What Is an AI BOM? How It Differs from SBOMs, Model Cards, and Data Cards—and When to Use Each
Yuta Matsuyama
Security Researcher, TECHFUND Inc.
Updated:
Who this is for: Development leads and security professionals who develop and operate AI features and are asked by business partners to explain their composition and safety.
The practical answer
An AI BOM is composition information that organizes the models, related data, software, and other elements that make up an AI system so that their relationships can be traced. An SBOM primarily describes software composition, a model card explains a model’s uses and evaluations, and a data card explains how a dataset was created and the conditions for its use. Although some fields overlap, their purposes differ, so combine the information you need.
First, decide what you want to know
Suppose someone asks about your company’s AI features: “Which models are used by which applications, and with what data?” Even if you know the model provider, the same document may not describe the application’s library dependencies, the internal documents used for retrieval, or the paths through which inputs reach the model.
An AI BOM is an approach to organizing these components and connections. What it should include varies by purpose and the format you adopt. Deciding first how you will use it—for security response, providing information during procurement, or internal asset management, for example—makes it easier to select the fields you need.
The four types of documentation answer different questions
CycloneDX ML-BOM provides a way to represent models, datasets, dependencies, and more. Meanwhile, research on model cards proposes a method for reporting a model’s intended uses, evaluation conditions, and limitations. Research on datasheets for datasets provides a framework for documenting collection, composition, intended uses, and other details.
These kinds of documentation are not mutually exclusive. For example, CycloneDX also includes fields that represent model card content. Different document names do not necessarily mean that you must create separate files.
| Documentation | Main question | Examples |
|---|---|---|
| AI BOM | What elements and relationships make up the AI feature? | Models used, related data, dependencies, where they are used |
| SBOM | What components make up the software? | Libraries, versions, dependencies |
| Model card | What can the model be used for, and what evaluations and limitations does it have? | Intended uses, evaluation results, known constraints |
| Data card / datasheet | How was the data created, and what can it be used for? | Collection methods, composition, conditions of use, constraints |
These fields are examples. The actual scope of the information provided varies by format and provider.
Consider an application that uses an external model
For example, consider an application that searches internal documents and passes their content to an external model to generate an answer. This is a fictional example for illustration. An AI BOM can record the relationships between the application, the ID of the model it calls, the data used for retrieval, and the software components it uses. An SBOM lists the libraries used by the application, while model cards and data cards, if available, can be consulted for uses and constraints.
However, if the external model’s training data is not public, do not fill in the datasets actually used based on guesswork. Mark information that the provider has not disclosed as “unverified,” and record the date you received a response and the contact for inquiries. The internal documents used for retrieval are data used by the application and should be treated separately from the data used to train the model.
Separate reported information from information you have actually verified
A card published by a model provider, dependencies detected in a repository, and a model ID confirmed in production settings are all “composition information,” but their supporting evidence differs. Keeping a record of the source, verification date, responsible person, and unverified items helps identify where updates or inquiries will be needed later. If an external service does not sufficiently disclose the actual model or its version, record what is available within those limits.
When reviewing code under development, it is important not to assume that a model name in a configuration file gives you a complete picture of all paths in operation. Environment variables or runtime switching may cause a different model to be called. Cross-check references in the code against configuration and usage records you are authorized to inspect, and ask the responsible person about any discrepancies.
Start with a single AI feature
First, identify the feature in scope and the person responsible, then organize the model provider and identifier, the data used, the main dependencies, and the deployment destination. Next, record the evidence for each item and leave undisclosed information marked as unverified. Assigning responsibility for updates when models or settings change also helps ensure that the inventory is not a one-time exercise.
In May 2026, CISA and relevant G7 agencies published guidance on minimum elements for AI to consider in addition to the general SBOM minimum elements. This is not presented as an exhaustive or mandatory list.
CycloneDX and SPDX are options for handling this information in machine-readable form. Check the actual tools and contractual terms to determine which formats and fields the recipient can handle. An AI BOM informs decisions; simply recording information does not prove safety, legal compliance, or the provenance of training data. Evaluation and human verification are separately required.
Common questions
Does having an AI BOM make an SBOM unnecessary?
Not necessarily. Software dependencies can be included in AI composition information, but whether all necessary information is covered depends on the format and tools used to create it. Check this against your intended purpose.
Is training data listed in a model card verified evidence?
A card contains information supplied by the provider. Distinguish whether it has been independently verified and how much has been disclosed, and treat items you cannot verify as unverified.
How Hi AUDIT Agent can help
Once you have organized the model and data configuration in an AI BOM, it is also worth checking the security of the code that powers those features. With Hi AUDIT Agent, you can request an analysis in natural language from an MCP-compatible development environment—for example, “Check the security of this API.” You can review the flagged locations, the reasoning behind the findings, and suggested fixes on the same screen, making it easier to move from taking inventory to reviewing code. Have a person validate the analysis results, and handle AI BOM creation and checks for configuration completeness separately.
Create a Hi AUDIT Agent account