Elon Musk’s Department of Government Efficiency (DOGE) is pushing to rapidly develop “GSAi,” a custom generative AI chatbot for the US General Services Administration, according to two people familiar with the project. The plan is part of President Donald Trump’s AI-first agenda to modernize the federal government with advanced technology.

One goal of the initiative, which hasn’t been previously reported, is to boost the day-to-day productivity of the GSA’s roughly 12,000 employees, who are tasked with managing office buildings, contracts, and IT infrastructure across the federal government, according to the two people. Musk’s team also seemingly hopes to use the chatbot and other AI tools to analyze huge swaths of contract and procurement data, one of them says. Both people were granted anonymity because they aren’t authorized to speak publicly about the agency’s operations.

Thomas Shedd, a former Tesla employee who now runs Technology Transformation Services, the technology arm of the GSA, alluded to the project in a meeting on Wednesday. “Another [project] I’m trying to work on is a centralized place for contracts so we can run analysis on them,” he said, according to an audio recording obtained by WIRED. “This is not new at all—this is something that’s been in motion before we started. The thing that’s different is potentially building that whole system in-house and building it very quickly. This goes back to this, ‘How do we understand how the government is spending money?’”

The decision to develop a custom chatbot follows discussions between the GSA and Google about its Gemini offering, according to one of the people.

While chatbots such as ChatGPT and Gemini have been adopted across corporate America for tasks like writing emails and generating images, executive orders and other guidance issued during the Biden administration generally instructed government staff to be cautious about adopting emerging technologies. President Donald Trump has taken a different approach, ordering his lieutenants to strip away any barriers to the US exerting “global AI dominance.” Heeding that demand, Musk’s government efficiency team has moved swiftly in recent weeks to bring aboard more AI tools, according to reports published by WIRED and other media.

Overall, the Trump administration may be engaging in the most chaotic upheavalof the federal bureaucracy in the modern computer era. Some Trump supporters have celebrated the changes, but federal employees, labor unions, Democrats in Congress, and civil society groups have heavily criticized them, arguing in some cases they may be unconstitutional. While DOGE hasn’t publicly changed its stance, the team quietly halted the rollout of at least one generative AI tool this week, according to two people familiar with the project.

The White House did not immediately respond to a request for comment.

For the past few weeks, Musk’s team has been working to swiftly cut costs across the US government, which has seen its annual deficit increase for the last three years. The Office of Personnel Management, which acts as the HR department for the government and is stacked with Musk loyalists, has encouraged federal employees to resign if they cannot return to the office five days a week and commit to a culture of loyalty and excellence.

DOGE’s AI initiatives dovetail with the group’s efforts to reduce the federal budget and speed up existing processes. For instance, DOGE members at the Department of Education are reportedly using AI tools to analyze spending and programs, The Washington Post reported on Thursday. A department spokesperson says that the focus is on finding cost efficiencies.

The General Services Administration’s GSAi chatbot project could bring similar benefits, enabling workers, as an example, to draft memos faster. The agency had hoped to use existing software such as Google Gemini, but ultimately determined that it wouldn’t provide the level of data DOGE desired, according to one of the people familiar with the project. Google spokesperson Jose Castañeda declined to comment.

It’s not the only DOGE AI ambition that hasn’t panned out. On Monday, Shedd described deploying “AI coding agents” as among the agency’s top priorities, according to remarks described to WIRED. These agents help engineers automatically generate, edit, and answer questions about software code in hopes of boosting productivity and reducing errors. One tool the team looked into, according to documents viewed by WIRED, was Cursor, a coding assistant developed by Anysphere, a fast-growing San Francisco startup.

Anysphere’s leading investors include Thrive Capital and Andreessen Horowitz—both of which have connections to Trump. Joshua Kushner, Thrive’s managing partner, has historically made political campaign donations to Democrats, but he is the brother of Trump’s son-in-law, Jared Kushner. Andreessen cofounder Marc Andreessen has said he’s advising Trump on tech and energy policy.

A different person familiar with the General Services Administration’s technology purchases says the IT team at the agency had initially approved the use of Cursor, only to retract it later for further review. Now, DOGE is pushing to install Microsoft’s GitHub Copilot, the world’s most well-known coding assistant, according to the other person familiar with the agency.

Cursor and the General Services Administration did not respond to requests for comment. Andreessen Horowitz and Thrive declined to comment.

Federal regulations require avoiding even the appearance of a conflict of interest in the choice of suppliers. And while there haven’t been any known widespread concerns about Cursor’s security, federal agencies are generally required by law to study potential cybersecurity risks before adopting new technology.

The federal government’s interest and use of AI isn’t new. In October 2023, then president Biden ordered the General Services Administration to prioritize security reviews for several categories of AI tools, including chatbots and coding assistants. But by the end of his term, none had made it through even the preliminary agency review processes, according to a former official familiar with them. As a result, no dedicated AI-assisted coding tools have received authorization under FedRAMP, a GSA program to centralize security reviews and ease the burden on individual agencies.

Though the Biden prioritization process didn’t bear fruit, several individual agencies have explored licensing AI software. In transparency reports published during Biden’s term in office, the Commerce, Homeland Security, Interior, State, and Veterans Affairs departments all reported they were pursuing the use of AI coding tools, including in some cases GitHub Copilot and Google’s Gemini. GSA itself had been exploring three limited-purpose chatbots, including for handling IT service requests.

Guidance from the personnel office issued under then president Biden stated that the efficiency gains of AI coding agents should be balanced against potential risks such as introducing security vulnerabilities, costly errors, or malicious code. Historically, the heads of federal agencies have been left to develop their own policies for using emerging technologies. “Sometimes doing nothing is not an option and you have to accept a lot of risk,” says a former government official familiar with these processes.

But they and another former official say that agency administrators generally prefer to conduct at least preliminary security reviews before deploying new tools, which explains why it sometimes takes a while for the government to adopt new technology. That is one reason why just five big companies, led by Microsoft, accounted for 63 percent of government spending on software across agencies surveyed by government researchers at the Government Accountability Office for a report to lawmakers last year.

Undergoing government reviews can require companies to invest significant time and staff—resources startups may not have. That may have been one challenge affecting Cursor’s ability to win business from the recent DOGE push, as the startup didn't have immediate plans to achieve FedRAMP authorization, according to one of the people familiar with the GSA’s interest in the tool.

Additional reporting by Dell Cameron, Andy Greenberg, Makena Kelly, Kate Knibbs, and Aarian Marshall