What Federal AI Inventories Reveal—and Conceal—About Government’s Algorithmic Turn

On October 7, 2026, the Justice Department updated a spreadsheet listing 315 AI use cases—30.7% more than in its previous inventory. The number is arresting, but the more revealing detail is buried in the notes: similar uses may be combined into a single department-wide entry; old identifiers may vanish; and some information is withheld under disclosure rules. That is the right way to approach the federal AI-use-case inventory: not as a definitive census of government algorithms, but as a public ledger with real evidentiary value and hard-built blind spots. Department of Justice, “AI Use Case Inventory”

A public record designed for management first

Federal agencies are required, with important exceptions, to inventory AI use cases at least annually, submit them to the Office of Management and Budget, and post public versions on their own websites. The Department of Defense is exempt from the requirement to inventory individual use cases, and elements of the Intelligence Community are excluded. OMB’s current policy also tells agencies to update public inventories on an ongoing basis where possible. In other words, this is a continuing administrative reporting regime, not a one-time disclosure project—and it covers only the publicly releasable part of a broader internal picture. OMB Memorandum M-25-21, “Accelerating Federal Use of AI through Innovation, Governance, and Public Trust”

The inventory’s basic unit is the use case: a defined problem or task for which an agency develops, buys or deploys AI. That is a more useful concept than a list of vendors, because citizens usually care what a system does in practice. GSA’s 2025 inventory, for example, distinguishes an internal writing and research chatbot from tools proposed to process public comments, classify spending, detect cyber threats, review procurement documents or help with remote identity verification. It also marks whether an entry is pre-deployment, a pilot or actively deployed. Those fields can prevent a common mistake: treating every row as a system already making real-world decisions. GSA, “2025 GSA AI use cases”

That modest-looking status field matters. A pre-deployment tool describes an intention, acquisition or development effort; a pilot is limited testing; a deployed use case is authorized or being used in support of an agency function. None of those labels, by themselves, tells a reader how many people are affected, how frequently the tool is used, whether staff follow its recommendations, or whether it has improved a service. Inventories are maps of declared activity, not performance audits.

What citizens can genuinely learn from the rows

Read carefully, the inventories offer a valuable answer to a concrete democratic question: where is AI entering the machinery of government? They reveal which agencies are experimenting, which kinds of work are changing, and whether the work is operational or still aspirational. A 2025 Brookings analysis of the public records found that 41 agencies documented more than 3,600 individual use cases, up from the prior year; it also found that many entries concerned back-office work, while others touched benefits, health services and law enforcement. The shift is not merely rhetorical: it makes it possible to identify agencies and functions that deserve closer reporting, public-comment scrutiny or congressional oversight. Brookings, “Assessing the state of AI adoption across the federal government”

The inventories can also reveal the difference between an AI system that assists a worker and one that may shape a consequential decision. Under OMB’s 2025 framework, “high-impact AI” is AI whose output serves as a principal basis for an action or decision with a legal, material, binding or significant effect on rights, safety, privacy, or access to domains such as education, housing, insurance, credit, employment or benefits. For such systems, agencies must conduct pre-deployment testing and AI impact assessments, monitor performance and adverse effects, arrange appropriate human oversight, and discontinue non-compliant functionality. That classification is not a guarantee of safety, but it is a crucial flag for readers: it identifies cases for which the government itself says higher-stakes governance should apply. OMB Memorandum M-25-21, “Accelerating Federal Use of AI through Innovation, Governance, and Public Trust”

They can further expose the scale of change without reducing it to hype. GAO found that, across 11 selected agencies with inventories, reported AI use cases nearly doubled from 571 in 2023 to 1,110 in 2024; reported generative-AI cases rose from 32 to 282. That does not demonstrate that government has been transformed by chatbots. It does demonstrate a rapid expansion of reported experimentation and gives the public a baseline against which agencies’ later claims of efficiency, modernization or risk control can be tested. GAO, “Artificial Intelligence: Generative AI Use and Management at Federal Agencies”

