An AI model inventory is the single, maintained register of every AI model an organization builds, buys, or relies on, recording each model’s owner, purpose, risk tier, and lifecycle status. It gives governance teams one authoritative list to answer basic questions: what models do we run, who answers for them, how risky are they, and are they still approved for use. Without it, AI governance rests on assumptions.
Table of contents
- Introduction
- What Is an AI Model Inventory and Why Do You Need One?
- AI Model Inventory vs Model Cards: Different Artifacts, Different Jobs
- Which Fields Belong in an AI Model Inventory?
- How Do You Discover Models Already in Use?
- Assigning Owners, Risk Tiers, and Lifecycle Status
- Building an AI Model Register Step by Step
- How Does an AI Model Inventory Support EU AI Act Compliance?
- Keeping the AI Model Inventory Current After Launch
- The Bottom Line on AI Model Inventories
- Frequently Asked Questions About AI Model Inventory
Introduction
Most governance teams I have worked with discovered their AI footprint the hard way. A vendor feature quietly switched on a scoring model, a analytics group fine-tuned an open-weight model on a laptop-sized cluster, and a business unit wired a large language model into a customer workflow with no review. None of these were malicious. They were simply invisible, because nobody had a current list of what existed.
Having spent years building governance programs in regulated environments, I have come to treat the AI model inventory as the foundation that everything else depends on. In financial services environments like Wells Fargo, model risk management has long required institutions to know every model in use, its purpose, and who is accountable for it. That discipline predates the current wave of generative AI, and it translates directly to AI governance. The inventory is the piece that makes policy enforceable rather than aspirational.
This article is a practical guide to building and maintaining one. I will explain how an AI model inventory differs from a model card, which fields actually earn their place, how to find models that are already in use without anyone declaring them, and how to assign owners, risk tiers, and lifecycle status in a way that holds up under audit. I will walk through a step-by-step build, explain how the inventory supports obligations under the EU AI Act, and close with how to keep the register accurate after launch.
If you are new to the broader discipline, it helps to start with what is AI governance, because the inventory is one of the operational controls that governance policy depends on. The inventory also sits on top of the data foundations discussed in Data Lineage - Leveraging the Power of Insight to, since you cannot assess a model’s risk without knowing what data trained it and where that data came from. Throughout, I will use “model” broadly to include machine learning models, statistical models used for decisions, large language model deployments, and third-party AI features embedded in software. Where the distinction matters, I will call it out.
What Is an AI Model Inventory and Why Do You Need One?
An AI model inventory is a controlled, versioned list of AI models and the metadata needed to govern them. “Controlled” means someone is accountable for the list and changes to it are traceable. “Versioned” means you can see what the register looked like on a given date, which matters during audits and incident reviews. “Metadata” is the set of attributes that turn a list of names into something you can act on, such as owner, use case, data sources, risk tier, and status.
The concept is not new. Model risk management guidance for banks has required inventories for years. According to the Federal Reserve’s supervisory letter SR 11-7 on model risk management (2011), banks are expected to maintain a comprehensive set of information on models in use, which is the foundation for validation, ongoing monitoring, and governance. Financial regulators in other jurisdictions have reached similar expectations. What has changed is the scope. Traditional model inventories focused on pricing, credit, and capital models built by quantitative teams. Today’s inventory must also capture vendor models, generative AI deployments, embedded AI features in SaaS products, and internally built automation that makes or supports decisions about people.
The Three Problems an Inventory Solves
In my experience, the inventory solves three distinct problems, and it helps to name them because teams often build a register to solve only one.
The first is visibility. You cannot govern what you cannot see. An inventory forces the question of what is running, and it surfaces models that procurement, IT, and business units did not think of as “AI” at all. A fraud rule engine that uses a gradient-boosted classifier is a model. A marketing propensity score is a model. A chatbot that summarizes support tickets is a model deployment.
The second is accountability. Every model needs a named owner who can answer for its behavior, approve changes, and decide when it should be retired. Without that, incidents bounce between data science, IT, legal, and the business. The inventory makes accountability a field that must be filled in, not a hope.
