IRS Had 126 AI Use Cases. Its Technical Workforce Fell 28%.

The Internal Revenue Service had 126 active artificial-intelligence use cases in its inventory in June 2025. Most were not finished products. Seventy-seven were still in development, and 49 were in operation.

At the same time, the workforce most likely to supply the agency's technical and analytical capacity was becoming smaller. FederalHiringData identified nine Office of Personnel Management occupation series covering information technology, computer engineering, computer science, data science, statistics, mathematical statistics, operations research, economics and mathematics. At IRS, that selected workforce fell from 8,691 employees in January 2025 to 6,232 in June 2026, a reduction of 2,459 employees, or 28.3%.

That is a workforce proxy, not an official count of IRS employees assigned to AI. OPM occupation codes do not reveal a person's project, proficiency or daily duties. The proxy does establish something the agency's AI strategy cannot avoid: the pool of federal IT and analytical employees was much smaller by mid-2026, while IRS still intended to expand its use of AI.

The loss was not only visible in broad occupation data. The Government Accountability Office reported that IRS's Research, Applied Analytics and Statistics organization lost 63 employees who had supported AI work, about 10% of that unit's January 2025 staff. The agency's AI governance team lost more than three-quarters of its staff, including contractor support. A Chief Technology Officer team that had supported 11 AI use cases paused or suspended work on 10 after losing staff and two supporting contracts.

This is the central constraint on IRS's AI ambitions. Artificial intelligence may automate a task, rank a workload or help answer a question. Building and governing it still requires people who understand data, tax administration, software, contracting, privacy, security and the operational consequences of a model's errors.

The portfolio was large, but most projects were not operational

The 126-use-case count describes applications of AI to a particular business need, not 126 mature systems producing verified benefits. GAO classified 77 as being in initiation, acquisition, development, implementation or assessment. The remaining 49 had reached operation and maintenance.

IRS placed 59 use cases in operational efficiency, 43 in tax compliance and fraud detection, and 24 in taxpayer services. The mix differed sharply by stage. Only 14 of the 59 operational-efficiency projects were in operation. Taxpayer services had the smallest portfolio, but 17 of its 24 use cases were operating.

The operational-efficiency category included automation, internal chatbots, data insights, IT modernization, IT security, digitization and workload prioritization. Some examples were relatively bounded: meeting summaries, transcription, translation, document review or an internal tool that checked contract records for policy and transparency requirements. Other projects reached closer to mission decisions, including research and models for selecting returns for audit.

GAO's subcategory counts show how much of this work remained unfinished. Automation was the largest group, with 15 use cases in development and three in operation. IRS had six internal-chatbot projects in development and three operating, while data-insight projects had the same six-to-three split. Only one of seven IT-modernization projects and one of four workload-prioritization projects was operating. None of four digitization projects had reached operation.

Those categories span very different levels of risk. A meeting-summary tool and an audit-selection model both use AI, but they do not require the same testing, legal review or monitoring. A portfolio count that treats them as equivalent can describe volume while concealing the effort needed to move each one safely into operation. That distinction matters when staffing plans assign scarce specialists across the portfolio.

Development also creates work before any productivity gain appears. Teams must define the problem, obtain and prepare data, test performance, document limitations, integrate the system, secure it and decide whether the benefit justifies the cost. Once deployed, a model must be monitored as tax law, taxpayer behavior and source data change. That life-cycle labor helps explain why losing technical and governance employees can affect projects even when the agency still owns the software or contract.

Taxpayer-service AI had the clearest public scale. IRS's inventory included 11 voice bots and two chatbots that answered questions on different tax topics. GAO reported more than 5 million taxpayer interactions with public-facing bots during the 2024 filing season and nearly 25 million during the 2025 filing season. That is evidence of use, not proof that every interaction was resolved correctly or that a bot caused an improvement in service.

