How Are Government Employees Using AI Sandboxes?
In California, CDT administers the state’s AI sandbox and supports other agencies with subject matter experts in security, data, infrastructure and architecture. The agency shares responsibilities with a state group seeking generative AI innovation and the Office of Data and Innovation, Porat says.
“ODI conducts user experience research and testing on the proof of concept, and the innovating department provides staff to use and evaluate the proof-of-concept solution. Each solution is different, and departments work with vendors to use the solution,” he says.
New Jersey also established an AI sandbox in 2024. The state enhanced the tool early in 2025 and then surveyed its employees as to how they were using it. Here are the top five ways that state employees say they are using the tool, according to New Jersey’s Dave Cole:
- Drafting and editing: Enhancing emails, memos and reports for clarity and professionalism
- Summarizing documents: Condensing lengthy texts and complex information for quick comprehension
- Proofreading: Ensuring communication is polished and error-free
- Brainstorming and idea generation: Aiding in the creation and refinement of content and projects
- Technical support: Assisting with coding, data analysis and creating presentation outlines, streamlining workflows, and boosting efficiency.
RELATED: Synthetic data is a valuable resource for testing AI use cases.
What Projects Do State Governments Support With AI Applications?
In California, the AI sandbox supports projects including Caltrans’ Traffic Management Insights and its Vulnerable Road User Safety Assessment, the Department of Tax and Fee Administration’s call center productivity, and programs from the state’s Health and Human Services Agency for language access and health care facility inspections.
State agencies work with the New Jersey Office of Innovation on several large ongoing projects that use AI sandboxes, Cole adds:
- Redesigning call center menus. “For example, we analyzed calls to generate better self-service menu options for property tax questions, leading to a 50% increase in successfully resolved calls without speaking to an agent,” Cole says.
- Analyzing public feedback data. The state tackled a backlog of over 430,000 ratings, 70,000 comments and 22,000 emails from the public to online widgets at the bottom of each page, allowing officials to quickly identify common issues.
- Drafting email and website copy to improve clarity. “We worked with the Department of Labor to help rewrite emails in plain language, resulting in 35% faster responses,” Cole says.
- A team created a glossary of Spanish terms related to unemployment insurance and used that to help translate documents and forms. Those translations then received a human review.
“Of course, the team reviews anything that the AI assistant creates, but the time and effort saved doing the initial work is substantial,” Cole says. “As a result, we have new Spanish-language materials that are enhancing comprehension, reducing form-completion times and driving toward more equitable experiences.”
How Do State Governments Build AI Sandboxes?
To build their AI sandboxes, California and New Jersey both turned to the cloud.
“The sandboxes are designed for publicly available, nonsensitive data to protect privacy while allowing for a realistic testing environment to explore new AI tools without putting sensitive information at risk. These cloud-based environments are separate from actual state systems and allow teams to experiment freely while protecting privacy and keeping costs down,” Porat says.
The sandboxes were built specifically for California's government testing environment with compliance to security, governance and transparency directives, Porat says.
In a similar manner, New Jersey built its AI sandbox with strong privacy protections and cost-effective measures. It is first and foremost a sandbox, an isolated and controlled environment available for experimentation, hosted on the state’s IT infrastructure.
“The tool ensures that no state data is used to train third-party AI models, avoiding putting internal information at risk. The tool also has several filters to prevent input or output of potentially harmful information, as well as protections to prevent attempts to circumvent safeguards. Importantly, the Office of Innovation had invested in our team and capacity prior to all of this, so we could make the transition to providing a tool for all state employees in a matter of weeks,” Cole says.
How Do State Governments Secure AI Sandboxes?
CDT provides authorized personnel with an isolated cloud environment, so there are no connections to state systems. The agency configures the cloud environment in accordance with the State Information Management Manual 140, Cloud Security Guide. Everyone who uses the sandbox is required to adhere to the guidance of SIMM 140.
CDT also assigns a senior cybersecurity administrator to each sandbox to monitor activity within it and work with the security teams of an innovating department and vendor.
New Jersey also follows cybersecurity principles in securing its sandbox while placing an emphasis on training.
“We released a training course for state employees, accessible on the tool itself. This helps ensure that the tool is used in ways that are safe, secure and respectful of residents’ privacy; free of bias; and in line with state policies. Notably, more than 25 states and localities now use this training,” Cole says.
About 20% of New Jersey state employees have used the AI assistant, which has been prompted over a half a million times in its first year, with a satisfaction rating of over 80%, Cole says. The cost has been about $1 per user per month versus about $20 per user per month for off-the-shelf commercial licenses, saving millions of dollars, he adds.