How to keep an AI agent inventory

An AI agent inventory records, for every agent in the company, its owner, purpose, the data it reads, the actions it can take, its model and vendor, where data goes, its cost caps, its logging and its next review date, all in one list. As Copilot Studio agents, agents embedded in SaaS tools and custom builds multiply, that list becomes the foundation of AI governance.

How to keep an AI agent inventory

Topic

AI

Published

16 November 2026

Updated

16 November 2026

Why are agents multiplying so fast?

An AI agent is software that plans its own steps towards a goal, calls tools and can take actions in systems. Unlike a chat tool, it does not just answer; it does work. That changes the risk.

Agents arrive by three routes. Employees build their own with low-code tools such as Copilot Studio. CRM, helpdesk and HR products ship embedded agent features in new releases, often switched on with a single admin setting. And teams commission custom agents for specific processes. In Gartner’s 2027 CIO and Technology Executive Survey, 37% of respondents said their organisation had already deployed AI agents, and a further 34% planned to within 12 months.

The result is familiar: within a year, dozens of agents, each in a different team, reading different data and running on a different vendor’s model. Nobody can say at a glance which one reads which customer records or which one can send email. As agentic AI spreads, an inventory becomes the first step of governance.

What are Microsoft and Google offering?

The large platforms have started building agent inventories into their products. Here is what they offer, in their own terms.

Microsoft Agent 365. Microsoft describes Agent 365 as a way to observe, govern and secure the agents in an organisation. According to Microsoft Learn, admins can view all their agents in a single, centralised registry, with lifecycle, access and compliance managed through the Microsoft 365 admin center, Entra and Purview. It has been generally available to commercial customers on a per-user licence since 1 May 2026. Microsoft’s announcement says it is included in Microsoft 365 E7 or sold standalone at USD 15 per user per month; check Microsoft for current pricing. The same announcement introduced, in public preview, registry sync with AWS Bedrock and Google Cloud connections.

Google Cloud. According to the Gemini Enterprise Agent Platform release notes, Agent Registry and Agent Gateway became generally available in June 2026. A 26 October 2026 update extended semantic governance policies to evaluate tool calls from agents that use model providers other than Gemini. Google’s VPC Service Controls announcement describes adding agent identities to perimeter rules and writing conditions on MCP attributes such as the tool name or whether a tool is read-only; its example grants an agent read access to Workspace data while explicitly denying it the ability to send email.

These tools help, but they have two limits. Each sees mainly its own ecosystem; an agent embedded in a non-Microsoft SaaS product or a script running on a developer’s laptop may not appear on its own. And no tool knows an agent’s business purpose, its business owner or which risks were accepted on purpose. That information belongs in your inventory.

What should you record for every agent?

The inventory does not need to be a complex system; a shared spreadsheet is a fine start. What matters is that the same fields are filled for every agent.

Field What to record Why it matters
Owner A business owner and a technical owner, two named people Who to call when something goes wrong
Purpose What the agent does, in one sentence Spotting scope creep
Data it reads Systems, folders, data types; any personal data Privacy and data protection review
Tools and actions Tools and MCP servers it calls; read or write; can it send email or start a payment Knowing the worst case of an error
Model and vendor Which vendor, which model or platform Tracking vendor changes and contracts
Where data goes Region, sub-processors, international transfers Transfer and retention obligations
Cost caps Monthly budget, usage and rate limits, alert threshold Stopping surprise bills and abuse
Logging What is logged, where, for how long Post-incident review
Human approval Which actions need a person to approve Controlling irreversible actions
Review date Date of the next review Keeping the inventory from going stale

The cost cap field is not only a budget matter. In its strategic predictions published in September 2026, Gartner predicts that by 2030, 80% of organisations with public-facing AI will have experienced a cost exhaustion attack, in which attackers deliberately drive excessive AI usage to inflate costs. Gartner recommends treating AI token costs as a cybersecurity indicator and putting cost-focused security controls in place. For every customer-facing agent, per-user and per-session rate limits, a monthly spending ceiling and a spike alert should be mandatory fields.

How do you find shadow agents?