The definition of active also needs care. Fourteen use cases were temporarily paused in June 2025, and nine others were permanently retired. The active count included the paused projects and excluded the retired ones. Sixty-five of the 126 were too sensitive for public reporting or were research-and-development efforts exempt from public disclosure, limiting independent review of the full portfolio.

A long technical build reversed after January 2025

The IRS technical workforce did not begin with generative AI. The agency has used analytical methods for decades and was an early federal user of large-scale automated data processing. Federal occupation classifications changed along the way, so a long historical comparison requires a bridge.

FederalHiringData combined the old series 0334 Computer Specialist with series 2210 Information Technology Management. The 0334 classification contained 4,512 IRS employees in September 2000 and 4,889 in September 2001. Series 2210 replaced it during the early 2000s; in September 2002, IRS had 215 employees still coded 0334 and 4,732 coded 2210. Treating the two as one continuous IT category prevents the classification change from looking like a sudden hiring surge.

The analysis adds selected analytical occupations: economists, computer engineers, operations-research analysts, mathematicians, mathematical statisticians, statisticians, computer scientists and data scientists. In September 2000, those IT and analytical series contained 4,873 IRS employees. The total rose unevenly to 6,778 in September 2014 and 8,439 in September 2024. It reached 8,691 in January 2025.

By June 2026, the selected count was 6,232. That was still above the 2000 level, but the 18-month decline erased much of the technical expansion recorded since 2011.

Information Technology Management accounted for most of the loss because it dominated the selected workforce. Series 2210 fell from 7,960 employees in January 2025 to 5,705 in June 2026, a decline of 2,255, or 28.3%.

Every other selected series with employees in both snapshots also became smaller. Operations research fell from 153 to 106; economics from 247 to 200; statistics from 115 to 76; mathematical statistics from 65 to 44; computer science from 58 to 39; computer engineering from 65 to 47; and data science from 28 to 15.

Those categories are intentionally broader than AI. A series 2210 employee might secure a network, support a legacy application or manage infrastructure without working on a model. An economist or statistician might perform tax analysis unrelated to AI. Conversely, people in other occupations and contractors may contribute directly to AI projects. The chart measures technical capacity surrounding an AI portfolio, not an AI headcount.

Personnel flows show how difficult rapid replacement would be

OPM's public personnel-action files reinforce the direction of the headcount series. Across the selected IRS occupations, they record 110 accessions and 2,160 separations during 2025. Through June 2026, they record seven accessions and 126 separations.

Personnel actions are not necessarily unique people. A person can generate more than one action, and the public files do not identify project assignments. The totals should not be read as an exact replacement ratio. They do show that recorded outflow greatly exceeded inflow in the same set of occupations.

IRS's own account to GAO explains why replacing specialized employees was not a simple hiring exercise. RAAS officials said a majority of the 63 employees they lost had directly designed, developed or overseen AI projects. More than a dozen had researched analytical methods, tools and infrastructure. Some probationary employees were later rehired, but officials said others had found work elsewhere, and some returning or remaining employees were reassigned away from AI because of attrition in other RAAS work.

GAO reported that IRS had not identified the skills it needed to support AI or developed a plan to close the gaps. Short-term hiring plans focused on customer support and audit employees. Requests included computer engineers, computer scientists and one data scientist, but none was specifically identified as contributing to AI. The agency also could not use its data-scientist direct-hire authority under the hiring freeze described in the report.

Training existing employees can build capability, and IRS officials described growing demand for AI training and data literacy. It cannot by itself restore lost project ownership, institutional knowledge or contractor-management capacity. A useful workforce plan would need to distinguish at least four functions: people who develop models, people who validate data and outcomes, people who own tax-administration decisions, and people who independently govern risk.

Public recruiting returned from a historic low, but not to its earlier scale

The FederalHiringData historical USAJOBS archive provides a second, imperfect view of the pipeline. We counted distinct IRS announcements in the same selected occupations, using the Jan. 1 through Aug. 14 window in each year from 2018 through 2026.

