San Francisco granted 30,000 city employees access to Microsoft Copilot in July 2025, and in the months since, more than 1 million messages have been sent or received through the AI chatbot. But adoption is sharply divided: while some departments and workers have become heavy users, many others rarely touch the tool, and overall uptake has hovered around 50%.
Adoption splits along desk work vs. field work
Departments with disproportionate numbers of white-collar staff - lawyers, political aides, analysts, and managers - recorded the highest usage rates. The Department of Police Accountability had the highest Copilot use rate among major departments since December. In contrast, employees in physically intensive roles, such as firefighters and police officers, consistently showed the lowest AI use.
The mayor's office itself saw an early spike in adoption last summer, likely driven by the mayor's personal advocacy for the technology. But that share has since fallen. Across the city, a small number of super-users are driving a large portion of interactions. One collections officer at the port, for example, logged more than 1,200 interactions between December and April - roughly 10 per day.
How departments are using Copilot
The Department of Public Health, the city's largest, had the highest raw number of interactions. A department statement said use cases include drafting communications, summarizing documents, generating presentations, and analysis. At the Department of Police Accountability, Chief of Staff Marshall Khine said AI helps with large-scale data analysis, including reviewing body camera footage, combing through records, and providing summary transcriptions.
"It's really beneficial for AI to be a tool to focus where to look," Khine said. "But we've reinforced with our folks that AI is not a replacement for their investigations, that they still need to review all the materials."
Guidelines put humans in charge
San Francisco's AI rules, updated last July, require employees to review and fact-check all AI-generated content, especially for public-facing or sensitive work. The guidelines prohibit creating official city documents without expert human review, generating deepfakes or impersonations of officials, fabricating survey respondents, and using AI to review legal or regulatory issues.
Jane Gong, the city's director of emerging technologies, said in a December interview, "Our top reminder for everyone is at the end of the day, no matter what you put out as a city employee, you are responsible. Whether it's AI or not, you always have to check your work." The city is now developing role-specific AI training and refining its policies. A pilot program that gave 2,000 staff access to ChatGPT found that 70% of respondents saved up to five hours a week. For agencies adopting similar tools, Microsoft Copilot Training can help teams build safe and effective habits.
Why this matters for government employees
San Francisco's experience shows that simply giving employees access to AI does not guarantee broad adoption - even in a tech hub. The gap between desk workers and field staff, the reliance on super-users, and the need for ongoing guidance all point to a larger truth: successful AI deployment in government requires deliberate change management, clear guardrails, and role-specific training. For public-sector professionals looking to stay ahead, AI for Government Courses offer structured ways to build those skills without waiting for agency-wide programs. The number of AI interactions in San Francisco rose over 200% between July 2025 and April 2026, suggesting that while not everyone is using Copilot, those who do are using it more and more.
Your membership also unlocks:
