What Is an AI Bill of Materials?
The National Institute of Standards and Technology calls AIBOMs “enablers for AI software transparency and security,” pointing to their ability to also “foster trust” and “facilitate innovation.” Specifically, it is a “repository or inventory,” Soni says, “that can be read by your machines, by your systems, and it comprises components including your data sets, your prompts, your models, your specific configurations, version history, pipelines and third-party dependencies.”
Katie Norton, research manager for DevSecOps and software supply chain security at IDC, says that for organizations already using SBOMs, AIBOMs are a logical next step: “While an SBOM provides visibility into application code, libraries and dependencies, an AIBOM captures the components that define AI behavior.”
What Goes Into an AI Bill of Materials?
AIBOMs typically have similar components, though they might vary a bit based on an organization’s size, needs and systems.
Norton says that these structured, machine-readable inventories document multiple layers of the AI systems in the organization, including components and questions to ask for each:
- The data layer includes training and validation data sets, provenance, licensing and sensitivity. This layer answers the questions, “Where did your training data come from? What are the licensing terms? Does it contain protected health information or personally identifiable information?”
- The model layer includes architecture, weights, hyperparameters, versioning and lineage. It answers the questions, “What architecture? What version? What was the training configuration?”
- The infrastructure and dependency layer includes frameworks and hardware required to run the model. It answers the questions, “What frameworks and libraries does it rely on, and where does it run?”
- The governance metadata layer includes intended use, known limitations and risk mitigations. It answers the questions, “What is this model supposed to do? What are its limitations? What safeguards are in place?”
Norton adds that organizations need an AIBOM in addition to their SBOM. “An SBOM alone is insufficient for AI systems because it only inventories code,” she says. “AI systems are data-driven and often nondeterministic; their behavior emerges from training data and model configuration rather than explicitly written logic. Without an AIBOM, IT leaders lack visibility into the ‘cognitive layer’ of the system, making it difficult to audit decisions, reproduce results or assess supply chain risk.”
EXPLORE: Inova forges forward on its AI journey with mindful moves.
Why the AI Bill of Materials Is Gaining Traction Now
AI is not new, so why is the AIBOM just starting to become more widely adopted? Because AI has to be “ethical, transparent and fair,” Soni says, a framework is needed to analyze those principles, and AIBOMs have begun to do just that — especially important at a time when “blind spots” have become apparent and the need for regulatory compliance has become mandatory.
Norton points to three things converging to create the need for AIBOMS now: “First, generative AI made it trivially easy for developers to drop open-source models into applications without anyone in security knowing about it. Organizations suddenly realized they had no idea what models were running in production. Second, regulators caught up. The EU AI Act and NIST AI Risk Management Framework now expect transparency around training data and model lineage — things SBOMs were never designed to capture.”
And last, “the tooling finally exists. Standards like the Software Package Data Exchange (SPDX) and CycloneDX now have AI-aware profiles, so generating an AIBOM is no longer a custom engineering project. The risk was always there; now we have the means to address it,” Norton says.
While Gartner predicts SBOM adoption will rise from 56% among large enterprises in 2025 to 85% by 2028, the adoption rate for AIBOMs is yet to be determined.
Click the banner below to sign up for HealthTech’s weekly newsletter.
