What is an AI inventory?

An AI inventory is a maintained record of every AI system the organization builds, buys or embeds, with enough information about each to decide whether it should be running. It is the input to everything else in a governance program. Risk assessment needs to know what to assess. Monitoring needs to know what to watch. Regulatory reporting needs to know what exists. Without the inventory, each of those is a guess.

The unit that matters is the use case. A single foundation model behind an API can power a marketing copy tool, a support chatbot and a résumé screener. The model is the same; the risk tiers are low, limited and high. An inventory keyed on models hides that, which is why regulations and auditors ask about uses and affected people.

What fields to capture

GroupFields
IdentityName, unique ID, type (use case, model, agent, AI feature in a SaaS product), version, business owner, technical owner
PurposeIntended use, prohibited uses, decisions it informs or makes, affected people (employees, customers, public)
RiskRisk tier with written reasoning, EU AI Act classification, applicable laws and policies, last assessment date, open findings
TechnologyUnderlying models and providers, hosting, tools and APIs an agent can call, MCP servers, integrations
DataTraining and input data sources, personal or sensitive data categories, data residency, where outputs go
LifecycleStage (idea, pilot, production, retired), approval record, deployment date, monitoring in place, retirement plan
EvidenceModel or system card, evaluation reports, impact assessment, vendor attestations such as SOC 2 or ISO 42001, contracts

Resist the urge to capture everything on day one. A first pass with identity, purpose, owner and a provisional tier for every system beats a perfect record for ten of them. Add depth in proportion to the tier: a high-risk system needs every field and linked evidence; a spell-checker with an AI feature needs four lines.

For agents, the technology row is the one that matters most. An agent's impact is the union of what it can reach. Record the tools, APIs, MCP servers, data stores and permissions, and treat a change to any of them as a change to the system that reopens the risk assessment.

Six discovery sources

Self-reported intake is necessary and never sufficient. People do not register a SaaS feature that switched on in a release, a script that calls a model API, or an agent a team built in an afternoon. Combine intake with automated discovery:

SourceWhat it findsBlind spot
Intake and procurementPlanned use cases and new AI vendorsOnly what people choose to report
Cloud AI servicesModels and agents in Amazon Bedrock, Azure AI Foundry, Google Vertex AI, Databricks and similar, via their APIsMisses AI outside your cloud accounts
Code and repositoriesModel SDK imports, API keys for AI providers, agent frameworks, MCP server configsNeeds a scanner tuned for AI dependencies
Network, SSO and browser telemetryEmployees using third-party AI tools; the raw material for shadow AIShows usage, not purpose
SaaS vendor reviewsAI features switched on inside tools you already pay forDepends on vendors disclosing honestly
Identity systemsNon-human identities and agent credentials with access to business systemsRequires agents to be issued identities in the first place

Discovery produces signals, not inventory entries. A detected API key or a spike in traffic to an AI tool becomes an entry only when someone confirms the owner and purpose. Route discovery hits into the same intake workflow, with the discovered metadata prefilled, so the owner's job is to confirm and correct rather than start from a blank form. For the employee-usage side, see our shadow AI guide.

Keeping it current

Inventories decay. The fix is not a quarterly reminder email; it is to attach updates to events that already happen and that people cannot skip:

A 30-day build plan

  1. Days 1 to 5: agree the minimum fields and the tiering rule. Pick a home: a spreadsheet is fine under about 20 systems, a governance platform above that.
  2. Days 5 to 15: pull from the cheapest discovery sources first: cloud AI service APIs, SSO and network logs, and procurement records of AI-enabled vendors.
  3. Days 10 to 20: interview each business unit with the discovered list in hand. Confirm owners and purposes; add what they built that discovery missed.
  4. Days 20 to 30: tier everything, flag high-risk candidates for full assessment, and wire the first update trigger (procurement or deployment) so the list stops decaying.

Once the inventory exists, the rest of the program has something to act on: NIST AI RMF risk work, EU AI Act classification and the AI audit checklist. When manual tracking stops scaling, an AI governance platform takes over the inventory and adds discovery.

Related questions

Frequently asked questions

What is an AI inventory?

An AI inventory, also called an AI registry or AI catalog, is a maintained list of every AI system an organization builds, buys or embeds, with an owner, purpose, risk tier, data sources, lifecycle stage and supporting evidence for each. It is the foundation of AI governance: risk assessment, approvals, monitoring and regulatory reporting all start from it.

What should an AI inventory include?

At minimum: a unique ID, business and technical owner, intended and prohibited uses, the people affected, the risk tier with reasoning, underlying models and providers, data sources including personal data, lifecycle stage, approval record and links to evidence such as model or system cards and evaluation results. For agents, add the tools, APIs and permissions the agent can use.

Is an AI inventory required by law?

Not by that name in most places, but it is hard to comply without one. The EU AI Act requires providers and some deployers of high-risk systems to register them in an EU database and keep technical documentation, which presupposes knowing what you run. ISO/IEC 42001 expects documented resources and impact assessments per system, and US federal agencies must publish AI use case inventories.

What is the difference between an AI inventory and a model registry?

A model registry such as MLflow or Unity Catalog tracks model artifacts and versions for the teams that build them. An AI inventory tracks use cases and their risk, including models you never trained: AI features in SaaS tools, third-party APIs and agents. Link the two so each inventory entry points to the model versions it uses, but do not treat the registry as the inventory.

What is an AI bill of materials?

An AI bill of materials (AI-BOM or ML-BOM) lists the components inside an AI system: models, datasets, libraries and their versions and licenses, similar to a software bill of materials. The CycloneDX standard supports machine learning BOMs. An AI-BOM is a technical artifact attached to an inventory entry; the inventory adds ownership, purpose and risk.

Which tools build an AI inventory automatically?

Most AI governance platforms now include discovery. Examples include OneTrust AI agent detection, ServiceNow AI Control Tower discovery, Holistic AI Identify, Credo AI shadow AI discovery, Truyo scanning of websites and code, and Cranium detection across code and cloud. Security tools for shadow AI and cloud AI posture management also feed inventories. Our partner site We The Flywheel compares the platforms in its AI governance platforms guide.

Sources