An AI inventory sounds like paperwork, and that's why most companies don't have one. But it's the list every other AI decision depends on. You can't write a usage policy for tools you haven't listed, answer a customer's security questionnaire about tools you don't know you have, or switch off something nobody owns. And the list goes stale faster than you'd think: vendors keep switching AI on inside software you already pay for, sometimes with a date and a default you didn't choose.
What is an AI inventory?
An AI inventory is a maintained list of every AI tool, built-in AI feature and agent your company uses, with an owner, a purpose, the data it can reach and a review date for each. It's the record behind your AI rules: the policy says what's allowed, and the inventory shows what's actually running.
This article is for leaders at companies of 20 to 1,000 people who use AI rather than build it. It isn't about AI for stock control, which people also search for under the same phrase. It covers what to record, a copy-ready template, who should own the list, how to keep it current and how to build the first version in a month.
Why does a company need an AI inventory?
Because every other AI rule is applied row by row against it, and because AI arrives in companies faster than anyone approves it.
The governance frameworks start here. The NIST AI Risk Management Framework, the free US framework most mid-size companies borrow from, sets the outcome plainly: "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities" (GOVERN 1.6). Its playbook adds that "partial inventories may not provide the value of a full inventory." Our article on the NIST AI RMF covers the rest of the framework.
The risk is concrete. In IBM's 2025 Cost of a Data Breach study, based on breaches at 600 organizations from March 2024 through February 2025, "One in five organizations reported a breach due to shadow AI, and only 37% have policies to manage AI or detect shadow AI." Organizations with high levels of shadow AI saw an average of $670,000 in higher breach costs. Our article on shadow AI covers how staff end up using tools nobody approved.
The less obvious reason is that vendors add AI for you. Two examples from November 2026:
- Microsoft says it "enables Claude models on by default for most customers in commercial cloud (excluding EU/EFTA and UK)" across Copilot and related products. Its Message Center post MC1546438 says that on December 9, 2026, Claude models are also enabled for EU and EFTA tenants that haven't made a selection.
- Anthropic says the new unified Chat and Cowork experience in Claude Enterprise is "off by default in existing organizations and turns on by default on November 16, 2026."
Neither is a problem in itself. But each changes what a tool you already listed can do, and a list built once and never updated wouldn't show it.
What should an AI inventory include?
Record, for each use: what it is, what it's for, who uses it, what data it touches, whether it affects decisions about people, the vendor's terms, who owns it, its status and when it's next reviewed. Anything less and the list can't answer the questions you'll be asked.
NIST's playbook suggests a policy that defines "model or system attributes to be inventoried", and gives examples aimed at teams that build models: "documentation, links to source code, incident response plans, data dictionaries, AI actor contact information." A company that buys its AI needs a different set of fields. The vendor's terms replace the source code, and the data each tool can reach replaces the data dictionary.
One decision matters more than the column list: make each row a use, not a product. The same assistant can draft marketing copy, summarize contracts and help screen applicants. Those are three rows, because each has different data, a different owner and, for screening, different law. That's what people mean by an AI use case inventory, and it's how the template below is built.
AI inventory template: copy this
Paste this into a spreadsheet, one row per use. The example row is illustrative: a meeting assistant inside a video-call tool you already pay for, the kind of AI feature that's easiest to miss.
| Field | What to write | Example |
|---|---|---|
| Use | The job it does, in a few words | Summarizes sales calls |
| Tool and vendor | Product name and who sells it | Meeting assistant in our video-call platform |
| Plan or license | The plan you're on, or "personal account" | Business plan, AI add-on on by default |
| How it arrived | Bought, built-in feature, staff brought it, agent | Built-in feature, switched on by vendor |
| Who uses it | Team and roughly how many people | Sales, 14 people |
| Data it can reach | The data classes it sees or stores | Customer names, deal terms, call recordings |
| Decisions about people | Does it score, rank or decide about applicants, staff or customers? | No |
| Vendor terms | Training on your data, retention, and the date you checked | No training on business data; recordings kept 90 days; checked Nov 2026 |
| Owner and backup | The person who answers for this use, and a deputy | Head of Sales; Sales Ops lead |
| Status | Approved, trial, under review, or retire by a date | Approved |
| Off switch | Where it's turned off and who can do it | Admin console, IT admin |
| Next review | A date, not "annually" | February 2027 |
Three fields need a word of warning.
- Decisions about people. Mark any use that screens, ranks or scores applicants or staff. Those rows are where hiring laws apply, such as New York City's Local Law 144 and the Illinois AI law, and they need the closest look.
- Vendor terms. Write the date you checked. Terms on business plans change, and a row that says "no training" without a date is a guess.
- Off switch. If nobody can say where a use is switched off, it isn't owned yet.
What counts as AI for the inventory?
Anything that uses a model to produce text, images, code, scores or actions for company work, whoever bought it and on whatever account. In practice that's four groups:
- Tools staff chose. Chat assistants, writing tools and browser extensions, approved or not, on company or personal accounts.
- AI inside software you already pay for. Features in your office suite, CRM, help desk, applicant tracking system and meeting software. Vendors often add these as updates, not purchases, so nobody signs off on them.
- Agents and automations. Assistants that act on their own: a scheduled workflow that drafts replies, an agent built in a no-code tool, or a bot with access to a shared drive.
- Things you built. Anything that calls a model through an API, however small.
