AI Bright Spot: Making zoning laws user-friendly with AI in Lebanon, New Hampshire

The Partnership for Public Service AI Center for Government® is publishing a series of blogs to celebrate how artificial intelligence and intelligent automation are being used by the government to serve the public.

We spoke with Melanie McDonough, chief innovation and AI officer for the City of Lebanon, New Hampshire, about her work developing an AI agent called Zoner, with the goal of making zoning laws user-friendly for residents.

This interview has been edited and condensed for clarity.

Please tell us more about Zoner. How did the idea for this tool come about?

“We are addressing a problem that’s been around for a long time, which is ‘How do we help people understand our zoning regulations?’ There’s a real disconnect when someone asks, ‘Can I have chickens in my backyard?’ and in response we give them a 200-page PDF of all the zoning ordinances.

“If people don’t get an easy answer to their questions, they’re going to call, which then means city staff are overloaded with phone calls and questions.

“As generative AI has come out and as it keeps improving over time, we really see that there are possibilities now to make that zoning laws search process better for everyone involved.”

How did you get started developing Zoner?

“I started around the time that OpenAI released its agent builder. It looked very accessible to me as a non-technical person. Within the city government, I’m more of a generalist who works between all the teams to connect all the tools they use together. When I first saw the agent builder, I thought ‘Oh, I understand this.’

“We started by replicating the process of someone coming to the counter or calling to ask about a zoning question. Normally when this happens, our representatives will ask a series of questions to narrow down the process — things like, ‘Are you calling the right office?’ and ‘Which zoning district do you live in?’

“We broke that process and workflow down into smaller chunks, then trained the AI agent to do the same. We scoped each part of the process down to very small, narrow tasks in the same way humans do it.

“In a nutshell, before we even started touching a generative AI tool, we really focused on mapping this whole process. Then we focused on the little tasks we could get AI to do very reliably and repeatably.”

How are you testing this tool and getting it ready for public deployment?

“Before we started migrating the tool over to the city’s network, a lot of this [testing] was handled directly by me. For example, I would tag someone in our planning department and say, ‘Hey, test this out and see how it’s doing.’ They would put in some questions based on what they’re dealing with from the staff side and then share feedback with me.

“As we move things more to the city side, we’re now developing evaluation questions to start measuring reliability. When we hit certain points of reliability, then we will decide whether we will allow it to go through [for official use]. For certain questions that we’re not comfortable with the chatbot answering, it’ll be trained to say, ‘You have to talk to a staff member now.’ We’re also making sure it’s very transparent at the beginning that this is not meant to be a legal source.”

What about dealing with risks that come from inputting sensitive information such as addresses or other personally identifiable information?

“This is something we’re still working through and scoping. The chatbot does ask for people’s addresses, which is public information. However, instead of a city employee having to go over to our GIS system, look up a caller’s address, get their zoning district and come back, we’re allowing the tool to do that task directly.

“When using these tools, you don’t want a toxic combination of information coming together. For just an address, we use that just to do the ordinance lookup, but that information isn’t really being stored anywhere besides that. But if you start layering different things that start adding additional details, that could expose something sensitive about their address that could then cause problems.

“To prevent this, some ideas we’re looking at include collecting information only on the client side instead of our servers and finding ways to mask or dump the data. These are things we need to address as we go through our final stages of preparation and this is a very important point for us.”

Are there any leaders, agencies or cities that have inspired this project?

“The first example that comes to mind for me is Boston. I have been really interested in the work they are doing with Model Context Protocol, or MCP, to connect AI tools to their open data portal and make public data more accessible. I also think their work to make permitting information more understandable and easier for the public to navigate is a great example. It’s not exactly the same as building an AI chatbot, but it gets at the same core of getting the data right, getting the content right and making government information easier for people to use.”

Thank you for joining us, Melanie. This is such wonderful work!

For more on the technical specs of Zoner, see Melanie’s detailed step-by-step project toolkit: https://melaniemcd.ai/projects/zoner-mcp.

---

Continuing the conversation

The AI Center for Government champions AI innovators across all levels of government. If your agency is taking steps to lead AI well, we’d love to hear from you. Join us as we highlight real-world AI use cases and convene public sector leaders from across the country to share tools and insights to lead confidently in the age of AI.

We’re here to help!

• Sign up for our newsletter.
• Follow us on LinkedIn.
• Get in touch! Email us at [email protected].