What the spreadsheet cannot prove

The inventory cannot establish completeness. It depends on agencies identifying, defining and reporting their own uses, and public versions exclude material that cannot be released. DOJ explicitly says it consolidates similar uses and withholds some information under existing law, policy and regulation. Those choices can be sensible: publishing operational details about sensitive investigations or security systems may create real risks. But the consequence is unavoidable. A missing row does not prove the government is not using AI for that purpose; it may indicate exclusion, aggregation, a reporting gap or a nonpublic system. Department of Justice, “AI Use Case Inventory”

Nor can a short description answer the questions most important to a person affected by an automated process. Which data were used? What model and version are in production? What error rate appears in the relevant population? Does a human reviewer meaningfully disagree with the system, or merely ratify it? Can someone obtain an explanation, correct bad data, appeal an outcome or learn that AI was involved? A use-case title and a two-paragraph purpose statement may identify where to investigate; they rarely supply enough evidence to judge fairness, reliability or due process.

Data quality remains a live concern, rather than a historical embarrassment that reporting rules have solved. GAO’s review of early inventories found incomplete or inaccurate entries at 15 of 20 agencies that reported uses, including missing required fields and cases later found not to be AI at all. Its updated recommendation tracker shows improvement at some departments but continuing issues in 2025 inventories at others, including missing dates and incomplete information about high-impact uses. The right conclusion is neither that the inventories are worthless nor that they are verified ground truth. They are administrative disclosures whose credibility has to be checked against source documents, procurement records, privacy assessments, audit findings and, where appropriate, FOIA releases. GAO, “Artificial Intelligence: Agencies Have Begun Implementation but Need to Complete Key Requirements”

The classification problem is also a governance problem

One especially important limitation is that categories can change while the underlying activity continues. The 2025 framework narrowed the trigger for heightened treatment from systems that “control or significantly influence” outcomes to systems whose output is a “principal basis” for an outcome. Brookings found that some use cases previously tagged as rights- or safety-impacting were no longer classified as high-impact, often because agencies said the AI did not form the principal basis for a decision. That may be a defensible judgment in individual cases. Yet it also means that readers should not compare high-impact counts across years as if they were a simple measure of danger. Definitions, disclosure practices and names all move. Brookings, “Assessing the state of AI adoption across the federal government”

The newest federal repository improves the raw material for scrutiny by consolidating agency-posted, machine-readable files and publishing data dictionaries, reporting instructions and sourcing information. But even the repository describes itself as a compilation of agencies’ publicly available inventories. It is an index, not an independent audit. That distinction is easy to lose when a CSV file becomes a chart, a headline or a claim that the public now knows how government AI works. OMB, “2025-Federal-Agency-AI-Use-Case-Inventory”

How to read an inventory like a citizen, not a spectator

• Start with function, not the word AI. Ask what action the tool supports: drafting, search, fraud detection, eligibility, enforcement, diagnosis, prioritization or identity verification.
• Check lifecycle stage. A pilot is evidence of testing, not proof of a permanent operational system.
• Look for the high-impact designation—and then look beyond it. If an entry touches benefits, employment, health, policing or identity, seek its impact assessment, privacy documentation, procurement record and appeal procedures.
• Treat counts cautiously. A larger inventory may reflect more adoption, better reporting, a broader definition or the unbundling of one program into several entries. A smaller one may reflect consolidation rather than retreat.
• Follow the trail outward. The inventory is the lead, not the verdict. Compare it with agency compliance plans, inspector-general and GAO reports, public notices, contracts and the agency’s own performance evidence.

The most surprising lesson of this unglamorous artifact is that its weakness is also its civic use. A spreadsheet cannot certify that an AI system is just, accurate or necessary. But it can force a scattered technical reality into names, owners, categories and dates—enough structure for the public to ask better questions. Federal AI transparency will become meaningful not when every agency produces a longer list, but when a listed use case reliably opens onto the evidence needed to evaluate what the system does to people, who remains accountable, and what happens when it gets something wrong.