AI Bill of Materials

A declared list says what teams intended to use. An observed list says what they used. The gap between the two is usually the finding.

An AI bill of materials is an inventory of what an AI system is made of — models and their versions, training and fine-tuning data sources, frameworks and libraries, prompts and system instructions, and the tools and external services the system can reach. It is the AI analogue of a software bill of materials, and it exists to answer one question quickly: when something is found to be compromised, what of ours is affected.

Why an SBOM is not enough

A software bill of materials covers code dependencies. An AI system's risk surface includes several things that are not code.

Model weights are artifacts with provenance, licenses and a serialization format that can carry executable content. Training and fine-tuning data determines behavior and is where data poisoning enters. Prompts and system instructions are configuration that changes behavior as decisively as a code change, and are usually versioned nowhere. Tool and MCP dependencies define what the system can reach — see MCP security.

The last two are the ones most inventories miss, and they are the two that change most often.

The inventory goes stale faster than it is maintained

A conventional SBOM is generated at build time from a dependency file, so it is accurate by construction. An AI bill of materials has no equivalent build step. A team swaps a model, adds an MCP server, edits a system prompt — none of it touches a manifest, and the document drifts from reality within weeks.

The structural answer is to derive the inventory from the traffic rather than from a document. A control plane on the request path already observes which models are called, by whom, through which tools, which makes the inventory a query rather than a maintenance task. A declared list says what teams intended to use; an observed list says what they used, and the gap between the two is usually the finding — see shadow AI and runtime governance.

What it is for

Three jobs, in order of how often they come up.

Incident scoping. A model, package or MCP server is reported compromised. Without an inventory, establishing exposure takes days of asking teams — see AI incident response.

Regulatory response. Model inventory is a standing request in supervisory examinations, and it is the first thing asked for rather than the last. NYDFS Part 500 requires a documented asset inventory tracking owner, location, classification and support expiry — a requirement whose transitional deadline passed on 1 November 2025 and which includes AI systems.

Supply chain assurance. Knowing the provenance of a model before it reaches a regulated workload, which is where it meets model allowlists.

ISO/IEC 42001 expects an AI system inventory as part of the management system, and examination readiness generates it from runtime records rather than from a spreadsheet.

Data poisoning · AI incident response · Model allowlist and deny-list · NYDFS Part 500 · Full glossary

Put this in the path of your own agents.

Policy enforced inline between your agents and every model and tool they reach, with a record bound to the human who owns it.

Request a Demo Read the docs