An AI inventory is a structured, maintained record of the AI systems an organization develops, buys, or uses. It lists each system and records who owns it, what it is for, what data it touches, how risky it is, and where it stands in its lifecycle.

Two terms are used for it. “AI system inventory” is the NIST AI Risk Management Framework’s term. “AI use case inventory” is the US federal government’s term, and it tracks specific applications of AI rather than the technology only.

The US government’s guidance for 2025 agency AI reporting says federal agencies are required to conduct an “annual inventory of their AI use cases.”

 

Why it matters

Adoption has outrun visibility

Bloomberg Law, reporting on Stanford University’s Human-Centered AI Institute’s 2026 AI Index, said about 88% of organizations used AI for at least one business purpose in the prior year. That is 10 percentage points more than in 2024.

 

It is a foundation for risk management

NIST’s AI Risk Management Framework includes a governance subcategory, GOVERN 1.6, which says that “Mechanisms are in place to inventory AI systems” and that they should be resourced according to organizational risk priorities. In the framework’s design, the inventory is a precondition for the rest of AI risk management. Basically, you cannot assess, measure, or manage systems you have not identified.

The framework also adds actions under the same subcategory, these include enumerating generative AI systems for the inventory and defining any inventory exemptions in organizational policy for generative AI embedded in other products.

Regulators and standards bodies expect it

  • US federal agencies. Agencies must inventory their AI use cases, submit the inventory to the Office of Management and Budget (OMB), and publish the releasable portion. Agencies including the Department of Justice, the Federal Trade Commission, and the Federal Reserve Board now publish their own inventories.
  • United Kingdom. The UK government’s Algorithmic Transparency Recording Standard (ATRS) asks public bodies to publish structured records of their algorithmic tools. According to a UK Department for Science, Innovation and Technology (DSIT) press release, the ATRS became mandatory for all government departments.
  • European Union. The EU AI Act hinges on classifying systems by risk, which cannot be done without knowing what systems exist. According to law firm Hunton, the Commission published draft high-risk classification guidelines on May 19, 2026. Hunton also reports that requirements for stand-alone high-risk systems now apply from December 2, 2027, and requirements for high-risk systems embedded in products from August 2, 2028.
  • International standards. According to ISO, ISO/IEC 42001 is the first international standard for an AI management system. It is designed for any entity that develops, provides, or uses AI-based products or services.

 

What is in scope

A useful inventory usually covers:

  1. Systems you build. This includes machine learning models, scoring engines, and internal generative AI applications.
  2. Systems you buy. Vendor products with AI features, and third-party models you call through APIs.
  3. AI embedded in other tools. Assistants and automation inside productivity, HR, CRM, and security software.
  4. Systems in every lifecycle stage. The Department of Justice’s 2025 inventory says it includes use cases in all stages: pre-deployment, pilot, deployed, and retired.

On third-party systems, the Cloud Security Alliance’s AI Safety Initiative notes that organizations putting high-risk systems into use, even as off-the-shelf purchasers, carry their own obligations under Article 26 of the EU AI Act. Buying a system does not remove it from your inventory or your accountability.

 

What to record for each entry

  • Identity and ownership. Name, business owner, technical owner, and department.
  • Purpose. What decision or task the system supports, and who is affected by it.
  • Lifecycle stage. GAO’s review of federal inventories identified missing lifecycle-stage data as one type of gap. Common stages are planned, in development, pilot, deployed, and retired.
  • Risk classification. OMB guidance requires agencies to flag use cases that are “safety-impacting” or “rights-impacting,” as summarized by law firm Mintz. Under the EU AI Act, the equivalent is the high-risk classification.
  • Data and model details. Training data sources, personal data involved, model provider, and version.
  • Source and supply chain. Whether it is built in-house or bought, and from whom.
  • Controls and documentation. Testing, human oversight, monitoring, and links to impact assessments.
  • Decommissioning plan. NIST’s GOVERN 1.7 covers phasing out AI systems safely.

The UK’s ATRS shows how a standardized template works. The UK government’s guidance describes it as a template that organizations complete with key information about their algorithmic tools, which is then published on GOV.UK as a record. The OECD’s public-sector innovation observatory notes that the approach resembles datasheets and model logs. The UK’s earlier announcement of the standard explained why it focused on public bodies, observing that algorithmic tools support some of the “highest impact decisions affecting individuals.”

 

How to prepare an AI inventory

1. Set the definition and scope

Decide what counts as AI for your purposes, and write it down. Align it with the definitions you will be measured against, such as the EU AI Act’s definition, OMB’s, or your sector regulator’s. GAO’s 2023 review found that some federal inventories included technology that agencies later determined was not AI. FedScoop reported that two inventories included technology that was not AI and that several contained duplicates. A definition prevents both over- and under-counting.

 

2. Assign ownership and resources

Name an accountable lead and an executive sponsor. OMB’s April 2025 guidance, as described by ExecutiveGov, directs agencies to establish oversight for high-impact systems and to improve acquisition practices to avoid vendor lock-in.

 

3. Discover what exists

Combine several sources:

  • Surveys of business units and project teams.
  • Procurement, contract, and expense records.
  • IT and security records of approved software and SSO or identity logs.
  • Engineering repositories and cloud environments.
  • Interviews with teams you suspect are experimenting.
  •  

4. Consolidate where it makes sense

The OMB reporting guidance lets agencies report widely used AI use cases collectively, so long as they meet certain criteria. The Department of Justice did exactly this, combining similar widely adopted use cases into single department-wide entries.

 

5. Classify by risk

Assign each entry a risk tier using criteria drawn from your regulatory context. Under the EU AI Act, Arthur Cox, an Irish law firm analyzing the Commission’s draft guidelines, highlights that “Intended purpose is a decisive classification indicator.” That means your purpose field has to be written precisely, because it can determine the classification. The Cloud Security Alliance’s guidance adds that prudent practice is to treat potentially high-risk systems as high-risk until a formal determination is made.

 

6. Validate and quality-check

Review entries for completeness, duplicates, and accuracy against your template. As reported by Global Government Forum, only five of the 20 agencies that reported use cases gave information for each one, and the rest had “data gaps and inaccuracies.” GAO’s primary recommendation, in the same coverage, was that agencies update inventories to include all required information and align the data with instructions.

 

7. Decide what to publish and what to withhold

OMB’s guidance permits agencies to withhold use cases whose public release would be inconsistent with applicable law and policy. Private organizations may publish selectively for transparency, or keep the inventory internal and share it with regulators, auditors, and customers on request.

 

8. Maintain it on a schedule

Federal agencies update annually, but a living inventory should be updated often. Provide updates for events such as new vendor approvals, new model releases, project stage-gates, and decommissioning. The UK’s guidance, as reported by UKAuthority, recommends starting the transparency record during development rather than waiting for deployment, which helps keep entries current from the start.

Read also: Shadow AI is outpacing healthcare email security

 

FAQs

Who should own the AI inventory inside an organization?

An accountable lead supported by an executive sponsor should own it, with business and technical owners named for each individual entry.

 

How does an AI inventory differ from a general IT asset register?

An AI inventory goes beyond listing assets by recording each system’s purpose, affected people, data, risk tier, and lifecycle stage.

 

Should small organizations or startups keep an AI inventory?

Yes, because any entity that builds, buys, or uses AI needs to know what it has in order to assess and manage risk.

 

How can an organization find “shadow AI” that employees adopted without approval?

Cross-checking procurement and expense records, SSO and identity logs, and cloud environments, and interviewing teams suspected of experimenting, can surface unsanctioned tools.