1,818 of 3,611 Federal AI Uses Are Piloted or Deployed. OMB Does Not Say Which

The 2025 federal AI use case inventory lists 3,611 AI uses across 56 agency submissions, and the Office of Management and Budget reports 1,818 of them as piloted or deployed. That is 50.3% of the list. It is not a production rate, because OMB's public summary adds pilots and live systems into a single column. Splitting them with OMB's own data-cleaning log puts the number in production somewhere between 878 and 1,410.

Short answer

How many federal AI projects reach production? In the 2025 inventory, between 878 and 1,410 of 3,611 reported use cases are deployed, which is 24% to 39%. At least 408 more are pilots. The other 1,793 are pre-deployment or retired. OMB publishes only the combined piloted-or-deployed figure of 1,818.

What the federal AI use case inventory actually counts

Every year, US federal agencies must list the AI systems they use or are building. OMB collects the submissions and publishes a consolidated version on GitHub. The 2025 edition carries a summary dated 13 April 2026, and it is the largest public census of AI projects by stage that exists anywhere.

The headline counts are specific. There were 56 agency submissions in total. Of those, 41 reported individual use cases, 46 reported consolidated commercial off-the-shelf tools, and 2 agencies affirmatively said they use no AI. The 3,611 figure covers only the individually reported entries, at every stage of development.

That last detail changes how you read the number. Widely used commercial tools sit in a separate file, so the 3,611 is mostly custom, mission-specific work. It is closer to a portfolio of projects than to a software asset register.

What counts as a use case

Each agency decides where one use case ends and the next begins. One department may log a document classification model as a single entry. Another may log the same pattern five times, once per bureau. Nothing in the public summary normalises this, so a raw count measures reporting granularity as much as adoption.

This is the same problem a count of AI use cases as a vanity metric creates inside a company. More entries can mean more work. They can equally mean a team learned that a long list reads as progress.

Agencies may remove a use case from the public version where publishing it would conflict with law, so security and enforcement work is underreported by design. No Department of Defense row appears in the README table we read. The inventory is a census of what agencies were willing and allowed to publish, not of everything they run.

The 2025 stage tally for federal AI deployment

The 2025 data dictionary gives every use case one of four development stages. OMB's own definitions are worth reading word for word, because the whole argument turns on the difference between the middle two.

Pre-deployment means the use case is in development or acquisition. Pilot means it has been deployed in a limited test or pilot capacity. Deployed means it is actively authorized or used to support the agency's functions or mission. Retired means it was in the prior year's inventory and has since been discontinued.

The data dictionary makes this split mandatory for every record. The public README then throws it away. Its summary table reports a single "Deployed and Piloted" column, which totals 1,818. That is the figure most coverage has repeated, and it answers a different question from the one buyers and boards ask.

How we bounded the production count

OMB also published a data standardization report. It logs every raw stage string that agencies submitted in a messy format and the clean value OMB rewrote it to. Entries submitted as "c) Deployed" with stray spacing, for example, were rewritten 629 times.

Because the log only covers strings that needed rewriting, each stage total in it is a floor, not a count. Summed, the log shows at least 878 entries rewritten to Deployed, at least 408 to Pilot, at least 460 to Pre-deployment and at least 114 to Retired.

Two subtractions follow. If at least 408 of the 1,818 are pilots, then at most 1,410 are deployed. And at least 878 are deployed. So production sits between 878 and 1,410 uses, or 24.3% to 39.0% of the inventory.

Method: floors are the sum of rewritten stage strings per clean value in OMB's 2025 standardization report. Entries that agencies submitted in an already-valid format do not appear in that log, which is why the floors sum to 1,860 rather than 3,611. The 1,818 and 3,611 totals are OMB's. Retrieved 7 November 2026.

The 50.3% figure is true and misleading at once. A pilot is, by OMB's own definition, a limited test. Counting it alongside authorized production systems inflates the apparent conversion rate by at least 408 entries, or about 11 percentage points.

My read is that roughly 1 in 3 federal AI use cases is in production, and the honest range is 1 in 4 to 2 in 5. That is still a substantial number. It is just not half.

Government AI deployment by agency

The inventory is concentrated. The ten largest reporters hold 2,892 of the 3,611 entries, which is 80.1% of every federal AI use case listed. Their combined piloted-or-deployed share is 49.4%, close to the government-wide figure.

Counts from the OMB 2025 README table as of 13 April 2026. Shares and the "all other" row are Zan Digital calculations against OMB's stated totals of 3,611, 1,818 and 445.

The agencies that report 100%

Commerce lists 223 use cases and reports all 223 as piloted or deployed. Education reports 56 of 56 and the Tennessee Valley Authority 59 of 59. An organisation with no work in development at all is unusual, so I read these as a reporting choice: list what runs, skip what is being built.

That choice matters for the national figure. Remove those 3 agencies and their 338 entries, and the government-wide piloted-or-deployed share falls from 50.3% to 45.2%. A small number of reporting habits move the headline by 5 points.

NASA's 11% is probably the honest number

NASA reports 425 use cases, second only to HHS, and only 47 are piloted or deployed. That is 11.1%. It is tempting to read this as a delivery failure. I think it is the opposite: an agency that lists research and early-stage work alongside live systems is showing you its real funnel.

Small agencies show both patterns. The Social Security Administration reports 31 of 33 in use, and the Small Business Administration 4 of 34. Treat any single agency's share as a fact about its reporting first and its delivery second.

Of the 445 high-impact use cases, Veterans Affairs reports 215 and Justice 114. Together that is 73.9% of all high-impact federal AI. HHS, the largest reporter, lists just 2. The high-impact label carries extra risk-management duties, so this column also reflects how willing each department was to apply it.