The third is prioritization. Governance resources are finite. An inventory with risk tiers tells you where to spend review effort. A low-stakes internal summarization tool does not need the same validation depth as a model that determines loan eligibility or medical triage. Without tiers, every model gets the same process, which means the important ones get too little attention and the trivial ones create bottlenecks.
Why the Inventory Comes Before Everything Else
Many organizations start with an AI ethics principles document or a policy, then wonder why it does not change behavior. The inventory is the bridge between principles and practice. A policy that says “high-risk models require independent validation” is meaningless until you can tell which models are high-risk and whether each has been validated. The inventory gives every other control a target.
It also gives leadership a defensible answer when someone asks how many AI systems the company operates. I have seen executives give a number in a board meeting that turned out to be off by a factor of several once discovery work began. An inventory that is maintained and reconciled against procurement and engineering records protects the credibility of that number.
AI Model Inventory vs Model Cards: Different Artifacts, Different Jobs
Practitioners frequently confuse the AI model inventory with a model card, and the confusion causes real gaps. They are complementary artifacts that answer different questions, and you need both.
A model card, a concept introduced by Mitchell and colleagues in the paper “Model Cards for Model Reporting” (2019), is a document describing a single model. It typically covers intended use, training data characteristics, evaluation metrics across relevant groups, limitations, and known failure modes. The audience is anyone who needs to understand a specific model in depth, including developers, reviewers, and downstream users. A model card is a deep, per-model artifact.
An AI model inventory is a portfolio-level register. It describes the entire population of models at once, and its job is to answer enterprise questions. Which models exist? Who owns them? Which are high-risk? Which are approved for production? Which are past end of life but still running? The inventory points to model cards and other documentation; it does not replace them.
How the Two Work Together
The clearest way I explain it to governance teams is this: the inventory is the index and the model card is the entry. Each inventory record should carry a link to the model’s card or equivalent documentation, and the inventory should flag when a model lacks one. A model with no card, no owner, and no risk tier is a governance gap regardless of how good the model is technically.
Consider a credit underwriting model. Its model card might describe the features, the training population, the performance across demographic segments, and the conditions under which it should not be used. The inventory record for the same model says the model is in production, owned by the consumer lending risk function, classified as high-risk under the institution’s tiering scheme, last validated on a given date, and scheduled for review. A risk committee reviewing the portfolio reads the inventory. A validator digging into the model reads the card and the underlying documentation.
Where Teams Get It Wrong
The most common mistake is treating the model card as the inventory, which leaves no way to see the portfolio. A folder of forty model cards is useful, but it does not tell you which models are missing cards. The second mistake runs the other way: a spreadsheet inventory with dozens of columns that duplicates the content of model cards. That creates two sources of truth that drift apart. The discipline I recommend is to keep substantive technical content in the model documentation and keep the inventory to governance attributes plus links.
The distinction also matters for vendor models. For a third-party model you cannot fully document, the inventory record may be the primary governance artifact you control, supplemented by whatever documentation the vendor provides. Your inventory should record what you know and what you have requested that you do not yet have.
Which Fields Belong in an AI Model Inventory?
The field list determines whether an inventory is useful or becomes shelfware. The temptation is to capture everything. In practice, a register with sixty mandatory fields gets filled with placeholder text within a quarter. I recommend a core set of mandatory fields that every record must carry, plus optional or conditional fields that apply to higher-risk models.
Core Identification Fields
Every record needs a unique identifier that does not change when the model is retrained or renamed. I use a stable identifier assigned at registration, separate from the model name and version. Models get renamed, and marketing names for vendor features change frequently. The record also needs a human-readable name, a short description of what the model does in business terms, and the business unit or product it supports.
Version information belongs here too. Record the current deployed version, the version identifier used by your model registry if you have one, and the date of the last material change. Be explicit about what counts as a material change, because retraining on new data may or may not require re-review depending on your policy.
Ownership and Accountability Fields
Three ownership fields are essential. The accountable owner is the individual, not a team, who answers for the model’s business outcomes. The technical owner maintains the model and its infrastructure. The validator or reviewer is the independent party who assesses it, where independence is required. Keeping these separate avoids the common situation where the person who built the model also signs off on its risk assessment.
