Wiki · Concept · Last reviewed June 15, 2026

AI System Inventory

An AI system inventory is a maintained catalog of AI systems, use cases, models, vendors, owners, risks, lifecycle status, and accountability paths, built so an institution can know what AI it actually uses before it claims to govern it.

Definition

An AI system inventory is a structured register of AI systems and AI-supported use cases inside an organization. At minimum, it records what the system is called, who owns it, what task it supports, which people or groups may be affected, which data and vendors are involved, what lifecycle stage it is in, whether it is high-impact or otherwise risky, and where review, appeal, incident, and retirement records live.

The inventory is not the same as an audit, model card, impact assessment, procurement file, or public transparency portal. Those can draw from it. The inventory is the map that lets the institution find the system in the first place. Without that map, AI Governance, AI Audits and Assurance, AI Incident Reporting, and Human Oversight of AI Systems become reactive rituals around systems that may already be deployed, copied, retired, or hidden in workflow tools.

How It Works

A useful inventory begins with discovery: procurement records, cloud accounts, model endpoints, vendor contracts, code repositories, spreadsheets, employee tools, help desks, agent integrations, and business-unit pilots. Each entry receives a stable identifier so later documents can point to the same system rather than re-describing it under a new name.

Typical fields include purpose, owner, sponsor, vendor, model or service, data categories, affected population, output type, human-review role, lifecycle stage, deployment date, risk classification, evaluation evidence, security boundary, logging rule, incident contact, and decommissioning path. Public inventories may omit sensitive technical details, but the internal record should still preserve enough evidence for review.

The inventory should feed other controls. A high-impact classification can trigger an Algorithmic Impact Assessment. A public-facing system can require a plain-language notice. A model update can trigger a change log. A serious failure can connect to incident reporting. The inventory turns scattered AI use into a governed object.

Current Context

As of June 15, 2026, the clearest public examples are in government. OMB Memorandum M-25-21 applies to new and existing AI developed, used, or acquired by covered U.S. federal agencies and continues the federal use-case inventory approach. OMB also maintains a public GitHub repository for the 2025 Federal Agency AI Use Case Inventory. The repository notes reporting scope changes aligned to M-25-21 and rules for retired use cases.

Agency inventories show both the promise and the messiness. The Department of Homeland Security publishes a full inventory and a library page for downloadable files. The Department of Justice says its 2025 AI Use Case Inventory includes 315 entries, a 30.7 percent increase from 2024. These numbers are not proof of safety. They are proof that discovery itself is becoming part of AI administration.

Outside the U.S. federal context, the EU AI Act establishes an EU database for high-risk AI systems listed in Annex III, while city-level projects such as Helsinki's AI Register and the Algorithmic Transparency Standard show a municipal form: residents can see system purposes, descriptions, governance information, and feedback paths.

Governance and Safety

The basic safety problem is invisibility. A chatbot, classifier, recommender, translation tool, scoring model, spreadsheet assistant, or agent workflow can affect people before the organization has named it as an AI system. That invisibility weakens risk classification, procurement review, privacy analysis, security testing, accessibility review, bias testing, human oversight, and appeal.

The basic governance problem is inventory theater. A weak inventory can become a static spreadsheet of vendor names, with no lifecycle status, no affected population, no risk category, no owner, no evidence, and no correction path. A stronger inventory is live infrastructure: it changes when systems change, and it connects to logs, audits, incident reports, procurement decisions, and public notice where disclosure is lawful.

Defense Pattern

Spiralist Reading

The AI system inventory is institutional memory under pressure.

AI enters organizations through contracts, pilots, browser tabs, spreadsheets, agents, vendor features, and exhausted workers trying to finish tasks. The inventory is the moment the institution admits that deployment is not only a launch announcement. It is the full set of places where machine output starts shaping work, judgment, and public life.

Open Questions

Sources

Return to Wiki