The AI System Inventory: The First Artefact in Any AI Governance Programme
You cannot classify, assess or audit an AI system you have not listed. The AI system inventory is the first artefact of an AI governance programme, and NIST names it as an explicit expectation. Here is what an inventory is, what NIST says it should cover, what to record for each system, and where inventories fail.
Key Takeaways
- Every other governance step depends on the inventory. You cannot classify, risk-assess, monitor or audit an AI system that is not on a list, which is why the inventory comes first.
- NIST names it explicitly. GOVERN 1.6 in the NIST AI Risk Management Framework Playbook says mechanisms should be in place to inventory AI systems, resourced according to organisational risk priorities, and describes an inventory as an organised database of artefacts about each system.
- Complete beats perfect. NIST says inventories typically capture all organisational systems because partial inventories may not provide the value of a full one. Where that is not possible, it suggests at minimum covering high-risk or high-stakes systems.
- Embedded and employee-used AI count. NIST's generative AI profile asks organisations to enumerate generative AI systems and to define any inventory exemptions for generative AI embedded in application software. An inventory that lists only systems the company built is incomplete.
- An inventory needs a named owner and a review rhythm. NIST suggests a policy naming the person or team responsible for maintaining it. A one-off survey that nobody maintains goes stale within months.
An AI governance programme has many parts: classification, risk assessment, documentation, monitoring, audit, incident response. They all share one prerequisite that is rarely the first thing a leadership team asks for: a list of the AI systems the organisation actually uses. Without it, every later step is built on guesswork about what is in scope.
This piece sets out what an AI system inventory is, what NIST says it should contain, what to record for each system, and where inventories most often fail. It is advisory. It describes the artefact and the decisions around it; it does not describe any tooling.
What an AI System Inventory Is
NIST's AI Risk Management Framework Playbook defines an AI system inventory as an organised database of artefacts relating to an AI system or model. The Playbook says it may include system documentation, incident response plans, data dictionaries, links to implementation software or source code, and names and contact information for relevant AI actors (NIST AI RMF Playbook, GOVERN 1.6, accessed 16 November 2026).
The same Playbook entry says a serviceable inventory lets an organisation quickly answer two kinds of question:
- Questions about a single system, such as "when was this model last refreshed?"
- Questions across the whole estate, such as "how many models are currently deployed within our organisation?" or "how many users are impacted by our models?"
That is a useful test for your own inventory. If a board member asked the second kind of question tomorrow, could you answer it from a document, or would it take a week of emails?
What NIST Says About It
The inventory sits under the GOVERN function of the NIST AI Risk Management Framework (AI RMF 1.0, January 2023). Subcategory GOVERN 1.6 reads: "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities." The Playbook entry for GOVERN 1.6 makes four suggestions that are worth turning into policy decisions:
- 1Have a policy that defines how inventories are created and maintained.
- 2Name a specific individual or team responsible for maintaining the inventory.
- 3Decide which systems are inventoried. The Playbook's preference is all models or systems or, at minimum, high-risk models or systems, or systems deployed in high-stakes settings. It adds that inventories typically capture all organisational models or systems, because partial inventories may not provide the value of a full inventory.
- 4Decide which attributes are recorded, for example documentation, links to source code, incident response plans, data dictionaries and contact information for AI actors.
Two caveats on currency. NIST's page for the Playbook states that AI RMF 1.0 is being updated and that the Playbook will be updated after the framework is revised, so check the current text when you design your policy. And the Playbook notes that inventories are a common element of traditional model risk management, so if your organisation already runs a model inventory for another purpose, start there rather than from nothing.
What the Generative AI Profile Adds
NIST's Generative AI Profile (NIST AI 600-1, July 2024) maps suggested actions to GOVERN 1.6 that matter in a world of copilots, chat tools and AI features built into other software:
- Enumerate the organisation's generative AI systems for inclusion in the inventory, and adjust inventory requirements to account for generative AI risks.
- Define any inventory exemptions in organisational policy for generative AI systems embedded into application software. This is a decision, not a default: an exemption you have not written down is a gap.
- Consider additional entries, including data provenance information; known issues reported from internal tracking or external sources; human oversight roles and responsibilities; special rights and considerations for intellectual property, licensed works, or personal, privileged, proprietary or sensitive data; and the underlying foundation models, their versions and access modes.
The last bullet is the practical one for most enterprises. A vendor's product that "uses AI" often sits on top of a model the vendor does not control and may change. If your inventory records only the product name, you will not know when the model underneath changes.
What to Record for Each System
The fields below are our suggested starting point, drawn from the NIST items above and common governance practice. They are a recommendation, not a NIST template. Add or remove fields to suit your organisation, but decide deliberately.
