---
title: What should you do in the first 30 days of building an AI agent inventory?
url: https://opencontroller.lyzr.com/blog/first-30-days-of-ai-agent-inventory
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What should you do in the first 30 days of building an AI agent inventory?

**Spend the first 30 days on four sweeps, one per week: cloud accounts, SaaS agent platforms, identity and network logs, then developer devices. Each week ends with a clear exit test, so by day 30 you have a named owner for most agents found and a list of what you still cannot see. Do not aim for perfect coverage in month one; aim for a defensible baseline you can re-run.**

Most teams stall because they start with a spreadsheet and a survey. Surveys find the agents people remember. Systems find the rest. The plan below uses systems first and people second.

For the broader method, see the pillar guide on [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).

## Why start with a 30-day plan at all?

An inventory is the base for every other control. NIST's AI Risk Management Framework names it directly: GOVERN 1.6 calls for mechanisms to inventory AI systems, resourced according to organisational risk priorities. You cannot risk-tier, review access or set runtime policy for an agent you have not recorded.

A fixed 30-day window does two things. It forces scope decisions early. And it gives security and audit a date by which they can expect a first answer to "how many agents do we have?"

## What is the week-by-week plan?

| Week | Sweep | Main sources | Exit criteria |
|---|---|---|---|
| 1 | Cloud accounts | AWS, Azure and Google Cloud AI services, model endpoints, API keys, audit logs | Every cloud account in scope has been scanned once; read-only discovery roles exist |
| 2 | SaaS agent platforms | Salesforce, SAP, Databricks, HubSpot, ServiceNow, Claude and ChatGPT enterprise workspaces | Admin export or connector run per platform; agents listed with platform owner |
| 3 | Identity and network | IdP sign-in and consent logs, OAuth app grants, egress and proxy logs to model APIs | Unknown OAuth apps and direct LLM API traffic mapped to a team |
| 4 | Developer devices and triage | MDM inventory, coding agents, local MCP servers, CI/CD | Device findings merged; every record has an owner or an "unowned" flag; gaps list signed off |

### Week 1: cloud accounts

Start where the evidence is cleanest. Cloud providers expose AI resources through APIs and record calls in audit logs.

1. List every account, subscription and project in scope. Include sandboxes.
2. Create read-only discovery roles. Our guide to [discovery permissions for AWS, Azure and Google Cloud](/blog/ai-agent-discovery-permissions-aws-azure-google-cloud) lists the minimum rights.
3. Enumerate agents, model deployments, endpoints and API keys.
4. Pull 30 days of audit logs for model invocation calls to catch usage that has no long-lived resource.

**Exit test:** every in-scope account scanned at least once, with failures listed by name.

### Week 2: SaaS agent platforms

Many agents never touch your cloud accounts. They live inside CRM, ERP, data and productivity platforms.

1. List the platforms that let users build agents or assistants.
2. Use admin exports or connectors to pull each platform's agents, versions and tools.
3. Record the platform owner, even where the agent owner is unknown.
4. Note which platforms give you inventory only, and which also expose usage or runtime data.

**Exit test:** one export or connector run per platform, dated and stored.

### Week 3: identity and network logs

This sweep finds what no admin console lists: scripts, side projects and third-party apps calling model APIs directly.

1. Review OAuth consent grants for AI apps in your identity provider. See [discovering agents through identity provider logs](/blog/discover-ai-agents-through-identity-provider-logs).
2. Query proxy, firewall or DNS logs for traffic to model provider domains.
3. Match source IPs, workloads and service accounts to teams.

**Exit test:** every unknown app or traffic source has either an owner or a ticket.

### Week 4: developer devices, then triage

Coding agents, agent skills and local MCP servers run on laptops. Use your MDM inventory and endpoint data to find them, then spend the rest of the week cleaning the data.

1. Merge device findings with the cloud and SaaS records.
2. Deduplicate: the same agent often appears in a platform export, a cloud log and an IdP grant.
3. Assign owners. Mark the rest "unowned" rather than guessing.
4. Write the gaps list: places you know exist but could not scan.

**Exit test:** a signed-off baseline count, an ownership rate, and a gaps list.

## What fields should each record have by day 30?

Keep the first version small. You can add fields later. A minimum set:

- Name and unique ID
- Platform or runtime where it runs
- Owner (person and team), or "unowned"
- Models used
- Tools, MCP servers and data sources it can reach
- Credentials or identity it runs as
- First seen and last seen dates
- Discovery source (cloud, SaaS, identity, network, device)

The full list of useful fields is in [AI agent inventory fields](/blog/ai-agent-inventory-fields).

## What are the common mistakes in month one?

- **Starting with a survey.** Use it in week 4 to confirm ownership, not to find agents.
- **Counting only "agents".** Models, API keys, MCP servers and skills are part of the same risk picture. See [AI asset types to inventory](/blog/ai-asset-types-to-inventory).
- **Treating the count as final.** Discovery must re-run. A one-off census is stale within weeks.
- **Blocking on day one.** Record first. Enforcement comes after owners and risk tiers exist.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to Salesforce, SAP, Databricks, HubSpot, and Claude Enterprise and OpenAI Enterprise workspaces. Opencontroller Edge, installed by IT via MDM, extends discovery to developer devices, coding agents, skills and local MCP servers. Findings land in one governed registry of agents, versions, tools and MCP servers, so weeks 1, 2 and 4 of this plan become connector runs instead of manual exports. To see it on your estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can we build an AI agent inventory in 30 days?**
You can build a defensible baseline: a count, owners for most records and a list of gaps. Full coverage takes longer and depends on how many platforms you run.

**Who should own the first 30 days?**
Usually security or an AI governance lead, with named contacts from cloud, identity, SaaS administration and endpoint teams.

**What happens after day 30?**
Schedule discovery to re-run, risk-tier what you found, and start closing the gaps list.

## Key facts

A first 30-day AI agent inventory plan runs four weekly sweeps: cloud accounts (week 1), SaaS agent platforms (week 2), identity and network logs (week 3), and developer devices plus triage (week 4). Each week has an exit test. NIST AI RMF GOVERN 1.6 calls for mechanisms to inventory AI systems. Day-30 output: baseline count, ownership rate, gaps list. Opencontroller by Lyzr auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, connects to SaaS agent platforms, and Opencontroller Edge covers developer devices. 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 discover AI agents and apps through identity provider logs?](https://opencontroller.lyzr.com/blog/discover-ai-agents-through-identity-provider-logs.md)
- [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)
- [More on Discovery and shadow agents](https://opencontroller.lyzr.com/blog/cluster/discovery-and-shadow-agents)
