What is an AI inventory?
An AI inventory lists every AI system an organization builds or uses, with its owner, purpose, providers, and review date. Built by survey, it goes out of date as soon as someone adopts a tool the form didn't ask about. Build it from observation: record what devices connect to and what agents report, then have people confirm ownership and purpose.
An AI system is an application or agent together with the model providers and tools it uses. A chatbot in your product that calls the OpenAI API and two internal Model Context Protocol (MCP) servers is one AI system. So is a team's use of Claude Code with the MCP servers in its settings. The inventory is the list of those systems, and each row carries the record that governance frameworks ask for.
Every framework starts here. The National Institute of Standards and Technology (NIST) AI Risk Management Framework (AI RMF) 1.0 asks for mechanisms to inventory AI systems in GOVERN 1.6. Arcjet's evidence page also maps the inventory to ISO/IEC 42001. The EU AI Act's deployer duties for high-risk systems in Article 26 assume you know which systems you deploy. For how the inventory fits the rest of a governance program, see what is AI governance.
What fields should an AI inventory record?
An AI inventory record needs enough to answer who is accountable, what the system is for, and what it touches. The following table lists the fields that the frameworks ask for, and where each one comes from in an observed inventory.
| Field | What it records | Where it comes from |
|---|---|---|
| AI system | The application or agent, with the providers and MCP servers it uses | Observed, then confirmed by a person |
| Providers | The model providers and clients the system reaches, such as the OpenAI API | Observed |
| Tools | The MCP servers and tools the system calls | Observed from captured tool calls |
| Owner | One named person accountable for the system | Assigned by a person |
| Purpose | What the system is for, with the data classes and tools it may use | Declared by a person |
| Provider decision | Whether each provider or client is approved or prohibited | Decided by a person |
| Risk tier and EU AI Act role | How much review the system gets, and whether you act as its provider or deployer | Decided by a person |
| Last review | When someone last confirmed the record, and when it's due again | Recorded when a person marks it reviewed |
The split matters. Observation answers what exists and what it does. People answer what it's for and who answers for it. An inventory that asks people for the first half is a survey, and an inventory that guesses the second half is a list nobody stands behind.
Why does a survey-built AI inventory go stale?
A survey-built AI inventory goes stale because it records what people remembered on the day they filled in the form. Adoption doesn't wait for the next survey. A developer adds an MCP server to a project, a team switches model providers, or someone installs a desktop AI app on a Monday, and the spreadsheet doesn't change until someone asks again.
A survey also can't say how much it missed. If 40 people reply and 60 don't, you don't know whether the 60 use no AI or use the most. The people most likely to skip the form are often the ones using tools they expect to be refused, which is the use the inventory most needs to find. For the methods that find that use, see how to detect shadow AI.
Surveys still have a place. They capture purpose, intended data, and business context, which no network observation can see. Use them to fill in the record for a system you already know exists, not to discover the system.
Which sources build an AI inventory from observation?
Three Arcjet sources feed the inventory, and each sees a different part of your AI use. The following table describes what each source contributes.
| Source | What it sees | What it adds to the inventory |
|---|---|---|
| Discovery on managed devices | Connections from macOS and Windows devices to AI services, with the signed app and the person behind the local account | The providers and clients people use, such as the ChatGPT app or a command-line agent |
| The Arcjet SDK in your applications | Each application or agent integration that reports to Arcjet, called an actor, and the Guard decisions on its tool calls, prompts, and model use | The AI systems you build, and what they call at runtime |
| Coding-agent hooks | Tool calls, prompts, and model switches from Claude Code, GitHub Copilot, Cursor, OpenAI Codex, and Muse Code | The coding agents you run, and the MCP servers and tools they use |
Each source proves something different. A Discovery connection shows that an app on a device reached a provider, not which account it used or what it sent. A hook event shows what an agent ran and what the policy decided. An SDK decision shows what your own application did. Keep them as separate evidence on the same record. For how the endpoint and hook sources apply to developer machines, see shadow AI coding agents.
How does Arcjet group observations into AI systems?
Arcjet groups an actor with the providers and MCP servers it uses and suggests that grouping as an AI system. People confirm or change the suggestion. A suggestion comes from what Arcjet captures for the actor, so a support agent that calls one model provider and a ticketing MCP server appears as one suggested system with those three parts.
Confirmation is a human step on purpose. Arcjet can see that two things are used together. It can't see that they belong to the same business purpose or the same accountable person. When a suggestion is wrong, correct it before you assign an owner, because the owner then sits on the system rather than on each of its parts.
Retiring a system works through its actor. Deleting an actor revokes its keys and keeps the record, so the inventory shows when the system was retired and who retired it. That record supplies evidence for NIST GOVERN 1.7, which asks that systems can be decommissioned safely.
How do you know whether your AI inventory is complete?
You know how complete an observed inventory is by measuring its sources. Arcjet reports this as the inventory.coverage signal: healthy enrolled Discovery devices as a share of enrolled devices. If 950 of 1,000 enrolled devices are reporting, the inventory's device view covers 95% of them, and the other 50 are the place to look for use you haven't recorded.
No survey can report its own coverage in this way. A form tells you how many people answered, not how much AI use their answers missed. An observed inventory states the limit of what it saw, which is what an auditor reading GOVERN 1.6 evidence wants to know.
