Skip to content
Article reader Listen + reading controls
LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

0:00 0:00
Reading settings
Text size
100%

A federal AI inventory is not an adoption strategy

In late 2023, twenty civilian agencies reported roughly 1,200 current and planned uses of artificial intelligence. It was an irresistible number. Read quickly, it sounded like proof that artificial intelligence (AI) was spreading across the federal government at scale.

The U.S. Government Accountability Office (GAO) found something more complicated. Only five of the twenty agencies provided complete information for every reported use case. Two inventories included systems that the agencies later determined were not AI at all. Many entries lacked basic information such as lifecycle stage or whether the use could be publicly released.

The inventory documented attention. It did not establish adoption, readiness, value, or control.

That distinction remains important because federal AI inventories have continued to grow. The number of ideas in a spreadsheet will always rise faster than the number of capabilities that survive acquisition, integration, evaluation, workforce adoption, and sustained operation. Leaders need to know which kind of growth they are looking at.

Revised and substantially expanded July 17, 2026, using the subsequent federal policy and oversight record.

Visibility is necessary

An agency cannot govern systems it cannot see. A useful inventory establishes a shared view of where AI is being researched, acquired, developed, tested, operated, or retired. It helps security, privacy, legal, acquisition, data, civil-rights, mission, and technology leaders find work that would otherwise remain fragmented across programs and contracts.

Public inventories also create a modest form of democratic visibility. They allow Congress, watchdogs, researchers, affected communities, and other agencies to see where automated capabilities may influence public services or government operations.

That is why the requirement has survived policy changes. Office of Management and Budget (OMB) Memorandum M-25-21, issued in April 2025, continued to require agencies to inventory AI use cases at least annually, submit the inventories to OMB, and publish versions appropriate for public release. It also required central tracking and additional practices for high-impact uses.

The persistence of the inventory is sensible. The problem begins when the count becomes the strategy.

A use case can mean almost anything

The phrase “use case” often collapses several very different states:

  • someone has described an opportunity;
  • a team is testing technical feasibility;
  • an agency is acquiring a product;
  • a system is being evaluated with representative users and data;
  • a capability is operating in a bounded environment;
  • a capability has become part of normal mission delivery; or
  • a system is being retired.

Treating all seven as equivalent makes a portfolio look larger while making it harder to manage. An idea should be inexpensive and easy to stop. An operational system needs ownership, monitoring, funding, incident response, user support, and evidence that remains current as the system changes.

GAO's original review of agency inventories made the data-quality problem visible. A later 2025 review of generative AI use showed how quickly the portfolio was changing: across eleven selected agencies, reported AI use cases nearly doubled from 571 in 2023 to 1,110 in 2024, while reported generative AI uses increased much faster. Agency officials still identified technical capacity, budget, data protection, and rapidly changing policy as material constraints.

Growth in reported demand and growth in operational capability are different measures.

The inventory should be a management instrument

A static annual spreadsheet is optimized for reporting. A useful registry is optimized for decisions.

For each use case, leaders should be able to answer:

  • What mission problem or service outcome is being addressed?
  • What baseline describes the current process?
  • What stage is the work actually in, and what evidence justifies that classification?
  • Who owns the mission outcome, technical system, data, risk decision, and operating budget?
  • Which models, data sources, vendors, infrastructure, and human workflows does it depend on?
  • Who is affected, and what happens when the system is wrong, unavailable, or misused?
  • Which evaluation supports expansion, and which threshold would stop the work?
  • When was the evidence last refreshed?

Those fields turn an inventory into a portfolio-control surface. Leaders can see duplicate investments, common infrastructure needs, high-impact uses without adequate evidence, pilots without operating owners, and capabilities whose value no longer justifies their cost.

The registry should also preserve decisions. If a team determines that a tool is not AI, that a use is too risky, or that ordinary software solves the problem better, the reasoning is organizational knowledge. Deleting the row to improve the portfolio's appearance guarantees that another team will pay to relearn the same lesson.

Count transitions, not entries

The most revealing portfolio measures describe movement and residue:

  1. Time to a credible decision. How long does it take to move from a proposed use to a well-supported build, buy, redesign, or stop decision?
  2. Pilot survival for the right reasons. Which experiments become operations because evidence supports them—not because a sponsor protected them?
  3. Reuse. Which data products, evaluation methods, platforms, acquisition clauses, and controls serve more than one mission team?
  4. Operational performance. Does the capability improve the mission outcome relative to the baseline, for the people and conditions that matter?
  5. Risk performance. Are consequential errors, incidents, overrides, complaints, and distributional effects visible and acted upon?
  6. Retirement quality. Can the agency stop a use cleanly, preserve necessary records, and return the work to a safe alternative process?

An agency with 40 well-understood systems may be more capable than one reporting 400 loosely defined uses. The smaller portfolio may show greater discipline, reuse, and operational ownership.

Governance should follow the lifecycle

Inventory fields alone will not manage a system. The registry has to connect to the work around it.

An acquisition should create or update the record. Security and privacy reviews should attach their decisions and conditions. Test results should be linked to the version of the system they evaluated. Significant model, data, vendor, or workflow changes should trigger new evidence. Incidents and user feedback should update the risk picture. Retirement should close access, contracts, data retention, and public reporting obligations.

This is where many inventories fail. They describe systems but do not participate in their lifecycle. Teams update the spreadsheet for a deadline while procurement, engineering, governance, and operations continue elsewhere. The inventory becomes a second, increasingly stale representation of the agency.

OMB's current policy points toward a more integrated model by tying inventories to high-impact determinations, impact assessments, risk practices, and accountability reviews. Agencies should carry that logic further: enter information once, at the point where the work occurs, and generate reporting views from a living system of record.

Do not reward AI theater

Senior leaders influence inventory quality through the questions they ask. “How many AI use cases do we have?” rewards relabeling, fragmentation, and speculative entries. “What mission capability improved, what did we learn, and what did we stop?” rewards evidence.

The 1,200-use-case figure was not proof that civilian agencies had operationalized AI. It was proof that government needed a better way to see and manage a rapidly expanding field of activity. GAO's finding that the underlying data were incomplete and sometimes inaccurate was not a clerical footnote. It was a warning about the quality of the management picture itself.

Federal AI adoption will become credible when agencies can explain not only what appears in the inventory, but also why each investment exists, what evidence supports it, who is accountable for it, and what happens next.

Sources

READER-NEUTRAL SUBSCRIPTION

Follow Field Notes via RSS.

Copy this address into the RSS reader you already use. New notes will appear there automatically—no account, email address, or tracking required.

Open the feed