The first version of any inventory is incomplete. To find shadow agents, look at:

  1. Platform admin consoles. The agent list in the Microsoft 365 admin center and Copilot Studio environments, or Agent Registry in Google Cloud.
  2. App consents. Third-party apps users have consented to in Entra ID or Google Workspace; AI apps with access to mailboxes and files show up here.
  3. SaaS admin settings. AI or agent features switched on in CRM, helpdesk, HR and accounting tools.
  4. AI vendor bills. Payments to AI providers on company cards and expense reports; API keys opened on personal accounts often surface this way.
  5. Automation tools. Steps in n8n, Make or similar workflows that call a language model.
  6. A short survey. Ask teams whether they use any AI tool that takes actions on its own, and do not penalise honest answers.

The goal is not to shut down everything you find. It is to make each agent visible, give it an owner and fill in its fields.

When should an agent be retired?

Agents age like any other software. Consider retiring one when its owner has left and nobody has taken over; when it has gone unused for a set period, say 90 days; when its purpose has gone or overlaps with another agent; when it costs more than the value it creates; or when a review finds an unacceptable risk.

Retirement is a process too: revoke the agent’s credentials and API keys, remove its connections and permissions, archive its logs under your retention rule, and mark it “retired” in the inventory. An agent deleted from the interface but with its permissions still open is the most dangerous row in the list.

An inventory works best inside a wider AI governance framework. To choose one, see our comparison of ISO 42001 and the NIST AI RMF; for the choice between a platform and a custom build, see Copilot Studio or a custom AI agent.

Where to start

  1. Define the inventory fields using the table above and keep them in one shared place.
  2. Build the first list by scanning platform consoles, app consents, SaaS settings and bills.
  3. Give every agent a business owner and a technical owner; suspend any agent nobody will own.
  4. Prioritise customer-facing agents and those with write access: complete their cost caps, rate limits and human approval rules.
  5. Add a short request step for new agents, and give each agent a review date three or six months out.

At Corrplus, our AI governance work sets up the agent inventory, data flows and approval rules with you. For companies on Microsoft 365, we cover Copilot Studio agents and admin settings through Microsoft Copilot consulting, and our managed services keep the inventory current and watch costs and logs over time. If you want to build a new agent with these rules from day one, see AI agent development or get in touch for a free first call.

Want help with this?

Book a call to apply what “How to keep an AI agent inventory” describes to your company. We will listen and agree the first step together.

Frequently asked questions

Is an AI agent inventory the same as an AI system inventory?

An agent inventory is one part of the wider AI inventory. Agents do not just produce answers; they call tools and act in systems, so the inventory needs extra fields such as permissions, connected tools and cost caps. If you work with ISO 42001 or the NIST AI RMF, link agent records to that same inventory.

If we use Microsoft Agent 365, do we still need our own inventory?

Microsoft describes Agent 365 as a single, centralised registry for seeing agents and managing their lifecycle. But a tool cannot know an agent's business purpose, its business owner, which risks were knowingly accepted, or every agent outside its reach. Use it as a technical source, and keep decisions and accountability in your inventory.

What is a shadow agent?

An agent set up without the IT or governance team knowing. An agent an employee built in Copilot Studio, an AI feature switched on in a SaaS product, or an automation running on a personal API key all count. It need not be malicious; the problem is that nobody is watching its data or permissions.

When should an agent be retired?

When its owner has left and nobody has taken over, when it has gone unused for a set period, when its purpose has gone, when it costs more than it delivers, or when a review finds an unacceptable risk. Retiring it means revoking credentials and connections, archiving logs under your retention rule and updating the inventory.

Related services

ServiceAI governance and complianceWe inventory your AI use, classify it under KVKK, GDPR and the EU AI Act, and build the rules and technical controls.
ServiceManaged servicesWe monitor and improve your live website, automations and AI features under a written service level agreement.
ServiceMicrosoft Copilot consultingWe roll out Microsoft 365 Copilot once permissions and data are ready, and build Copilot Studio agents for specific teams.

Related articles

Let’s talk about your transformation

Tell us which work repeats every week. We will tell you honestly whether it is worth automating.

How we start

  1. A free first callWe pin down the problem and the goal together.
  2. A 2-week discovery sprintWe look at the data, the process and the tools, and decide where to start.
  3. A written proposalScope, architecture and timeline, in writing.

Example questions

Go straight to

CorrAI answers with AI from the content of this site and can be wrong. Please do not enter personal details. Privacy

Book a free call