How do government employees use AI sandboxes?
In California, CDT manages the state's AI sandboxes and supports other institutions with subject matter experts in security, data, infrastructure and architecture. Agents share responsibility with state groups seeking generative AI innovation and offices of data and innovation, Porat says.
“ODI conducts research and testing user experiences on proof-of-concept, and the innovation department offers staff members the use and evaluation of proof-of-concept solutions. Each solution is different, and the department works with vendors to use the solution,” he says.
New Jersey also established the AI Sandbox in 2024. The state strengthened its tools in early 2025 and then looked into how employees are using them. Here are the top five ways state employees use the tool, according to Dave Cole of New Jersey.
- Drafting and editing: Enhanced email, notes and reports for clarity and professionalism
- Document Summary: Condensing long text and complex information for quick understanding
- Calibration: Make sure communication is refined and error-free
- Brainstorming and Idea Generation: Helping to create and refine content and projects
- Technical support: assisting with coding, data analysis, and creating presentation overviews, streamlining workflows, and increasing efficiency.
Related: Synthetic data is a valuable resource for testing AI use cases.
What projects does the state government support with AI applications?
In California, AI Sandbox supports projects that include Caltrans traffic management insights and its vulnerable road user safety assessments, the Department of Tax and Fee Management call center productivity, and the state's health and human resources agency programs for testing state health and medical facilities.
State agencies are working with the New Jersey Department of Innovation on several large, ongoing projects using AI sandboxes, Cole adds:
- Call center menu redesign. “For example, we analyzed calls to generate better self-service menu options for property tax questions, and we've increased calls successfully resolved without talking to agents by 50%,” says Cole.
- Analysis of public feedback data. The state worked on a backlog of over 430,000 ratings, 70,000 comments and 22,000 emails at the bottom of each page, from the public to online widgets, allowing civil servants to quickly identify common issues.
- Draft emails and copies of the website to improve clarity. “We worked with the Department of Labor to help rewrite emails in plain language, bringing 35% faster responses,” Cole says.
- The team created Spanish terms related to unemployment insurance and used them to translate documents and forms. These translations were subsequently reviewed by humans.
“Of course, the team reviews anything the AI assistant creates, but the amount of things saved you time and effort at the beginning is pretty much,” Cole says. “The result is new Spanish material that increases understanding, reduces the completion time of the form and drives towards a more equitable experience.”
How does the state government build an AI sandbox?
To build an AI sandbox, both California and New Jersey have been transformed into clouds.
“The sandbox is designed for publicly-exposed insensitive data while protecting privacy, allowing realistic testing environments to explore new AI tools without putting sensitive information at risk. Apart from real-world state systems, these cloud-based environments allow teams to protect privacy and experiment at their own cost,” says Porat.
The sandbox is specifically built for the California government's testing environment, in compliance with security, governance and transparency instructions, Porat says.
Similarly, New Jersey has built an AI sandbox with strong privacy protection and cost-effective measures. This is, first and foremost, a sandbox, an isolated, controlled environment that can be used for experiments hosted on the state's IT infrastructure.
“The tool avoids risking internal information and ensures that state data is not used to train third-party AI models. The tool also has several filters to prevent the output of potentially harmful information, and also provides protection to prevent attempts at protection.
How will the state government protect AI sandboxes?
CDT provides an isolated cloud environment for certified staff, so there is no connection to state systems. The agency configures the cloud environment according to the State Information Management Manual 140, Cloud Security Guide. All sandbox users must adhere to the guidance of SIMM 140.
CDT also assigns a senior cybersecurity administrator to each sandbox to monitor its internal activities and work with innovative departments and vendor security teams.
New Jersey also follows cybersecurity principles when securing sandboxes, with a focus on training.
“We have released a training course for state employees who have access to the tool itself, ensuring that the tool is safe, secure and used in a way that respects the privacy of its residents. It is unbiased and in line with state policies. More than 25 states and regions use this training in particular.”
Approximately 20% of New Jersey employees use AI assistants. This was encouraged over 500,000 times in the first year, with a satisfaction rating of over 80%. The cost was around $1 per month, but with ready-made commercial licenses, it saves millions of dollars at around $20 per month per user.