I also recommend a field for the executive sponsor when the model supports a material business process. This is not bureaucracy for its own sake. When a model misbehaves in a way that reaches regulators or customers, someone senior needs to be on the hook.
Purpose and Use Case Fields
Record the intended use in plain language, along with explicit out-of-scope uses. The out-of-scope field is underused and extremely valuable. A sentiment model built for product reviews should not be repurposed to rank job applicants, and writing that down prevents scope creep. Include the decision type the model influences, such as recommendation, classification, scoring, generation, or automation, and whether a human reviews the output before action.
Data fields belong here as well: the categories of data the model consumes, whether it processes personal data, and the data sources. This is where the inventory connects to data governance. If you cannot say where a model’s training data came from, you have a lineage gap that will surface during any serious review.
Risk, Compliance, and Lifecycle Fields
Risk tier, lifecycle status, and the date of the next scheduled review belong in the core set. I cover tiers and status in depth in a later section. For regulated use cases, add fields for the applicable regulatory categories. Under the EU AI Act, that means whether the system is likely to fall into a prohibited category, a high-risk category, a transparency-obligation category, or none. The classification should be recorded with the rationale, not just the outcome.
Also capture the locations where the model runs, including cloud region and whether it is hosted by a third party. For EU operations, data residency and the legal entity deploying the system can determine which obligations apply. Record the vendor and contract reference for any third-party model, and the date the contract or license renews, because that often triggers a re-review opportunity.
Documentation and Evidence Fields
Each record should link to the model card or equivalent, the most recent validation or assessment report, the monitoring dashboard, and the approval record. A useful test: if an auditor selects a model at random, can you produce each linked artifact within a day? If not, your evidence fields are either missing or pointing at documents that do not exist.
Fields I Would Leave Out Early
Early in a program I recommend deferring fields that are expensive to populate and rarely drive decisions, such as detailed hardware specifications, exhaustive feature lists, and full performance metrics in the inventory itself. Keep those in the model documentation. Start with the fields that drive action: owner, purpose, tier, status, and evidence links. You can add depth once the habit of maintenance is established.
If you are building the governance structure around these fields, the data governance RACI matrix template is a practical way to map who is responsible, accountable, consulted, and informed for each inventory field. It keeps the owner fields honest.
How Do You Discover Models Already in Use?
Discovery is the hardest part of the first inventory and the part most teams underestimate. Models that someone deliberately registered are the easy ones. The difficult ones are the models nobody thinks of as models: the forecasting spreadsheet with a regression embedded in a macro, the vendor’s “smart” routing feature, the chatbot a team stood up on a personal account over a weekend. The phrase practitioners use for the last category is shadow AI, and the discovery process is where you find it.
Start with Authoritative Sources
Begin with records you already have. Procurement and accounts payable hold software contracts and subscriptions, and filtering for AI-related vendors is usually productive. Cloud provider billing exports show which AI services are consuming spend and in which accounts. Engineering teams maintain service catalogs and repositories, and searching code for common machine learning and large language model library imports reveals a surprising amount. Model registries, feature stores, and experiment tracking tools, where they exist, are primary sources for internally built models.
Do not assume these sources are complete. Each one covers a slice. The reconciliation between them is where the real picture emerges, because a model that appears in cloud billing but not in the engineering catalog is a strong signal of an unmanaged deployment.
Survey the Business, Carefully
A structured survey of business units and functional leaders surfaces use cases that never touch IT systems, such as analysts using a general-purpose chat tool to draft client communications or process documents with AI features. Design the survey to ask about outcomes and tasks rather than about “AI,” because many people do not identify their tools as AI. Ask what decisions they make with software assistance, what they automate, and what they have tried in the last six months.
The survey must be framed as an inventory exercise, not an enforcement exercise. In my experience, if people fear that disclosure will get their tool shut down, they under-report. I have found it works better to publish a clear amnesty period in which teams can register existing tools without consequence, and to make the process of registering fast. Teams that feel safe disclosing will tell you about the models you most need to know about.