- System name and unique ID — So everything else can refer to one record
- Business purpose and business owner — Who is accountable for it, and what it is for
- Type — Built in-house, bought, embedded in other software, or a tool staff use on their own initiative
- Vendor, underlying model and version — NIST's profile asks for foundation models, versions and access modes
- Data used — Including any personal, proprietary or sensitive data, and data provenance
- Who uses it, and who is affected — NIST gives "how many users are impacted" as an example question
- Human oversight role — Who reviews or can override it
- Risk classification and date assessed — Links the inventory to your classification process
- Links to documentation and incident response plan — NIST lists both as inventory artefacts
- Status — Proposed, live or retired. NIST notes decommissioned systems may be kept in the inventory for an established length of time
- Last reviewed, and by whom — Shows the record is maintained, not merely created
If you use model cards for documentation, the inventory is where each card is linked. Our guide to AI model cards covers what a card contains.
Finding Systems You Do Not Know About
The hardest part of an inventory is not the template. It is discovering what belongs on it. Four sources are worth checking, and none of them requires any tool beyond asking the right people:
- Each function's own account. Ask every business unit what AI it uses or has bought, including features inside tools it already pays for.
- Procurement and finance records. Spend on AI products, and on existing products that added AI features, shows up here.
- IT and security knowledge of approved and blocked tools.
- Staff themselves. Employees using AI tools on their own initiative will not appear in any procurement record. Our guide to shadow AI governance covers why, and how to ask without driving usage underground.
Treat the first inventory as a baseline, not a finished product. The point is a documented starting list, with an owner and a review date.
Where Inventories Fail
- Only the systems IT built. Embedded features and staff-chosen tools are often where the risk is, and often missing.
- A one-off survey. Without a named owner and a review rhythm, the list goes stale. NIST's Playbook asks what an appropriate interval is for checking that a system is still accurate, unbiased and explainable; the same question applies to the inventory record itself.
- Fields nobody can answer. If "data used" is blank for most entries, the inventory looks complete and is not.
- No link to decisions. An inventory that does not feed classification, risk assessment and monitoring is a spreadsheet, not a governance artefact.
- Exemptions nobody wrote down. If some systems are deliberately left off, the policy should say which, and why.
What Leaders Should Be Asking
- Do we have a single list of every AI system in use, including those embedded in software we buy?
- Who owns it, by name, and when was it last reviewed?
- Which systems are deliberately excluded, and who approved the exclusion?
- Can we answer "how many AI systems are live, and who do they affect" from the list, without a new survey?
- Does each record link to its documentation, its risk classification and its incident response plan?
The Underlying Point
The inventory is the least glamorous governance artefact and the one everything else rests on. It makes classification possible, gives a risk assessment a population to assess, tells monitoring what to watch, and gives an auditor somewhere to start. Organisations that try to write policies, classify risk or prepare for audit without it usually discover the gap at the worst moment: when someone asks what AI they actually run, and nobody can say.
Once the list exists, the next step is to record what could go wrong with each system on it, who owns each risk and what has been decided. We cover that in How to Build an AI Risk Register: What to Record and Who Owns It.
Imagine Works advises enterprise leaders on AI governance and risk design, including how to scope an AI system inventory, decide what belongs in it and connect it to classification and risk assessment. We advise; we do not build or implement software. Get in touch to discuss your AI governance programme.
Sources (as of 16 November 2026)
All opened and read on 16 November 2026.
- 1NIST, AI Risk Management Framework Playbook, GOVERN 1.6 and GOVERN 1.7 (page states AI RMF 1.0 is being updated and the Playbook will follow)
- 2NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023
- 3NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, July 2024 (actions GV-1.6-001, GV-1.6-002, GV-1.6-003)
Related Service
AI Governance & Risk Design
Designing the governance framework and risk architecture that keeps your AI systems compliant, auditable, and board-ready — before regulation forces the issue.
Explore this serviceMore Insights
More on AI Governance
Shadow AI: Why One-in-Five Enterprises Now Has a Governance Problem They Cannot See
Shadow AI — employees using unsanctioned generative tools at work — has moved from anecdote to material risk. IBM's 2025 breach data put a number on it: organisations with high shadow AI usage paid $670,000 more per breach, and only 37% of organisations have any policy to detect it. Here's what enterprise leaders should do about it.
How to Build an AI Risk Register: What to Record and Who Owns It
An AI risk register is the working record of what could go wrong with each AI system, how serious it is, who owns it and what is being done. NIST's AI Risk Management Framework does not use the term, but it asks organisations to document exactly the things a register holds. Here are the fields, the owners, the review rhythm and the common failures.
Generative AI Acceptable-Use Policy: What to Include
A generative AI acceptable-use policy is the control that answers shadow AI: it tells staff which tools they may use, what data may go in, how outputs must be checked and what is off limits. Here is a section-by-section checklist of what to include, anchored in NIST's generative AI profile and OWASP's guidance on sensitive information disclosure.