The United States federal government ran the AI workforce replacement experiment that private companies are still planning. It did not happen quietly inside a tech company with a press release and a stock surge. It happened at a federal agency managing government buildings and procurement contracts across the country, under the direction of DOGE, and the results are now documented well enough to evaluate. The General Services Administration lost nearly 40% of its total workforce. Then it deployed an AI chatbot to handle the work those employees used to do. Staff who remained told Wired the chatbot performs “about as good as an intern.”
That gap, between the scale of the workforce reduction and the capability of the tool deployed to cover it, is the story. It is not unique to government. As we covered in our analysis of Klarna’s AI experiment, the pattern of replacing workers with AI tools that handle volume but fail at complexity has now been documented at scale in both private and public sectors. The GSA case is different in one significant respect: the disruption was faster, less reversible, and the workers affected had no severance negotiation or market alternatives of the kind private sector employees at least theoretically have.
What DOGE Did at the GSA
The General Services Administration manages the federal government’s real estate portfolio and most of its procurement contracts. It is not a glamorous agency but it is an essential one: when federal agencies need office space, vehicles, or supplies, GSA is the mechanism through which those things are acquired and managed.
Beginning in late 2024, DOGE targeted GSA for significant workforce reductions as part of its broader effort to cut the size of the federal government. The cuts were substantial and rapid. According to the Government Accountability Office, the Public Buildings Service, the division of GSA responsible for federal properties, lost approximately 45% of its employees between September 2024 and November 2025. Across the agency as a whole, Federal News Network reported that nearly 40% of the total workforce was gone within that period.
The GAO noted in its assessment that the deep staffing cuts at the Public Buildings Service had “created challenges in accessing and preparing properties for sale,” which was one of DOGE’s stated goals. In other words, the agency tasked with selling federal property could no longer easily process the sales because the people who knew how to do it had been cut.
The Chatbot Deployed to Fill the Gap
Alongside the workforce reductions, DOGE deployed a custom AI tool called GSAi to approximately 1,500 GSA employees. Computerworld reported that the tool was built by combining Anthropic’s Claude Haiku 3.5, Claude Sonnet 3.5, and Meta’s Llama 3.2. Staff were told they could use it to draft emails, create talking points, summarise documents, and write code.
The instruction to employees came with a significant caveat: do not enter any sensitive or confidential government information into the system. For an agency that handles federal procurement contracts, property records, and interagency financial transactions, that restriction covers a substantial portion of the actual work. The chatbot was deployed to handle government work but could not be trusted with government data.
The performance assessment from staff was not encouraging. A government employee told Wired that the tool’s output was “about as good as an intern,” producing “generic and guessable answers.” Inc reported on the Wired findings and noted that the tool’s rushed deployment raised security concerns about whether its architecture was adequately protected against data leaks or external compromise.
The Million Hours Challenge
Despite the early performance assessments, GSA has announced an internal initiative it is calling the “million hours challenge.” The goal, as described by GSA Deputy Director Michael Lynch and reported by Federal News Network, is to identify one million work hours across the agency that can be automated using AI tools. GSA says it has already identified approximately 400,000 hours of automatable work, roughly 40% of the target.
To put that number in concrete terms: one million work hours is approximately what 500 full-time employees working standard eight-hour days would produce in a year. GSA is aiming to replace the equivalent of 500 roles with automation, in an agency that has already cut nearly 40% of its headcount. The agency’s stated framework is “eliminate, optimise, automate” and it intends to apply the model to other federal agencies if the internal pilot is considered successful.
What the Evidence Actually Shows
The GSA case is an unusually transparent view into what AI workforce replacement looks like when it is deployed rapidly and at scale, rather than in the controlled pilots and press release announcements that dominate most coverage of this topic.
Three things are documented. First, the workforce cuts created immediate operational gaps. The GAO found that the Public Buildings Service could no longer efficiently perform core functions after losing 45% of its staff. The institutional knowledge, process familiarity, and relationship context that experienced employees carry did not transfer to the AI tool deployed to replace them. Second, the AI tool that was deployed could not handle sensitive data, which for a government agency means a large share of its actual workload. Third, the employees who remained assessed the tool’s output as entry-level at best.
This is consistent with the pattern documented at Klarna, where AI handled high-volume routine queries but failed at the complex, contextual, and emotionally sensitive interactions that determined customer satisfaction. It is consistent with what the Stanford 2026 AI Index found: AI performance on structured benchmarks has improved dramatically, but the gap between benchmark performance and real-world institutional complexity remains significant. As we covered in our analysis of the Stanford findings, the productivity gains are real in bounded, structured tasks and much less reliable everywhere else.
Why This Case Is Different From Private Sector Experiments
When Klarna replaced 700 customer service workers and later acknowledged the experiment had gone too far, the company had the option to hire people back. It did. When Block cut 4,000 staff and the stock surged 24%, those workers entered a private labour market where, in principle, equivalent roles exist elsewhere. The feedback loops, however imperfect, exist.
Federal workers cut under DOGE do not have the same options. Civil service roles do not have direct private sector equivalents. The institutional knowledge they carried, how to manage federal property transactions, how to navigate procurement regulations, how to maintain relationships between agencies, was not documented in a way that an AI tool can access. When it is gone, it is gone. The GSA is now reportedly resuming limited hiring after realising the depth of the operational gaps the cuts created, but rebuilding institutional capacity takes years, not quarters.
The broader implication is that the fastest and most aggressive AI workforce replacement experiments are being run not in tech companies with venture capital and quarterly earnings flexibility, but in public institutions where the cost of failure lands on the people those institutions are supposed to serve. A federal agency that cannot efficiently manage property sales or procurement contracts is not a failed stock. It is a broken public service.
The “million hours challenge” will likely reach its target, at least on paper. The question worth tracking is whether the hours automated are the hours that mattered, or the hours that were easiest to count.
This article draws on reporting by Federal News Network, Computerworld, Inc, and the Government Accountability Office. Analysis and interpretation reflect the author’s reading of publicly available information.