Address Shadow AI Directly
Shadow AI inventory work needs its own approach. Browser-based tools, personal accounts, and browser extensions do not show up in procurement. Network and endpoint telemetry can identify traffic to common AI service domains, subject to your privacy obligations and employee monitoring policies, which vary by jurisdiction and should be cleared with legal and works councils where applicable. The goal is not surveillance of individuals; it is identifying which services are in use so you can assess them.
Once you have identified shadow usage, triage it. Some tools are low-risk productivity aids that can be approved with simple conditions, such as prohibiting input of personal or confidential data. Others are used in ways that touch regulated decisions and require full registration and assessment. A short acceptable-use standard, with examples, helps people self-classify before they ask.
Embedded and Vendor AI
The fastest-growing category, in my view, is AI embedded in software you already license. Enterprise applications increasingly ship with recommendation, scoring, summarization, and generation features that are switched on by default or by a configuration change. Review your major SaaS contracts and release notes for AI features, and ask each vendor directly what models power those features, where data is processed, and whether it is used for training. Where vendors cannot answer, record the gap in the inventory.
A practical rule: if a feature influences a decision, generates content that leaves the organization, or processes personal data, it belongs in the inventory regardless of whether anyone calls it AI.
Keep Discovery Continuous
Discovery is not a one-time project. Build triggers into procurement, change management, and architecture review so that new AI-related purchases or deployments prompt registration before go-live. Schedule periodic re-scans of the sources above. A discovery exercise that completes in a quarter and then stops will be out of date within a year, because the rate at which AI features appear in ordinary software is high.
Assigning Owners, Risk Tiers, and Lifecycle Status
These three attributes turn the inventory from a list into a control. Owners establish who acts, risk tiers establish how much scrutiny applies, and lifecycle status establishes what the model is currently allowed to do. Getting them wrong produces an inventory that looks complete and governs nothing.
Assigning Accountable Owners
The accountable owner should be a named individual with authority over the business outcome the model supports. This is usually a business leader, not a data scientist, because the business leader decides whether the model’s output is acceptable in practice. The technical owner, typically a data scientist or machine learning engineer, maintains the model. Confusing the two is a frequent source of failure. When the technical owner is also the accountable owner, nobody with business authority is looking at the model’s effects.
Ownership must survive personnel change. I recommend requiring an owner review whenever someone leaves the role, and treating an unowned model as a governance incident that blocks further changes until an owner is assigned. Orphaned models are where the worst surprises live.
Designing Risk Tiers
Risk tiers group models so that controls scale with potential harm. Most organizations use three to five tiers. The tier should reflect the consequences of error or misuse, the population affected, the degree of automation, and the regulatory exposure, not the sophistication of the algorithm. A simple logistic regression that determines credit limits can be higher risk than a sophisticated recommender that suggests articles.
A workable tiering approach uses a short set of questions that everyone can answer consistently. Does the model influence a decision about an individual’s access to a service, employment, credit, insurance, or safety? Does it operate without meaningful human review? Does it process sensitive or regulated personal data? Could a failure cause financial loss, legal exposure, or physical harm, and at what scale? A “yes” to the first question or two should generally push a model into the top tier, regardless of its technical complexity.
Map each tier to specific control requirements: validation depth, approval authority, monitoring frequency, documentation standard, and review cadence. The tier only has value if it changes what happens. The data classification policy template can help align tier definitions with the data sensitivity categories you already use, which keeps your governance vocabulary consistent.
Writing Tier Definitions That People Can Apply
The most common failure I see is tier definitions that are technically precise but unusable. “Models with significant societal impact” tells a product manager nothing. Replace abstractions with examples drawn from your own portfolio. Show the top tier with two or three real use cases, the middle tier with examples, and the bottom tier with examples. When a new model arrives, the team can compare it to examples rather than debating principles.
Document the tiering decision and who approved it. When a model is reclassified, record the reason. Reclassification is a normal part of the lifecycle, and the audit trail of decisions is often more valuable than the current tier itself.
