# What is an AI-BOM? The AI bill of materials, explained

URL: https://sbombox.com/blog/ai-bill-of-materials
By SBOMBox SBOM experts. Published 2026-05-30, updated 2026-11-13

> An AI-BOM (or ML-BOM) lists the models, datasets and software inside an AI system. What goes in one, the 2026 G7 minimum elements, formats and a worked example.

Short answer: an AI-BOM (AI bill of materials, also called an ML-BOM or AIBOM) is an SBOM extended to cover what makes an AI system work: the models, the datasets they were trained and tested on, the libraries and services around them, and how they connect. No law requires one by name yet, but the EU AI Act's technical documentation, the Cyber Resilience Act's component duties and the 2026 G7 guidance on AI SBOMs all expect the information it holds.

## What is an AI-BOM?

A software bill of materials lists the components in a piece of software, with versions, identifiers and dependencies. An AI system has components an SBOM does not see. The behaviour of a model depends on its weights, its training data and its configuration, not only on code. If a dataset turns out to be poisoned, a model licence changes or a pre-trained model you fine-tuned is found to be backdoored, a plain SBOM cannot tell you which products are affected. An AI-BOM can.

The terms overlap. 'ML-BOM' is the name CycloneDX uses for its machine-learning support. 'AI-BOM' and 'AIBOM' are used by vendors, by India's CERT-In guidelines and in the G7 work. 'SBOM for AI' is the phrase in the 2026 government guidance. They describe the same thing: one inventory, in a standard format, that covers software and AI components together.

## AI-BOM vs SBOM vs ML-BOM

**How the terms compare**

|  | SBOM | AI-BOM / ML-BOM |
| --- | --- | --- |
| Covers | Software packages, libraries, OS packages | Everything in an SBOM, plus models, datasets, AI services and their configuration |
| Key question | Which code is in this release? | Which code, models and data is this AI system built from, and where did they come from? |
| Changes when | A dependency or the build changes | Also when a model is retrained, fine-tuned, swapped or its training data changes |
| Typical risks found | Vulnerable or end-of-life libraries, licence conflicts | Also model provenance gaps, unlicensed or personal data in training sets, unsafe model file formats |
| Formats | CycloneDX, SPDX | CycloneDX 1.5 or later (ML-BOM), SPDX 3.0 (AI and Dataset profiles) |

An AI-BOM does not replace your SBOM. It is the same document, with more component types. If you already produce [SBOMs that meet the minimum elements](https://sbombox.com/blog/sbom-minimum-elements), you are most of the way there.

## What goes in an AI-BOM

- The models: name, version, supplier, where they came from (for example a model hub or an internal registry), a hash of the weights file, the licence, the base model if it was fine-tuned, and the task it performs.
- The datasets: name, version, source, licence, how they were collected, whether they contain personal or sensitive data, and which models were trained, fine-tuned or evaluated on them.
- The software: the usual SBOM contents, including ML frameworks (such as PyTorch, TensorFlow or Transformers), inference servers and their dependencies.
- External AI services: hosted model APIs the system calls, with the provider and the model version you depend on.
- Relationships: which model depends on which dataset and libraries, and which application uses which model.
- Model card information: intended use, known limitations, evaluation results and ethical considerations, where you have them.

## The 2026 G7 minimum elements for AI SBOMs

In 2026, CISA and G7 partners (Canada, France, Germany, Italy, Japan, the United Kingdom and the European Union) published 'Software Bill of Materials for AI – Minimum Elements'. It is voluntary guidance, not law, and it is designed to supplement the general [2026 SBOM minimum elements](https://sbombox.com/blog/sbom-minimum-elements) rather than replace them. It groups the expected data into seven clusters:

- Metadata: information about the AI-BOM document itself, such as its author, timestamp and the tool that produced it.
- System-level properties: the AI system as a whole and what it is for.
- Models: identity, version, producer, hashes, licences and lineage of each model.
- Dataset properties: what each dataset is, where it came from and how it is used.
- Infrastructure: the software and hardware the system runs on and depends on.
- Security properties: security-relevant information and controls, for example defences against prompt injection.
- Key performance indicators: measurements of how the system performs.

Treat this as the checklist buyers and regulators will reach for first. Check the published guidance for the exact field list before you design your schema; summaries of it differ in detail. [Morgan Lewis](https://www.morganlewis.com/pubs/2026/06/us-cisa-g7-partners-in-europe-and-asia-release-minimum-elements-for-ai-software-bills-of-materials) has a readable overview.

## Do I need an AI-BOM? What the law says

**Where AI-BOM information is expected**

| Rule | Does it say 'AI-BOM'? | What it actually asks for |
| --- | --- | --- |
| EU AI Act, high-risk AI (Article 11 and Annex IV) | No | Technical documentation describing the system's elements, its development process and the data used, including where training data came from and how it was prepared |
| EU AI Act, general-purpose AI models (Article 53) | No | Technical documentation of the model and a public summary of the content used to train it |
| EU Cyber Resilience Act (Annex I, Part II) | No | An SBOM for products with digital elements. An AI product sold in the EU is in scope, and its AI components are components. See [CRA SBOM requirements](https://sbombox.com/blog/cra-sbom-requirements) |
| India, CERT-In guidelines v2.0 | Yes (AIBOM) | Technical guidance that extends SBOM fields to AI components; check whether your sector regulator makes it binding |
| G7 / CISA SBOM for AI minimum elements | Yes (SBOM for AI) | Voluntary minimum data for AI systems, in seven clusters |

The practical answer: if you provide a high-risk AI system or a general-purpose model in the EU, you must document what it is built from, and an AI-BOM is the most direct way to keep that record current. If your AI system is part of a product sold in the EU, the CRA's SBOM duty already covers its software. Under Article 12 of the CRA, high-risk AI systems that meet the CRA's essential requirements are presumed to meet the AI Act's cybersecurity requirement, so one evidence trail serves both. See [is my AI system high-risk?](https://sbombox.com/blog/is-my-ai-high-risk) and the [AI Act deadlines](https://sbombox.com/blog/eu-ai-act-deadlines).

