Raleigh, North Carolina, has an AI agent named Alli working its IT help desk. The desk has four people supporting about 4,400 city employees. Alli handles routine requests from start to finish, such as adding someone to a distribution list or walking a user through a password reset. It works from the city’s knowledge base. A second agent, Ral-E, sorts incoming tickets and sends them to the right team.
On October 8, 2026, ServiceNow, which sells the product, announced that Raleigh is the first city government to put its help desk agent into production. The release claims 66 percent lower service desk costs from redirected staff time, nearly half of IT requests resolved without a person, 98 percent of tickets routed correctly on first contact, and 98 percent of AI-written incident summaries accepted as written. ServiceNow has a commercial interest in these numbers. The release gives no baseline, period, or method. Treat them as directional.
The city’s own account is more useful. CIO Mark Wittenburg told Smart Cities Dive that testing began in February. He limited Alli’s permissions the way he would for a new hire, who does not get administrator rights on day one.
Early on, after Alli gained access to the knowledge base and old tickets, it began telling employees to change registry settings, something the help desk does not ask users to do. Wittenburg said the city had given it too much autonomy and too much information. They pulled it back and widened its data sets slowly. He said the mistakes traced to inconsistencies in the city’s own data, which could be fixed.
Help desk staff now oversee Alli and keep its knowledge accurate. The city built AI training courses for staff and a cross-departmental group to assess the technology. Wittenburg wants agents to resolve up to 85 percent of IT requests. The city plans to connect both agents to its Ask Raleigh resident portal.
The agent’s first mistakes came from the city’s own documentation. Fixing the data was part of the deployment.
Why it matters
Raleigh’s sequence does not depend on the vendor. Start with routing, which is easy to check. Give the agent narrow permissions. Widen them as accuracy holds, and pull them back when it does not.
The registry episode shows why. Wittenburg said the agent was not wrong. It was handing users steps the help desk would normally take itself. An agent turns whatever is in the knowledge base into action, including material never meant for the person asking. Someone has to own that knowledge base.
The job changed too. Help desk staff now supervise an agent and curate what it knows. That is different work from answering tickets, and it needs training. The resident portal raises the stakes. Mistakes inside city hall reach employees who can call the help desk. Mistakes on a resident portal reach the public.
Questions before you deploy an agent
- What can the agent do on day one, and who approves each expansion?
- Is every action logged, and can you reverse it?
- Who owns the knowledge base it works from, and how often is it reviewed?
- What are your own baselines for cost and resolution, measured before the agent arrived?
- What does the agent do when it does not know the answer?
Treat the 66 percent as the vendor’s number. Treat the probation period as the city’s lesson.
Sources
- Raleigh, North Carolina, reduces IT service desk costs by 66% with ServiceNow AI agents (ServiceNow press release, Oct. 8, 2026)
- How do you train a new AI service agent? A lot like you would a new employee, a city CIO says (Smart Cities Dive, July 24, 2026)
- This new Raleigh, N.C., IT help desk agent is, well, agentic (GovTech)