
AI-BOM: a bill of materials for AI systems
In short: An AI bill of materials (AI-BOM) lists the models, datasets, prompts, tools and libraries inside an AI system, so risk can be traced when one changes.
An AI bill of materials (AI-BOM), also written ML-BOM, is a structured inventory of what an AI system is built from: base models and versions, fine-tuned adapters, training and evaluation datasets, prompts, retrieval sources, tools and libraries. It extends the software bill of materials (SBOM) idea to AI, so teams can answer "where do we use this model or dataset?" when something changes or goes wrong.
AI-BOM in plain words
It is the ingredients label on a food packet. When one ingredient is recalled, you know at once which products contain it.
How it works
Each AI system records its components in a machine-readable format, generated as part of the build and deployment pipeline. Two open standards have a place for it, as read in November 2026. CycloneDX's ML-BOM capability covers models, datasets and configurations, and describes its purpose as model and dataset transparency for security, privacy, safety and ethical considerations. The SPDX 3.0.1 specification includes dedicated AI and Dataset profiles.
Entries carry versions, licences, sources and owners. A central inventory then aggregates them across the organisation.
Example
As an illustration, a provider announces that a model version will be retired, or a library used by an MCP server has a vulnerability. With AI-BOMs in place, the platform team queries the inventory and gets the list of affected features and their owners in minutes, where the alternative is a week of emails.
Where you'll meet it
- Security and supply-chain programmes extending SBOM practice
- Governance reviews that ask what an AI feature is built from
- Procurement questionnaires asking vendors what models they use
- Incident response for model, data or dependency issues
What it changes for you
For security teams, AI components become trackable supply-chain items. For engineering leads, recording models, prompts and tools per release turns audits and migrations into queries. For buyers, asking vendors for an AI-BOM reveals which third-party models handle your data.
AI-BOM vs model card
A model card documents one model's purpose, training, performance and limits. An AI-BOM lists all the components in a system, including models, data, prompts and tools.
Common mistake
My view: listing only the base model. A prompt change, a retrieval source, a tool or a library can change behaviour just as much, and the published formats centre on models and datasets, so you have to add those parts yourself. If it can change behaviour, it belongs in the BOM.
Related terms
Model card · AI governance · Model Context Protocol · Shadow AI · EU AI Act
From the archive: AI BOMs need behavioral inventory
If a component can change the system's behaviour, it belongs in the AI-BOM.


