AI inventory template: fields, example entries and how to keep it current

An AI inventory is a list of every AI system and agent your organisation uses, with a named owner and enough detail on each to say what it does, what data it uses and who it affects. Every AI framework assumes you know what AI you use. The NIST AI Risk Management Framework asks for mechanisms to inventory AI systems. ISO/IEC 42001 starts from the organisation’s context and the AI systems within it. The EU AI Act’s obligations attach system by system. Public sector policies, such as the Australian Government’s policy for the responsible use of AI, expect agencies to keep a register of their AI use cases with accountable owners. This page gives a set of fields you can adopt, three worked examples and the habits that keep an inventory current.

Scope: what goes in

Include anything that uses machine learning or generative AI to produce outputs, make or support decisions, or take actions. That covers more than people first expect.

• Systems built in-house, including models trained or fine-tuned on your data.
• Third-party AI products bought outright.
• AI features inside software you already use, once switched on.
• General-purpose assistants used for work, whether licensed centrally or not.
• Agents that call tools or act in other systems on your behalf.

Leave out deterministic automation that follows fixed rules, such as macros, scripts and rules engines, unless they feed a decision a law treats as automated. Write the scope rule down so entries are consistent.

The fields to record

Group the fields so each group answers one question. The groups below follow the shape most frameworks ask about.

Identity

• Name. The name people use for it.
• Kind. AI system or agent.
• Description. One or two sentences on what it does.
• Business unit. The part of the organisation that uses it.
• Internal reference. Your own identifier, used to join the inventory to other records.

Ownership

• Designated owner. The person accountable for the system day to day.
• Executive accountable. The senior person who answers for it.
• Technical contact. Who to call when it breaks.

Where it comes from (provenance)

• Origin. Developed in-house, third-party product, or a hybrid.
• Vendor. The company that supplies it, if any.
• Models. The model or models it runs on, with versions where known.
• Third-party data processing. Whether data leaves your environment to the vendor.

Data

• Processes personal information. Yes or no.
• Data categories. For example contact details, financial, health, biometric.
• Data assets used. Named datasets, with how each is used, such as for training, as reference material the system answers from, or as input it processes while running.
• Data retained. Whether inputs or outputs are retained, and by whom.

Decision and impact

• Outputs. Content, classification, recommendation or score, autonomous decision, task execution.
• Affects individuals. Whether its outputs bear on decisions about people.
• Impact categories. For example access to credit, employment, services or legal rights.
• Cross-border. Whether it affects people in, or data from, other jurisdictions.

Oversight and transparency

• Human oversight level. In the loop (a person approves each output), on the loop (a person monitors and can step in) or out of the loop (no routine human check).
• People interact directly. Whether individuals use or talk to it.
• People are told it is AI. Whether that is disclosed.
• Used as intended. Whether the use matches the vendor’s intended purpose.

Stage and risk

• Stage. Proposed, in development, testing, live, superseded or retired.
• Risk rating. From the most recent completed assessment, or “not assessed”. Never assume low.
• Last reviewed and next review due.

For agents

• Tools. The tools and systems it can call.
• Runtime identities. The identities it acts under.
• Action boundary. What it may do without a person, and what always stops for approval.
• Suspension. How it is stopped, and who can stop it.

Three example entries

These examples show the level of detail that makes an entry useful. The organisations and systems are illustrative.

Applicant screening assistant

• Kind: AI system. Business unit: People and Culture. Reference: HR-014.
• Owners: Head of Talent Acquisition (designated owner), Chief People Officer (executive), HR Systems Lead (technical).
• Origin: third-party product. Model: vendor’s proprietary ranking model. Data leaves environment: yes, to the vendor’s cloud.
• Personal information: yes. Categories: employment history, education, contact details. Data assets: applicant records (input).
• Outputs: recommendation or score. Affects individuals: yes. Impact: employment.
• Oversight: in the loop, a recruiter reviews every shortlist. Candidates told AI is used: yes, in the job advert.
• Stage: live. Risk rating: high. Next review: quarterly.

Customer service assistant

• Kind: AI system. Business unit: Customer Service. Reference: CS-003.
• Owners: Customer Service Manager, Chief Customer Officer, Digital Platforms Lead.
• Origin: hybrid. Model: a hosted large language model that answers from the help centre articles.
• Personal information: yes. Categories: contact details, account information. Data assets: help centre articles (reference material), chat transcripts (logging and retention).
• Outputs: content generation. Affects individuals: indirectly. People interact directly: yes. Told it is AI: yes.
• Oversight: on the loop, with escalation to a human agent on request or low confidence.
• Stage: live. Risk rating: moderate.

Accounts payable agent

• Kind: agent. Business unit: Finance. Reference: FIN-021.
• Owners: Accounts Payable Manager, Chief Financial Officer, Finance Systems Lead.
• Origin: developed in-house on a hosted model. Tools: read invoices, look up purchase orders, draft payment batches.
• Runtime identity: a dedicated service identity with read access to the ledger and draft-only access to payments.
• Action boundary: may match and code invoices on its own. Payment release always stops for a person.
• Personal information: limited, supplier contact details. Outputs: task execution.
• Oversight: in the loop for payments, on the loop for coding. Suspension: Finance Systems Lead can revoke the identity.
• Stage: testing. Risk rating: not assessed.

Seeding the inventory

Start from three sources at once. Procurement and expense records show what has been bought. A short survey of business unit leads, asking what AI their teams use and for what, shows intent. A discovery scan across device management tools, cloud accounts, code repositories and your configuration database shows what is actually running. Expect each source to find things the others miss. Reconcile them into one list, then go to each owner to fill the fields.

Keeping it current

An inventory decays from the day it is finished unless updates are tied to events.

• Intake. When a request for new AI is approved, the request becomes the entry. See the guide on AI use case intake.
• Change. When the model, data, purpose or oversight changes, the owner updates the entry and the change is versioned.
• Discovery. Regular scans surface new tools. Each finding is added, linked to an existing entry or dismissed with a reason.
• Scheduled review. As a backstop, review entries on a schedule set by risk.
• Retirement. Retire entries rather than delete them, so the history stays.

Version every change. A regulator or auditor asking about a decision made last March needs the system as it stood in March, not as it stands today.

Where Aicura fits

The field list above is close to the form Aicura’s AI Register uses, section for section, from identity and ownership through provenance, data profile, decision and impact, human oversight and transparency. At Pro, agents also get their own tabs for tools and runtime identities. Picking a product from the library of known AI products fills the entry for you to check and correct, so the vendor-side fields are not typed from memory.

The register, discovery, filling details from the product library, audit log and system and agent reports are all on the free Register tier. Your owners keep the entries true, and the version history shows what each one said on any past date.