Lifecycle Status
Lifecycle status tells everyone what state a model is in. I recommend a small, fixed vocabulary. A common version includes proposed (under evaluation, not yet built or deployed), in development, in validation, approved for production, in production with restrictions (approved with conditions that must be met), under review (issue identified, use may be suspended), retired, and decommissioned (removed from infrastructure and data stores). Keep the list short enough that people use it correctly.
Status should drive behavior. A model marked “under review” should trigger an alert to the accountable owner and, for top-tier models, a notification to the governance committee. A model in production without a current validation should be flagged automatically. Transitions between states should require a named approver, and the approval should be captured in the record.
Build in expiry. Approvals should not be indefinite. An approval that lapses after a defined period, or when a material change occurs, forces re-review and prevents the common situation where a model approved three years ago is still running on data it was never designed for. I describe how to operationalize this in the section on keeping the inventory current.
A Note on Thresholds and Judgment
No tiering scheme eliminates judgment. Edge cases will arise, such as a model that is low-risk in one market and high-risk in another because of different regulations or customer populations. Give the governance committee a defined route for escalation and override, and require that overrides be recorded with rationale. The aim is consistency and transparency, not mechanical perfection.
Building an AI Model Register Step by Step
The build itself is a sequence of decisions that is easy to get out of order. Teams that rush to a tool before defining fields, or that try to inventory everything before setting tiers, usually rebuild the register within a year. The sequence below is the one I return to, adjusted to the size of the organization.
Step 1: Define Scope and Governance Ownership
Start by writing down what the inventory covers. Decide whether it includes only machine learning models, or also statistical models, rules-based decision engines, generative AI deployments, and embedded vendor features. Be explicit about the boundary, because ambiguity at this stage creates disputes later. Name the owner of the inventory itself, typically a function such as model risk, data governance, or an AI governance office, and define who can approve changes to the register.
Agree on the purpose statement. A sentence such as “The AI model inventory records every AI system used by or on behalf of the organization, its accountable owner, risk tier, and approval status, to support governance, audit, and regulatory obligations” gives every later decision a reference point.
Step 2: Set the Core Field Standard
Using the field discussion above, write a data standard that defines each mandatory field, its allowed values, and its format. Specify which fields are required at registration and which can be completed later, with a deadline. The standard should be short enough to read in one sitting. Publish it alongside examples of a completed record for a low-risk and a high-risk model, because examples do more to standardize quality than definitions do.
Step 3: Define Tiers and Lifecycle Vocabulary
Finalize the tier scheme and lifecycle states before you populate records. Pilot both on a small set of known models and revise definitions based on where people disagreed. This pilot is one of the most valuable steps in the build. Disagreement among reviewers during the pilot tells you exactly where the definitions are ambiguous.
Step 4: Choose the System of Record
You can start in a controlled spreadsheet or a governance platform, and many organizations do. What matters more is that the system supports unique identifiers, version history, role-based edit rights, approval workflows, and reporting. A spreadsheet can meet these needs for a small portfolio if access is controlled and changes are logged. As the portfolio grows, a dedicated governance tool or a model registry integrated with your development platform becomes necessary. Choose based on the integrations you need, not on feature lists. The platform selector on this site can help you narrow the tool category before you commit to a vendor, and it is worth doing that before any procurement begins.
I would caution against selecting a tool and then designing the inventory around its defaults. Design the inventory first, then confirm the tool can support it. Vendor defaults often omit the approval workflows and evidence linking that governance requires.
Step 5: Populate Tier-One Models First
Begin population with the highest-risk models. These are the ones that matter most if something goes wrong, and building their records forces you to resolve the hard questions about ownership, evidence, and data lineage. Completing them also creates the reference examples that make later records faster to write. Expect this stage to take longer than planned. Each high-risk record may surface a model nobody had documented properly.
Step 6: Run Discovery Against the Register
Once the first records exist, run the discovery activities described earlier and reconcile what you find against what is registered. Each discrepancy becomes either a new record or a documented decision that a system is out of scope. Track the reconciliation rate as a metric. A rising number of models found through discovery and not through self-registration suggests the intake process is not working.