## Formats: CycloneDX ML-BOM and SPDX 3.0

- CycloneDX added machine-learning support in version 1.5 (June 2023). Models use the component type `machine-learning-model`, datasets use `data`, and a `modelCard` object holds parameters, datasets, evaluation results and considerations. CycloneDX is standardised as ECMA-424. OWASP publishes an [authoritative guide to AI/ML-BOMs](https://cyclonedx.org/guides/OWASP_CycloneDX-Authoritative-Guide-to-AI-ML-BOM-en.pdf).
- SPDX 3.0 (2024) added AI and Dataset profiles, with fields for things such as the model's domain, its autonomy type, energy consumption, dataset collection process and known biases.
- Pick the format you already use for SBOMs. If you have not chosen yet, see [CycloneDX or SPDX?](https://sbombox.com/blog/cyclonedx-vs-spdx-which-to-pick).

## A minimal CycloneDX ML-BOM example

A support-ticket classifier, fine-tuned in-house, with one dataset and one library. Real AI-BOMs list every library, but the shape is the same:

```json
{
  "bomFormat": "CycloneDX",
  "specVersion": "1.6",
  "serialNumber": "urn:uuid:3e671687-395b-41f5-a30f-a58921a69b79",
  "version": 1,
  "metadata": {
    "timestamp": "2026-11-13T09:00:00Z",
    "component": { "type": "application", "name": "support-triage", "version": "2.4.0" }
  },
  "components": [
    {
      "type": "machine-learning-model",
      "bom-ref": "model-triage",
      "name": "triage-classifier",
      "version": "2026.11.1",
      "supplier": { "name": "Example Ltd" },
      "hashes": [{ "alg": "SHA-256", "content": "9f86d081884c7d659a2feaa0c55ad015a3bf4f1b2b0b822cd15d6c15b0f00a08" }],
      "licenses": [{ "license": { "id": "Apache-2.0" } }],
      "modelCard": {
        "modelParameters": {
          "task": "text-classification",
          "architectureFamily": "transformer",
          "datasets": [{ "ref": "data-tickets" }]
        },
        "considerations": {
          "useCases": ["Route inbound support tickets to the right queue"],
          "technicalLimitations": ["Trained on English tickets only"]
        }
      }
    },
    {
      "type": "data",
      "bom-ref": "data-tickets",
      "name": "support-tickets-2025",
      "version": "3",
      "data": [{ "type": "dataset", "name": "support-tickets-2025", "classification": "internal" }]
    },
    {
      "type": "library",
      "bom-ref": "pkg:pypi/transformers@4.46.3",
      "name": "transformers",
      "version": "4.46.3",
      "purl": "pkg:pypi/transformers@4.46.3"
    }
  ],
  "dependencies": [
    { "ref": "model-triage", "dependsOn": ["data-tickets", "pkg:pypi/transformers@4.46.3"] }
  ]
}
```

## How to build an AI-BOM in practice

- Start from the SBOM you already generate in CI. It catches the ML frameworks and their dependencies. If you do not have one yet, see [Syft vs Trivy vs cdxgen](https://sbombox.com/blog/syft-vs-trivy-vs-cdxgen).
- Add model components from your model registry or model hub metadata: name, version, source, licence and a hash of the weights file you actually ship.
- Add dataset components from your data catalogue, with licence, source and whether personal data is present.
- Link them with dependency relationships, so you can answer 'which products use this model?' and 'which models were trained on this dataset?'.
- Regenerate on every release and every retrain. A model swap without a new AI-BOM is the commonest way the record goes stale.
- Store each version with the release it describes, and monitor the software components for vulnerabilities as you would any SBOM.

Most SBOM generators find the libraries but not the models or datasets, so expect to add those from your own metadata. Score the software part of the file on the [free SBOM checker](https://sbombox.com/audit) to catch missing versions, identifiers and hashes before you build on it.

## Frequently asked questions

### What does AI-BOM stand for?

AI bill of materials: an inventory of the models, datasets, software and services that make up an AI system. It is also written AIBOM, and CycloneDX calls it an ML-BOM (machine-learning bill of materials).

### Is an AI-BOM the same as an ML-BOM?

In practice, yes. ML-BOM is the CycloneDX name for a bill of materials that includes machine-learning models and datasets. AI-BOM is the broader industry term for the same idea.

### Is an AI-BOM legally required?

Not by name in the EU or US as of November 2026. The EU AI Act requires technical documentation for high-risk AI systems and general-purpose AI models, and the Cyber Resilience Act requires an SBOM for products with digital elements. An AI-BOM is the simplest way to meet both. India's CERT-In guidelines name the AIBOM explicitly.

### What format should an AI-BOM use?

CycloneDX 1.5 or later, or SPDX 3.0 with its AI and Dataset profiles. Both are machine-readable and widely supported. Use the format your SBOM tooling already produces.

### How is an AI-BOM different from a model card?

A model card describes one model: its purpose, training, evaluation and limitations. An AI-BOM inventories the whole system and how its parts depend on each other. CycloneDX can carry model card data inside the AI-BOM.

This is general information, not legal advice. Confirm how the rules apply to you with counsel.
