to cloud scans.** If business teams build agents in a CRM or service-management platform, only that platform's admin view shows them.
- **One API key hides many agents.** Shared keys collapse several workloads into one usage line. See [how to find leaked or unused LLM API keys](/blog/find-leaked-and-unused-llm-api-keys).
- **Developer tooling is not in the cloud at all.** Coding agents and local MCP servers run on laptops and only show up in device inventory.

The NIST AI Risk Management Framework expects organisations to have mechanisms to inventory AI systems (GOVERN 1.6). An estimate is the first step toward that inventory, not a substitute for it. For the full discovery method, read the pillar guide 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 connects to third-party agent platforms such as Salesforce, SAP, Databricks and HubSpot, plus Claude Enterprise and OpenAI Enterprise workspaces. Discovered assets land in one governed registry, so the count becomes a live inventory rather than a spreadsheet. Opencontroller Edge extends discovery to coding agents, skills and local MCP servers on developer devices. To see the count for your own estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should we count Copilot or ChatGPT seats as agents?**
No. Count seats as users of an AI tool. Count agents only where users can build and deploy their own agents inside the product, and then count those agents.

**How long does a first estimate take?**
It depends on how many accounts and platforms you can read. The slow part is usually getting read access to each source, not the counting itself.

**Is a survey of team leads still useful?**
Yes, as a cross-check. Use it to find owners and purpose for agents you already found, and to spot sources you missed. Do not use it as the count.

## Key facts

Estimating AI agent count: define the unit (agent vs AI app vs version vs deployment); count per source (cloud AI services, SaaS agent platforms, code repositories, model API keys, developer devices, expense data, identity provider OAuth apps); record low/high per source; de-duplicate on owner, repo URL, cloud resource ID, key ID; report by tier (production, pilot, personal); repeat after 30 days for growth rate. NIST AI RMF GOVERN 1.6 expects AI system inventory mechanisms. Opencontroller (Lyzr) auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and connects to Salesforce, SAP, Databricks, HubSpot, Claude Enterprise and OpenAI Enterprise. Pricing starts from $250K per deployment per year.

## Related

- [Which AI asset types should an AI inventory cover?](https://opencontroller.lyzr.com/blog/ai-asset-types-to-inventory.md)
- [How do you find AI agents in your code repositories?](https://opencontroller.lyzr.com/blog/find-ai-agents-in-source-code.md)
- [How do you find AI agents using cloud audit logs?](https://opencontroller.lyzr.com/blog/find-ai-agents-with-cloud-audit-logs.md)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