Step 7: Connect Intake and Change Control
Make registration a precondition for a few specific events: procurement of AI-related software, promotion of a model to production, material change to a registered model, and retirement. Embed these checkpoints in existing processes such as architecture review, change advisory boards, and vendor onboarding rather than creating a parallel bureaucracy. In my experience, controls embedded in existing workflows get followed, while standalone registration requests get skipped under deadline pressure.
Step 8: Establish Reporting and Review Cadence
Define the reports the inventory must produce and who receives them. A typical set includes the portfolio count by tier and status, models past their review date, models without a named owner or current validation, and newly discovered models pending registration. Review the top tiers with the governance committee at a regular cadence, and review the full portfolio less frequently. Reporting is where the inventory earns its keep, because it turns records into decisions.
Step 9: Test with an Audit Sample
Before declaring the register operational, select a sample of models and test whether each record is accurate and whether the linked evidence exists and supports the recorded status. This dry run reveals gaps that look fine on paper. Fix the gaps, then repeat the test on a different sample. I have found that the first audit sample typically uncovers at least one model where the recorded owner has left the organization, and that discovery alone justifies the exercise.
How Does an AI Model Inventory Support EU AI Act Compliance?
The inventory is not itself a compliance obligation under the EU AI Act in the sense of a single named register, but it is the practical mechanism through which most organizations will meet several of the Act’s requirements. Anyone building an inventory for European operations should understand the connection, and anyone reading this should verify the current status of the Act’s timelines against official sources, because the application dates have been subject to proposed amendment. [VERIFY: current application dates for high-risk AI system obligations under the EU AI Act, including any amendments adopted through the EU’s 2025–2026 simplification proposals.]
For a broader treatment of the Act’s data requirements, see the EU AI Act data governance requirements guide, which covers the data-specific obligations in more depth than I will here.
Classification Comes First
The Act sorts AI systems by risk. Certain practices are prohibited outright. High-risk systems, defined in the Act’s Annex III categories and in certain product-safety contexts, face extensive requirements before and after deployment. Systems with specific transparency obligations, such as those that interact directly with people or generate synthetic content, face disclosure duties. Most other systems face few specific obligations beyond general requirements such as AI literacy among staff.
An inventory with a documented classification rationale for each system is the evidence base for these determinations. When a regulator or auditor asks why a system was not treated as high-risk, the record should show the analysis, the date, and the reviewer. Without that, you are reconstructing your reasoning after the fact.
Obligations the Inventory Makes Manageable
For high-risk systems, the Act requires among other things risk management, data governance, technical documentation, record-keeping, transparency to deployers, human oversight, and accuracy and robustness requirements. It also requires registration of certain high-risk systems in an EU database before they are placed on the market or put into service, subject to specified exceptions. Providers and deployers have distinct obligations, and many organizations are both, depending on whether they build or only use a system.
Each of these obligations needs to be traceable to a specific system. The inventory provides that traceability. The owner field identifies who is accountable. The classification field records the regulatory category. The evidence links point to technical documentation, risk management records, and monitoring output. The lifecycle status shows whether the system is in a state where the obligations apply. A quality inventory is effectively a compliance index for the Act.
Distinguishing Provider and Deployer Roles
The Act distinguishes between providers, who develop a system or place it on the market under their name, and deployers, who use a system under their authority. The same organization can hold both roles for different systems. Your inventory should record the role for each system, because the obligations differ. A deployer using a third-party high-risk system must meet deployer obligations, such as using the system according to instructions, ensuring human oversight, monitoring operation, and keeping logs. A provider has broader obligations including conformity assessment and documentation.
For third-party systems, the inventory should record what the vendor has provided about its compliance status and what you have requested. If a vendor cannot confirm its classification or provide technical documentation, that gap is material and should be visible in the register rather than buried in a contract file.
Logs, Retention, and Monitoring
The Act includes requirements for automatic logging of high-risk systems and for retention of certain logs and documentation for defined periods, and it requires post-market monitoring by providers. [VERIFY: exact retention periods and log obligations for deployers under the current text of the EU AI Act.] Your inventory should record where logs are held, who can access them, the retention period applied, and the date of the last monitoring review. These fields are easy to skip and expensive to reconstruct, so build them in from the start.