How the 2024 and 2025 stage labels differ

You cannot simply compare the 2025 stage mix with 2024's. The labels changed. The 2024 data dictionary used five system development life cycle stages: Initiated, Acquisition and/or Development, Implementation and Assessment, Operation and Maintenance, and Retired.

The 2025 dictionary uses four: Pre-deployment, Pilot, Deployed and Retired. Most of the mapping is obvious. Implementation and Assessment is not, because it could mean a pilot or a system finishing its authorization. No public crosswalk settles it.

OMB's cleanup log shows a handful of 2025 submissions still using the old vocabulary. One "Operation and Maintenance" entry became Deployed, which is reasonable. One "Acquisition and/or Development" entry became Pre-deployment, also reasonable. One "Initiated" entry became Pilot, which I would dispute, since initiated means work has begun, not that anything is in test.

It is one record, so the effect on the totals is trivial. It does show that stage assignment involves judgement at the centre as well as at the agency.

The one clean 2024 benchmark

HHS published its own 2024 breakdown. Of 271 HHS use cases in 2024, 59 were Initiated, 57 in Acquisition and/or Development, 35 in Implementation and Assessment, 104 in Operation and Maintenance and 16 Retired. That puts 38.4% of HHS AI in operation, which lands inside our 24% to 39% band for the whole government a year later.

The consolidated count rose from 710 in 2023 to 2,133 in 2024 and 3,611 in 2025. The 2025 total is 69% above 2024. You will also see 105% quoted, which uses the first 2024 release of 1,757 use cases from 37 agencies. The 2024 repository later settled at 2,133 across 41 submissions. Use the settled number.

Federal AI pilots to production versus private-sector surveys

The reason this dataset matters is that private-sector conversion figures are all surveys. They ask executives to recall or estimate. The federal inventory is a record-level census with a stage on every line, published with its data.

The two most quoted private figures point the same way. IDC research done with Lenovo found that for every 33 AI proofs of concept a company launched, only 4 reached production. S&P Global Market Intelligence found the average organisation scrapped 46% of AI proofs of concept before production, and 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier.

The last row in this table does not mean 54% reached production. Proofs of concept that were neither scrapped nor shipped are still in flight. The rows measure different things, which is the point of the next two sections.

Why the denominators differ

A private survey's denominator is every proof of concept started. The federal denominator is every use case alive enough to appear on a list at the inventory date. A pilot that started and died within one reporting year may never appear at all, and Retired only covers entries that sat in the prior year's inventory.

So the federal list is a survivor population. Survivors convert better than starters by construction. Comparing 24% to 39% with IDC's 12% and concluding the government is three times better at shipping AI would be a mistake.

It can support one narrower claim. Among AI projects that an organisation is prepared to name in public, roughly a third are in production use. That is a useful prior when a vendor tells you most of its customers are "in production", and it sits well above the 5% success figure behind the MIT NANDA pilot failure sample, whose own sample has problems of its own.

Survey design is the variable doing most of the work here. The methodology gap behind AI ROI surveys explains why two credible studies of the same market can land 40 points apart.

Where this tally is weakest

The bounds in this post are only as good as three inputs, and each has a soft spot. I would rather list them than have you find them.

The floors depend on a cleanup log

The 878 and 408 floors come from OMB's record of rewritten strings, read through the published report rather than tabulated row by row from the CSV. If the log includes rewrites from rows later dropped from the final file, the floors would be slightly too high. A direct count of the 2025 CSV would replace both bounds with exact numbers, and it is the obvious next step for anyone with an afternoon.

There is also a one-record discrepancy. The agency rows in the README, as we transcribed them, sum to 1,819 deployed or piloted, not the stated 1,818. It does not change any percentage at one decimal place, but it is there.

Self-reporting sets the stage

Every stage is chosen by the agency that owns the system. The 100% agencies and NASA's 11% show how far reporting habits can swing a share. Pilot at one agency may be production at another. No auditor checks the label before it is published.

The counter-case: production may be undercounted

The strongest argument against this post is that the 3,611 excludes the consolidated commercial tools reported by 46 agency submissions. Those are almost all in live use. Add them, and the share of federal AI in production would rise, possibly well above 39%. I have not tried to count them, because they are reported at a different grain from individual use cases.

What the OMB AI inventory should publish next

The data already exists. Every record carries a stage. The fix is a reporting choice, and it costs OMB almost nothing.

First, split the README column into Deployed and Pilot. That single change turns the most-quoted number in federal AI from an ambiguous 50.3% into a real production rate. I think the merged column is the most damaging design decision in an otherwise unusually transparent dataset.

Second, carry stable identifiers across years, so a use case can be tracked from Pre-deployment to Pilot to Deployed or Retired. Transition rates are what a production funnel actually is. A stage snapshot is only a photograph of it.

Third, record the date each use case entered its current stage. Time in pilot is the most useful early warning a portfolio owner has, and it is the same signal that good kill criteria for AI projects are built around. The ledger of canceled agentic AI projects shows how rarely private companies disclose this either.

Frequently asked questions

Where to start if you run an AI portfolio

Borrow the federal schema for your own list this week. Give every AI project one of four stages, pre-deployment, pilot, deployed or retired, and report pilots and production in separate columns from day one. If your current dashboard merges them, your conversion rate is overstated in exactly the way OMB's is.

Then add one field the government does not have: the date each project entered its current stage. Anything that has sat in pilot for more than two quarters needs a ship-or-stop decision, and the production readiness tests for AI proofs of concept give you the checklist for making it.

Related reading