The archive recorded 177 such announcements in 2018, 237 in 2019, 461 in 2021, 289 in 2023 and 501 in 2024. The count fell to 16 in 2025, then rose to 64 through Aug. 14, 2026. Of the 64 most recent announcements, 49 were tagged to series 2210 and 15 to the other selected analytical occupations.

The 2026 list included data scientists, economists, operations-research analysts, mathematical statisticians and an AI/machine-learning data-science role. It demonstrates that technical recruiting had resumed. It does not show 64 hires or 64 vacancies: one announcement can cover multiple locations or grades, an announcement may be canceled, and a selection is not visible from the posting count.

The rebound was also small relative to the same period in 2024. The 64 announcements through Aug. 14, 2026 were four times the 2025 low but 87% below the 501 recorded in 2024. That comparison cannot establish how many skilled employees IRS needs. It shows that the public recruiting signal had not returned to its earlier scale.

Readers can review current IRS openings separately. The counts in this article come from the historical archive and should not be interpreted as jobs that remain open today.

Contractor dependence makes federal capacity more important, not less

GAO estimated that about 80% of the 126 AI use cases involved contracts to some degree. Contractors can provide specialized tools, development capacity and infrastructure. They do not eliminate the need for federal staff who define requirements, understand the tax program, review performance, manage access to sensitive data and decide whether a contract should be renewed.

That oversight capacity also contracted. IRS procurement officials told GAO that their office lost more than 40% of its staff between May and June 2025. They warned of delays in new awards and modifications or renewals, including contracts supporting mission-critical programs. RAAS officials said the loss of procurement employees had already reduced institutional knowledge about technical contract details.

The CTO AI team's experience made the dependency concrete. It supported 11 use cases in January 2025. By May, only half of its staff remained and two contracts providing staffing support had not been renewed. The team paused or suspended support for 10 of the 11 projects because of resource limits.

FederalHiringData reviewed IRS-related transactions in its local USAspending contract extract. The records identify broad information-technology and professional-administrative obligations, but they do not support a defensible AI-specific dollar series. Contract descriptions can omit the technology involved, one award can support several functions, and an obligation is not a worker. This article therefore does not label broad IT spending as AI spending or convert dollars into contractor headcount.

The more defensible conclusion comes from GAO's project-level evidence: contracts touched most use cases, and federal teams had less capacity to oversee and sustain them.

The inventory itself showed the cost of reduced governance

An AI inventory is a control system. Managers cannot prioritize investments, compare outcomes or identify high-impact applications if entries do not accurately describe what a tool does, who owns it, where it operates and whether it is paused or retired.

Related research

Put this finding in context

All articles

Data investigations/August 18, 2026

IRS Workforce Fell 29% as GAO Warned of Filing-Season Risks

The IRS had 29,663 fewer covered civilian employees in June 2026 than in January 2025. GAO says workforce-planning gaps and post-filing pressures pose risks to future operations.

Data investigations/August 18, 2026

FEMA's Public Workforce Count Fell 22% as GAO Warned of Readiness Risks

FEMA had 5,793 fewer covered civilian employees in June 2026 than in January 2025. GAO says the agency cut staff without the workforce planning needed to assess mission capacity.

GAO found at least one quality issue in 62 of the 126 June 2025 entries. More than 25% did not state the expected benefit. Nearly 10% omitted status or life-cycle stage. At least 10% did not list all participating business units.

Some entries also blurred distinct deployments. One model originally used to select returns for audit research had expanded to risk-based audit selection in two other business units but remained a single inventory record. Another record combined a code-translation tool and a chatbot that answered questions about code, even though the tools operated independently.

The count was incomplete as well as inconsistent. GAO identified contracted identity-proofing AI and Criminal Investigation tools that were not listed. Forty-three use cases begun before the first public inventory in August 2022 did not enter the internal inventory until 2023 or later. Eleven took one to two years or longer to be reported.

