I’m thrilled to sit down with Donald Gainsborough, a political savant and leader in policy and legislation, who heads Government Curated. With his deep expertise in government innovation, Donald has been instrumental in driving New Jersey’s pioneering efforts in adopting artificial intelligence to transform state operations. In this conversation, we dive into the state’s journey with AI, exploring how it began with modernizing unemployment insurance communications, the strategic decision to build tools like the NJ AI Assistant in-house, and the tangible impacts on state employees’ daily work. We also touch on the training processes, ongoing support for staff, and the future potential of AI in government workflows. Let’s get started.
How did New Jersey first venture into using AI, particularly with the unemployment insurance modernization project?
Our journey with AI really kicked off during the pandemic when we were tackling unemployment insurance modernization. One of the biggest hurdles was communication—emails and printed mail often prioritized legal compliance over clarity, leaving recipients confused about what to do next. My team in the Office of Innovation used human-centered design principles to rework those templates. Around the same time, ChatGPT emerged, and we saw an opportunity. By feeding it specific instructions, we slashed the time to update a template from weeks to just an hour, allowing us to scale the effort across hundreds of documents. That was our first concrete, user-facing project with generative AI, and it showed us the potential to transform how we operate.
What were some of the communication challenges you encountered in that project, and how did you address them?
The core issue was that our communications were written in dense, legalistic language that didn’t resonate with the average person. People receiving these notices often had no idea what they meant or what action to take. We tackled this by rethinking the tone and structure using human-centered design, making the language simpler and more actionable. When we brought in AI tools like ChatGPT, it turbocharged the process. We could input prompts to rewrite text in a clear, empathetic way while ensuring legal accuracy, cutting down revision time dramatically and helping us reach more people effectively.
Why did the state choose to develop the NJ AI Assistant in-house rather than opting for an off-the-shelf solution?
Building the NJ AI Assistant internally was a strategic call. While we didn’t create our own large language model—we use the GPT-4 model via Microsoft Azure—we did develop the website and chat interface ourselves. This gave us control over critical aspects like security and integration with state accounts. It also allowed us to embed training requirements directly into the access process, ensuring responsible use. Plus, presenting it as a familiar internal state website boosted adoption—about 20% of our workforce are active users. Financially, it’s a win too; at roughly $1 per user per month, we’re saving millions compared to enterprise plans that start at $20 per user.
How does having a custom-built interface enhance security and accessibility for state employees?
Developing our own interface means we can tie access directly to official state accounts, ensuring only authorized personnel get in. When users log in, they’re greeted with alerts about data and security constraints, reinforcing trust and caution. It’s all hosted within our state infrastructure, so it feels like a seamless part of their workday tools. This setup not only protects sensitive information but also makes the tool more approachable—employees aren’t navigating an unfamiliar third-party platform, which lowers barriers to adoption and encourages consistent use.
What are some practical ways state employees use the NJ AI Assistant in their daily tasks?
The NJ AI Assistant supports a range of tasks that leverage the strengths of large language models. Document summarization is a big one—employees can extract key points from lengthy reports for analysis or decision-making. Another common use is processing public feedback on policies or regulations. Agencies upload comments, and the AI pulls out key themes or even specific quotes, which helps make sense of large volumes of input quickly. We also see it being used to identify overlapping regulations or reporting requirements, though it remains a general-purpose tool adaptable to various needs across departments.
How do you ensure the AI’s outputs, like summaries or quotes, are reliable and free from errors or fabrications?
Accuracy is critical, so we guide employees on crafting effective prompts to minimize errors. For instance, when summarizing feedback, we encourage asking the AI to pull specific quotes, so users can cross-check against the original material for accuracy. This helps catch potential hallucinations—where the AI might invent information. We also pair tool access with training on responsible use, emphasizing the need to verify outputs. It’s about maintaining a human-in-the-loop approach, ensuring that AI is a support tool, not a final decision-maker, to keep trust in the results.
What kind of impact has the AI Assistant had on the day-to-day work of state employees?
The impact has been significant and largely positive. We’ve surveyed users, and a strong majority report that it saves time and boosts productivity. Initially, we supported PDF attachments, but based on feedback, we’ve expanded to include Word documents and Excel files, showing how we adapt to real needs. Agencies like the Department of Labor have fully embraced it for unemployment insurance tasks, integrating it into their workflows. Employees appreciate how it streamlines repetitive tasks, freeing them up for more strategic work, and we’re seeing that reflected in their growing reliance on the tool.
Can you walk us through the training process for state employees getting started with the AI Assistant?
Training is a cornerstone of our AI rollout. We use the state’s existing learning platform, where employees access a series of video modules on responsible AI use. These were developed in partnership with a nonprofit focused on public sector upskilling. The modules cover practical applications and potential pitfalls. To drive participation, the governor released a video strongly encouraging staff to engage, and agency heads have reinforced that message. It’s integrated into our mandatory training framework, ensuring broad reach—about 25 states have even adopted similar programs from the same consortium.
How do you keep employees updated on AI advancements as the technology evolves?
We’re committed to keeping our workforce current with AI’s rapid changes. We’ve planned annual refreshes of training content to incorporate new developments. Right now, we offer a quick-start course and a deeper dive, with plans to tailor the advanced training to specific job roles over time. Beyond videos, we’re building interactive resources like websites with updated use cases and effective prompts. We also maintain an evolving acceptable use policy, ensuring guidelines stay relevant as we learn from agency experiences and emerging best practices.
What additional resources or support do you offer to help employees maximize the AI Assistant’s potential?
Beyond initial training, we provide ongoing support to ensure employees can use the tool effectively. We share practical resources, like tips for writing strong prompts and examples of successful use cases, often through interactive platforms for easy updates. Our acceptable use policy acts as a living guide, and we’re constantly adding new scenarios based on real-world applications within agencies. The goal is to foster a community of learning where employees can exchange ideas and solutions, ensuring they feel supported as they integrate AI into their work.
Looking ahead, what is your forecast for the role of AI in state government operations over the next few years?
I’m both excited and cautious about AI’s future in state government. The potential for positive disruption is huge—think automated bulk processing for tasks like grant document reviews, where AI could scan uploads, extract data, and flag issues in real time, cutting down wait times for citizens. Embedding AI deeper into systems, always with human oversight, could revolutionize how we serve the public, making processes faster and cheaper. But we must stay vigilant about risks like biases or hallucinations, much like we approach cybersecurity. My forecast is that with proper training and guardrails, AI will become a transformative ally in government efficiency, provided we keep our eyes wide open to both opportunities and challenges.