General-Purpose AI Models
Obligations for providers of general-purpose AI models were set to apply from August 2025, separate from the high-risk regime. Organizations that build or fine-tune general-purpose models, or that place them on the market, should record these models in the inventory with their provider status and any documentation the provider has published. [VERIFY: current status of the general-purpose AI code of practice and provider documentation expectations as of 2026.] For organizations that only use a general-purpose model through an API, the inventory should record the upstream model provider and the deployed application as separate entries, because the governance questions differ.
Practical Steps for EU Readiness
Start by identifying every system with an EU nexus, meaning placed on the EU market, used in the EU, or producing outputs used in the EU. Classify each one and record the rationale. Prioritize the systems that appear to be high-risk and confirm their roles, owners, and evidence. Then map the obligations to the fields and linked artifacts the inventory already holds, and identify gaps. This approach converts a legal analysis into an operational work plan, which is what most governance teams need.
Do not treat the regulation as a one-time exercise. Classification can change when a system’s use changes, and the inventory’s change-control checkpoints are where those changes should trigger reassessment.
Keeping the AI Model Inventory Current After Launch
An inventory that is accurate on launch day and stale six months later is worse than none, because it creates false confidence. Keeping the register current is an ongoing operational discipline, and it is where most programs struggle. The mechanisms below are the ones I have seen work.
Build Review Cadences by Tier
Assign each tier a review frequency. Top-tier models typically warrant a formal review at least annually and whenever a material change occurs, with monitoring output reviewed more often. Lower tiers can use longer cycles. The exact frequencies should come from your own risk policy and regulatory expectations, not from a generic benchmark, so set them deliberately and record the rationale. The review should confirm that the owner is still correct, the purpose is still accurate, the risk tier still applies, the evidence is current, and the status is right.
Make the review a checklist rather than a conversation. A short set of questions answered in the record is easier to sustain than a meeting that gets rescheduled repeatedly. Escalate exceptions to the governance committee rather than letting them accumulate.
Use Expiry and Triggers
Approvals should expire. When an approval lapses, the model’s status should automatically move to a state that prompts action, such as “under review,” and the owner should be notified. Material changes should also trigger reassessment. Define material change explicitly. Changes to the training data source, the model architecture, the intended use, the population affected, or the vendor typically qualify. Routine retraining on the same data with the same architecture may not, depending on your policy, but the decision should be recorded either way.
Monitor Drift and Performance Against Records
Monitoring output belongs in the governance picture, not only in the data science workflow. Where monitoring shows material degradation, bias shifts, or unexpected behavior, the inventory record should reflect the current status and link to the incident or review. A model whose monitoring dashboard shows a problem but whose inventory status still reads “approved” is a governance failure, even if the technical team knows about the problem. Automate the link between monitoring thresholds and status changes where you can.
Handle Retirement Deliberately
Retirement is the most neglected part of the lifecycle. Models are often left running after the business process they supported has ended, or after a replacement is in place. Require a retirement plan for every model in the top tiers, covering the decommissioning date, data disposition, and confirmation that downstream systems no longer depend on the model. Mark the record as retired only when the model is no longer running, and keep the record for the period your retention policy requires. The data retention policy template is a useful starting point for setting disposition rules for the artifacts a retired model leaves behind.
Reconcile Regularly and Measure Accuracy
Schedule periodic reconciliation between the inventory and your discovery sources. Track metrics such as the proportion of models with a current owner, the proportion with a current validation, the number of models found through discovery that were not self-registered, and the average age of open reviews. These measures tell you whether the inventory is trustworthy and where to invest. If you are assessing the broader governance program’s maturity at the same time, the data governance maturity assessment can help you judge whether inventory maintenance is a defined, measured process or still ad hoc.
Plan for Scale
As the portfolio grows, manual maintenance becomes the bottleneck. Integrate the inventory with your model registry, deployment pipeline, and procurement system so that records are created and updated from the source rather than typed in later. Use automation carefully: a field populated by a script from an unreliable source can be wrong with a veneer of authority. Validate automated fields against a sample periodically.
