Raleigh Set a High Bar for Its AI Assistant
One of Raleigh’s biggest concerns was language access. About one-third of people visiting the city in person speak a language other than English, Leos said, and the city found that their customer service experience was far less predictable.
Daisy was tested in English, French, Arabic and Spanish and given information about Raleigh Parks programs.
The city also established a firm benchmark: Daisy needed to answer at least 90% of questions correctly before Raleigh would consider spending hundreds of thousands of dollars on a broader deployment.
Early results weren’t close.
During internal testing with 25 city employees, the assistant posted an accuracy rate of about 30% and a satisfaction score of roughly 4 out of 10.
“Clean data is key when it comes to AI,” Bosley said.
Raleigh learned that lesson after discovering that one source document fed to Daisy was an outdated strategic plan that still identified an old mayor and city council. The conflicting information caused the assistant to return incorrect answers.
“Garbage in, garbage out,” Bosley said. “It is very true with AI and customer service.”
The system improved considerably before and during public testing. By the end of the pilot, only about 5% of Daisy’s answers were incorrect. But it still failed to meet Raleigh’s goal of 90% correct answers because too many responses amounted to, “I’m sorry, I don’t have that information for you.”
There was another problem.
“People just were not coming up to talk to the avatar,” Bosley said.
READ MORE: Here is a guide to AI governance for state and local government agencies.
Does Government AI Need a Human Face?
Raleigh found that Daisy’s human appearance presented another unexpected challenge.
“A face is controversial,” Bosley said.
About half of the residents participating in focus groups strongly preferred having a face associated with the assistant. They liked the visual interaction. The other half strongly disliked it.
Some residents became fixated on the avatar’s appearance. Bosley recalled one reaction: “What is wrong with her hair? It’s sticking out of the screen.”
Raleigh even asked focus group participants to choose among 24 possible faces for a digital assistant. Every participant selected a different one.
That led residents to suggest a distinctly Raleigh alternative: Don’t make the AI look human at all.
“Use something like an acorn,” Bosley recalled hearing. “We’re the city of oaks.” Others suggested rotating among different human avatars so residents would encounter a variety of appearances.
The feedback underscored a larger lesson for the city: Deploying resident-facing AI isn’t simply a technical decision. Officials also must consider cultural, psychological and behavioral factors when deciding how — and where — residents should encounter the technology.
DIVE DEEPER: ServiceNow uses AI to improve citizen services.
Residents Want AI To Make Government Easier
That reluctance didn’t translate into blanket opposition to AI.
Raleigh found that residents were willing to use AI when it offered an obvious advantage. If 10 people were waiting to speak with a city employee while an AI kiosk could answer a simple question immediately, Bosley said, residents were much more interested.
They also saw value when AI could make government services more convenient, surface difficult-to-find information or provide assistance outside normal business hours.
Still, Raleigh identified a tension between AI adoption and public comfort. The city cited the 2025 PayIt Consumer Digital Government Adoption Index, which found that 83% of governments are implementing AI while consumer comfort with AI stands at 50%.
For complex problems, Raleigh residents repeatedly told the city that they preferred people.
“If it is an emotional question, if it is something that’s complex, if I need something but I don’t actually know what the question is, I just want to talk to a person and work through it with them,” Bosley said, summarizing what the city heard in its focus groups.
That finding is pushing Raleigh to think beyond resident-facing AI. For complicated services, Leos said, staff-facing AI may offer more value by handling tedious or time-consuming work while employees focus on problem-solving and personal interaction.
“Not every service need is equally suited to that same type of technology and AI tool,” Leos said.
And that may be Raleigh’s biggest takeaway from Daisy: The question isn’t simply whether a city should use AI. It’s where AI actually makes the resident experience better.