The second and third groups are the ones first inventories miss. The platforms' own consoles help. Microsoft, for example, says its Defender for Cloud Apps catalog "includes a Generative AI category" with "more than a thousand generative AI-related apps", which lets IT see which ones are used. It also lists Entra Agent ID (in preview) as a directory of agent identities created in Copilot Studio and Foundry. Use whatever your own platforms offer, but don't treat any one console as the whole list: each shows what its own product can see.
Who should own the AI inventory?
One named person owns the list, and every row has its own owner. NIST's playbook suggests policies "that define a specific individual or team that is responsible for maintaining the inventory." In a company of 20 to 1,000 people that's usually someone in operations, IT or risk, with a named backup.
The row owners matter more. The list owner keeps the spreadsheet; a row owner answers for one use. That person approves changes, knows where the off switch is and turns up at the review. A row without a name is a risk you've written down, not one you've dealt with.
An inventory is an ownership list, not a discovery report
The point of the inventory isn't finding tools. It's putting a name next to each one.
Detection tools and surveys are good at producing a long list of what's running. On their own they leave you with visibility and no owners: a spreadsheet nobody changes after the meeting where it was presented. This is why our framework treats the inventory as part of Own, not as a security exercise. Every use gets an owner and a backup, an off switch and an undo, and a review date.
The course lesson Your AI list walks through building the list from three sources, with a decision on every item. The lesson before it, Keep, replace or retire, covers the decision itself. Both are free.
How often should you update an AI inventory?
Update a row whenever something changes, review the whole list on a fixed date, and give each row its own review date. NIST's playbook frames the question as "what is an appropriate interval for checking whether it is still accurate", given each use's purpose. A quarterly review of the whole list suits most mid-size companies, with a closer interval for anything that touches decisions about people.
Between reviews, these events should change a row straight away:
- A new purchase, trial or renewal.
- A vendor announcement that switches on an AI feature or changes a default, like the two examples above.
- A new use of an existing tool, which is a new row.
- A change to the vendor's training or retention terms on your plan.
- Retiring a use. Mark it retired and keep the row; NIST's playbook notes that "Decommissioned models or systems may be stored in a model inventory along with active models, for an established length of time."
How to build your AI inventory in 30 days
01 · Name the owner and the scope
Pick the person who'll keep the list and their backup. Write the scope in one sentence: every AI tool, feature, agent and build used for company work, on company or personal accounts. This gives everyone a clear answer to "does this count?"
02 · Collect from three sources in one week
Ask staff what they use, with a promise that the point is to list tools, not to punish anyone. Pull what IT and finance can see: single sign-on, expense reports, browser extensions and the admin consoles of the software you already pay for. Then read each major vendor's recent release notes for AI features switched on by default. Our AI audit article covers each source in more detail, and the free Shadow AI Audit gives you a starting list in about five minutes. This gives you the raw list.
03 · Split it into uses and fill in the template
Turn each tool into one row per use, then fill in the twelve fields. Leave a field blank rather than guess, and mark it to check. This gives you the inventory and a short list of gaps.
04 · Put a name on every row
Ask each team lead to take the rows for their team, and agree a status for each: approved, trial, under review or retire by a date. Anything nobody will own goes to "retire". This gives every use an owner and a decision.
05 · Set the review dates and the triggers
Put the quarterly review in the calendar, give each row its own date, and tell IT, finance and team leads which events should change a row between reviews. Link the inventory from your AI usage policy so staff know where approved tools are listed. This gives you a list that stays current.
The inventory is the first step in deploying AI on purpose rather than by accident. If you'd like help building it and turning it into decisions, our AI implementation team starts from a consultation.
What's still open
- Vendor defaults keep moving. Microsoft's December 9, 2026 change for EU and EFTA tenants and Anthropic's November 16, 2026 default are two of many. Expect more AI features to arrive in software you already pay for, often on by default.
- Agents are hard to see. Microsoft lists its agent directory as a preview. Most companies have no single place that shows every agent and what it can reach, so ask teams directly.
- State laws on AI in decisions. More states are writing rules on AI used in hiring and other decisions about people. The inventory's "decisions about people" column is how you'll know which rows a new law reaches.
We'll update this article when NIST revises its guidance or a law makes an inventory a legal duty for companies like yours. The date at the top of the page shows when we last checked.
Sources
- NIST AI RMF Playbook: Govern, National Institute of Standards and Technology, November 2026: the GOVERN 1.6 outcome, its suggested actions on ownership, scope and attributes, the review-interval question, and GOVERN 1.7 on decommissioned systems.
- Claude models from Anthropic in Microsoft Online Services, Microsoft Learn, November 2026: Claude models on by default for most commercial customers outside the EU, EFTA and UK.
- MC1546438: Claude models in Copilot are coming into the EU Data Boundary, Microsoft 365 Message Center (public mirror), November 2026: the December 9, 2026 default for EU and EFTA tenants without a selection.
- Set up and manage the new unified Claude experience for your Enterprise organization, Claude Help Center, November 2026: the November 16, 2026 default for existing Enterprise organizations.
- How do I discover AI apps and the sensitive data these use in my organization?, Microsoft Learn, April 2026: the Generative AI category in Defender for Cloud Apps and Entra Agent ID (preview).
- IBM Report: 13% of organizations reported breaches of AI models or applications, IBM, July 2025: the shadow AI breach share, the share with policies, the added cost and the study base.