Sustain Ownership of the Register Itself
The inventory needs a team that owns it, with capacity to chase owners, reconcile sources, and produce reports. In my experience, programs fail when the inventory is treated as a side task for a busy analyst. Name the function, fund the time, and make inventory health a standing agenda item for the governance committee.
The Bottom Line on AI Model Inventories
An AI model inventory is the operational backbone of AI governance, and the teams that get it right treat it as a living control rather than a compliance document. The essential ingredients are simple to describe and hard to sustain: a clear scope, a disciplined set of fields, accountable owners who are named individuals, risk tiers that change what happens, lifecycle states that trigger action, and a discovery process that keeps finding the models nobody registered. In financial services, where model inventories have underpinned model risk management for years, the lesson holds: the institutions that stay in control are the ones that can answer basic questions quickly and with evidence.
If I were starting again tomorrow, I would define the owner and tier fields first, populate the highest-risk models before anything else, and build the intake checkpoints into existing procurement and change processes before buying a tool. I would also treat shadow AI discovery as an amnesty and an intake channel rather than an enforcement exercise, because the disclosures you most need depend on people feeling safe to make them. Keep the inventory lean, keep the detailed evidence in the model documentation, and keep the links between them tested. That combination is what turns a register from a spreadsheet into a governance control that holds up under regulatory scrutiny.
Frequently Asked Questions About AI Model Inventory
What is an AI model inventory?
An AI model inventory is a controlled, versioned register of every AI model an organization builds, buys, or uses. Each record captures the model’s accountable owner, purpose, risk tier, lifecycle status, and links to supporting evidence. It gives governance teams one authoritative list for oversight, audit, and regulatory response.
How is an AI model inventory different from a model card?
A model card documents a single model in depth, covering intended use, training data, evaluation, and limitations. An AI model inventory describes the whole portfolio and records governance attributes such as owner and risk tier. The inventory should link to each model’s card rather than duplicate its content.
What fields should an AI model register include?
A core register needs a stable identifier, name, business purpose, out-of-scope uses, accountable and technical owners, risk tier, lifecycle status, data sources, and links to validation and monitoring evidence. Add regulatory classification and provider or deployer role for regulated systems. Keep detailed technical content in model documentation.
How do you find AI models that teams are already using?
Start with procurement records, cloud billing exports, engineering service catalogs, and model registries, then reconcile them. Survey business units using outcome-focused questions, since people often do not label their tools as AI. Review SaaS release notes for embedded AI features, and offer a registration amnesty to surface shadow AI without fear of enforcement.
Who should own an AI model in the inventory?
The accountable owner should be a named individual with authority over the business outcome the model supports, usually a business leader rather than a data scientist. A separate technical owner maintains the model. Assign an owner before any model is promoted to production, and treat orphaned models as governance incidents.
How many risk tiers should an AI model inventory use?
Most organizations use three to five tiers, which is enough to scale controls without creating confusion. Tiers should reflect potential harm, affected population, degree of automation, and regulatory exposure rather than algorithmic sophistication. Define each tier with examples from your own portfolio so teams can classify new models consistently.
Does the EU AI Act require an AI model inventory?
The EU AI Act does not name a single inventory as a requirement, but it imposes classification, documentation, logging, and oversight obligations that an inventory is well suited to manage. Organizations should verify current application dates against official sources, since timelines have been subject to proposed amendment. [VERIFY: current EU AI Act high-risk application dates as of 2026.]
How often should an AI model inventory be reviewed?
Review frequency should follow risk tier, with top-tier models reviewed at least annually and whenever a material change occurs. Approvals should expire automatically to force re-review. Reconcile the inventory against discovery sources on a regular schedule, and report inventory health to the governance committee, because stale records create false confidence.
What is shadow AI and why does it matter for the inventory?
Shadow AI refers to AI tools and models used without formal approval or registration, such as personal accounts on general-purpose chat tools or AI features switched on inside existing software. It matters because these uses often touch sensitive data or decisions without governance. Discovery and a clear acceptable-use standard bring them into the register.