Read the number carefully. It measures the health of devices that are enrolled, not the share of your fleet that's enrolled. Compare the enrolled count with the device count in your mobile device management (MDM) tool to find laptops that never received the agent. For the agents you build, coverage depends on whether each one reports to Arcjet, so check your deployment list against the actors in the inventory.
How do you build an AI inventory, step by step?
The following procedure builds an inventory from observation and then adds the human record. Steps 1 and 2 run in parallel.
- Turn on Discovery on managed devices. Deploy the Arcjet endpoint agent through your MDM to macOS and Windows devices, and map local accounts to people. Start with a pilot group that includes your heaviest AI users, then widen it. Accounts that aren't mapped to a person show as Unknown user instead of inheriting the device owner.
- Connect the agents you build and run. Add the Arcjet SDK to each application or agent that calls a model, so it reports as an actor. Install coding-agent hooks through managed settings for Claude Code, GitHub Copilot, Cursor, OpenAI Codex, and Muse Code.
- Review the suggested AI systems. Confirm or correct each suggestion so that one AI system matches one business purpose. Retire actors that no longer run.
- Assign owners and purposes. Give each system one owner, who must be a member of your Arcjet team, and write its purpose with the data classes and tools it may use. For what an owner is accountable for, see how to assign owners to AI systems.
- Decide on providers. Approve or prohibit each observed provider or client, such as approving the OpenAI API and prohibiting the ChatGPT app. For how decisions are enforced, see how to approve or prohibit AI providers.
- Start the review cycle. Mark each system reviewed once its record is complete. Arcjet sets the next review 180 days later and lists the system as overdue after that. For what a review checks, see how to run periodic AI system reviews.
After the first pass, new observations appear as unreviewed records. Use the No owner and Review overdue filters on the inventory lists to find records that need attention, and let the weekly owner digest remind each owner of what's waiting.
What can't an observed AI inventory see?
An observed inventory covers only what Arcjet observes: enrolled devices, actors that report to Arcjet, and coding agents with hooks installed. The following cases are outside it:
- Devices without the agent. Personal phones, unmanaged laptops, and devices that never enrolled make no observations.
- Local models. A model that runs on the device with no network traffic makes no connection for Discovery to see.
- AI features inside approved SaaS. An AI assistant built into a tool you already use often shares the tool's hostnames, so it needs the vendor's admin settings.
- Personal versus company accounts. A hostname can't separate a personal ChatGPT account from a company workspace. Use the vendor tenant controls described in block personal AI accounts on work laptops.
- Agents that don't report. An application without the SDK, or a coding agent without hooks, isn't an actor.
- Models, prompts, and datasets as assets. Arcjet records which providers and models a system uses. Model cards, training data, and prompt versions belong in your machine learning registry, linked from the inventory record.
Record systems from these gaps by hand, and mark them as declared rather than observed so a reader knows the difference.
Where Arcjet fits
Governance frameworks ask for the same four things: a list of the AI systems you run, a named owner and purpose for each, decisions about which providers are allowed, and evidence that those decisions hold. Arcjet builds the first of these from what it already sees. Discovery finds the AI providers people use on managed macOS and Windows devices. The SDK and coding-agent hooks report the agents you build and run. Arcjet suggests how those observations group into AI systems, and people confirm them.
The owner, purpose, provider decisions, risk tier, and review date sit on that observed inventory. Only team administrators edit them, other members read them, and every change is written to the WorkOS audit log. The evidence page maps the inventory and its coverage to NIST AI RMF GOVERN 1.6, also maps them to ISO/IEC 42001, and states what it doesn't cover.
The next step after the inventory is ownership. See how to assign owners to AI systems, and for the full governance picture, see what is AI governance.
Frequently asked questions
What is the difference between an AI inventory and an AI register?
They're usually the same thing under two names. Both list the AI systems an organization builds or uses with an owner, purpose, providers, and review date. Some organizations call the record kept for regulators a register and the working list an inventory, but the fields are the same.
Should models, prompts, or datasets be inventoried separately?
Inventory them as attributes of an AI system, not as separate rows. Arcjet records which providers and models each system uses. Model cards, training data, and prompt versions belong in your machine learning registry, linked from the AI system's record, so one owner answers for all of it.
How does NIST AI RMF GOVERN 1.6 describe an AI inventory?
GOVERN 1.6 in the NIST AI Risk Management Framework 1.0 asks that mechanisms are in place to inventory AI systems, resourced according to the organization's risk priorities. An observed inventory with a stated coverage figure supplies evidence for it. It doesn't cover systems the organization hasn't connected or enrolled.
What about AI systems that Arcjet can't see?
Add them by hand and mark them as declared rather than observed. Arcjet sees enrolled macOS and Windows devices, applications that report through the SDK, and coding agents with hooks installed. Personal phones, unmanaged laptops, local models, and AI features inside approved SaaS tools need other sources or a manual record.
How often should an AI inventory be updated?
An observed inventory updates as Arcjet records new providers, actors, and MCP servers, so new use appears as unreviewed records without a survey. The human fields, such as owner and purpose, are confirmed at each review, which Arcjet sets 180 days after the last one.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
Coding agents are often the AI systems a survey misses. Install Arcjet hooks and each agent's tool calls, models, and MCP servers appear in your inventory.