Staffing and inventory quality were connected. IRS governance officials had introduced reviews of each entry in 2024, but suspended them in early 2025 after staffing reductions. By May, the team allowed partial submissions because it preferred an incomplete record to no record and lacked capacity for the time-intensive meetings needed to validate each use case.

IRS has since strengthened policy. Its current AI governance manual, effective Feb. 10, 2026, sets responsibilities for the Chief Data and Analytics Officer, Responsible AI Official, governance bodies and use-case owners. It requires inventory maintenance, annual validation and controls for high-impact AI. GAO also noted that IRS restored independent subject-matter review for high-impact AI while the report was being finalized.

Policy is meaningful counterevidence to the idea that oversight simply disappeared. Execution remains the test. GAO made eight recommendations covering a workforce plan, inventory quality assurance, guidance, communications, contract identification, coordination, alignment with strategic goals and performance metrics. IRS agreed with all eight. The public GAO record listed them as open when this article was checked.

AI cannot be separated from tax-administration capacity

The riskiest interpretation of an AI workforce story is that a model straightforwardly replaces a tax employee. The reviewed evidence does not support that claim.

One RAAS official told GAO that IRS might not deploy a model designed to prioritize returns for audit because the program might no longer have enough staff to conduct the audits. Officials also said fewer completed audits would produce less outcome data for training or retraining models. In that example, the model depends on the workforce it is supposed to assist: auditors create both the action and the feedback needed to evaluate selection quality.

Taxpayer service illustrates a different relationship. Voice and chat bots can absorb common questions at high volume, but interaction counts do not disclose whether callers resolved their issue, needed a human later or received a correct answer. A responsible performance measure would connect use to outcomes instead of treating sessions as success.

Operational automation may reduce document handling or help developers understand code. It still requires owners who can define the task, engineers who can integrate it with old systems, security staff who can protect tax data, procurement staff who can manage vendors and independent reviewers who can challenge the result.

IRS's 126-use-case inventory therefore describes both ambition and workload. Sixty-one percent of the portfolio was still in development. Moving those projects into reliable operation is not the removal of labor from tax administration. It is a demanding form of tax-administration labor in its own right.

Methodology and limitations

FederalHiringData analyzed OPM public workforce records for IRS agency subelement TR93. The long-run technical proxy includes occupation series 0110, 0854, 1515, 1520, 1529, 1530, 1550, 1560 and 2210. Series 0334 is combined with 2210 to bridge the federal computer-occupation classification transition. The series uses September snapshots from 2000 through 2025 and June 2026 as the latest available observation. January 2025 and June 2026 support the detailed occupation comparison.

These occupation series do not identify AI assignments, project ownership, skills, full-time-equivalent effort or contractor labor. They omit employees in other series who may support AI. Headcount is not productive hours.

Accessions and separations come from OPM public personnel-action files for 2025 and January through June 2026. They are action counts, not necessarily unique people, and are not presented as an exact replacement ratio.

The USAJOBS analysis counts distinct IRS announcement control numbers tagged to the selected occupation series and opened from Jan. 1 through Aug. 14 in each year. FederalHiringData historical coverage begins in March 2017, so the same-window chart starts in 2018. Announcements are not hires, vacancies or selections.

AI portfolio, staffing, governance, inventory-quality and recommendation findings come from GAO-26-107522, released March 24, 2026. GAO's inventory snapshot is June 2025 and is not a current count of every IRS AI system. The current IRS manual and other official sources were checked through Aug. 24, 2026.

The analysis does not establish that workforce reductions caused a particular model failure, revenue result, audit rate or taxpayer-service outcome. It does not establish that AI replaced revenue agents, auditors or service representatives. For the agencywide workforce and filing-season context, see FederalHiringData's earlier investigation of IRS staffing and filing-season risk. Readers can also browse federal workforce statistics, the historical jobs archive, the agency directory and more data investigations.