Using AI to Improve Public Services in New Jersey: An interview with Dave Cole

In early January 2026, Governor Phil Murphy signed legislation establishing the New Jersey Innovation Authority, making New Jersey the first U.S. state to codify an innovation office in law.

The move comes just after the State of New Jersey was awarded a Public Benefit Innovation Fund grant to scale a set of practical AI tools designed to reduce delays and administrative burden in programs such as Medicaid, Unemployment Insurance, and Summer Electronic Benefit Transfer.

To mark the moment, Beth Simone Noveck sat down with Dave Cole, New Jersey’s Chief Innovation Officer, to reflect on what it takes to modernize government in practice, from using AI to reduce benefit delays to designing institutions that last beyond any single administration.

What follows is an edited transcript of their conversation, focused on real-world lessons for public leaders.

Beth Simone Noveck: I’m delighted to be here with Dave Cole today, Chief Innovation Officer of the State of New Jersey, and I want to congratulate you on this fantastic new grant. Maybe you could just tell us very briefly: what is this winning proposal? What has the Office of Innovation won, and what problem are you solving?

Dave Cole: Thank you. And of course, as folks will likely know, you were our first Chief Innovation Officer, so a lot of what we’ve been able to do builds on work you helped put in motion.

For a number of years, we’ve had AI-related projects focused on improving how government operates and delivers services. Often those projects have been one-off efforts with individual agencies. With the Public Benefit Innovation Fund grant, we’re able to take some of those projects and build them into a more generally available platform, so that across state government we can use the same tools we’ve been piloting.

It also gives us resources to make this work available to the broader civic technology community, especially as more responsibility for running critical programs shifts to the states. We can work collectively across states to solve shared problems.

Beth Simone Noveck: Could you unpack in plain English what the components of the project are? The proposal uses some technical language, like entity resolution, which may not be obvious to everyone. What problems are these tools designed to solve?

Dave Cole: We looked across many points in government benefits processing, especially healthcare and human services, where things slow down.

One of the first issues is documents. People upload a document, and then what happens? Depending on the program, it might get printed and put in a physical file, or scanned and manually typed into a database or even a mainframe. Staff often have multiple windows open, copying information field by field.

One clear use case we’ve seen, working with the Economic Development Authority, is using an AI model with computer vision to extract information from uploaded documents. These documents aren’t always structured, and we don’t always know the format in advance, but the models are very good at pulling information into structured data.

That also lets us validate the document: are the dates current, are required fields present? We can then give applicants feedback about whether the document looks accurate, so they can fix errors before submitting. In a pilot with EDA, this saved about 20 days in processing time, because applications no longer had to go through lengthy manual review.

You can imagine how that applies to SNAP, Medicaid, labor programs, and many others. The programs differ, but the processing bottlenecks are often the same.

Beth Simone Noveck: So from a resident’s perspective, they don’t have to learn new technology. They can still apply on paper if they want, but on the back end you’re using AI to process information faster and more accurately. What are the other components of the project?

Dave Cole: Another common challenge is synthesizing feedback. Agencies get a lot of information through public websites and comments, but it’s hard to process. We have tools that take that input and synthesize it into topics, trends, and sentiment, producing reports people can actually use.

With documents, there are really two parts. One is extracting the data. The other is validating it and providing real-time feedback so applicants can correct errors before delays occur.

We’ve also helped agencies with memo generation. In government, moving things forward often requires memos that pull from multiple sources. Writing those can take a week or more. We have a tool that takes source information and templates and generates draft memos in minutes.

In all of these cases, there is always human review. The tools reduce repetitive work, not decision-making. Editing, sign-off, and final decisions always rest with people.

Another tool, developed with our AI for Impact Fellows, analyzes public-facing government websites and recommends ways to simplify language and improve accessibility, so people can better understand programs and access benefits they’re entitled to.

Beth Simone Noveck: These are tools that can be acted on right now, and you’ve already piloted them with agencies. Could you share an example of what you’ve accomplished using these approaches?

Dave Cole: The example that always stands out to me is the Summer EBT program. A couple of years ago, New Jersey participated for the first time. The program provides food assistance to families whose children qualify for free or reduced-price meals during the school year.

The challenge is that eligibility data lives across many systems—school records, SNAP, Medicaid, TANF, foster care—and there’s no common link between them. People move, addresses don’t match, and Social Security numbers aren’t reliable or always available.

We built a tool that applies a series of matching tests, starting with exact matches and then moving to probabilistic matching that assigns confidence scores. Human experts still review the results.

The net effect was that we identified about 100,000 families each year who were eligible but would not have received the benefit otherwise. We’ve now done this for two years.

That matching problem isn’t unique to agriculture or nutrition programs. You see it across government.

Beth Simone Noveck: When you identify people who may be entitled to a benefit, how do you design the handoff between the technology and human judgment?

Dave Cole: There are always human inputs. Staff set the confidence thresholds they want to see, and then experts review the files just as they would with a manual process.

We also choose technology based on the decision involved. For matching, we use machine learning, not generative AI. Generative AI is useful for other things, like memo drafting, but it’s not appropriate everywhere.

With generative AI, the tool assembles a draft memo from source documents. Staff review the draft alongside the sources, edit it, and sign off. Whether AI is involved or not, state employees take ownership of the work.

In many cases, review is actually better because people aren’t exhausted from repetitive drafting. They’re focused on verification and accuracy.

Beth Simone Noveck: Sometimes systems are slow because the policy itself is flawed. How do you balance simplifying processes with advocating for deeper policy change?

Dave Cole: We have to do both. Policy moves slowly, and people can’t wait in the meantime. It’s very hard to justify telling someone their benefits are delayed because the policy isn’t written well.

If technology can simplify the process and get people help faster, we should use it, while continuing to advocate for better policy. The time between recognizing a policy problem and fixing it is where people fall through the cracks.

Beth Simone Noveck: You’ve also talked about making feedback easier to process. How do you encourage agencies to ask for feedback when that can feel overwhelming?

Dave Cole: Having tools helps, but we also start with research. One of the first things we do is look at places like Reddit to understand what people are actually experiencing. Some of it is tough to read, but it’s real.

We also use surveys and unstructured feedback. When we synthesize feedback, we always tie insights back to direct quotes so humans can verify them and understand the real sentiment.

There’s also a positive loop. When people see improvements reflected in services, they say so, and that reinforces the work.

Beth Simone Noveck: Since we recorded this conversation, the Office of Innovation has been institutionalized as the New Jersey Innovation Authority. What does that mean for this work?

Dave Cole: It gives us a stable foundation. We started with a small team and budget and grew over time. Codifying this work in law means it doesn’t depend on any one person and can continue across administrations.

It provides infrastructure so future teams can move quickly on priorities like improving customer experience, reducing the time it takes to start a business, and delivering higher-quality services.

Beth Simone Noveck: Critics sometimes ask how you fix government with more government. How do you respond?

Dave Cole: The track record shows that targeted investment delivers better outcomes. Faster benefit delivery means people get money sooner and get back into the economy sooner.

One example is Business.NJ.gov. An analysis showed that in 2024 alone, businesses that started earlier through the platform generated over $160 million in economic activity. Sometimes it takes investment in capacity to produce outsized gains.

Beth Simone Noveck: As you look ahead, what’s the most important thing you hope this work carries forward?

Dave Cole: Staying connected to the people we serve. New Jersey is incredibly diverse, and no single experience represents everyone.

Continuous research and feedback keep us grounded. If we keep building with people, not just for them, we’ll continue to find simple solutions to complex problems.