AWS Public Sector Blog
Anne Arundel County integrates generative AI into case management to empower staff and improve citizen services
Every day, the Anne Arundel County, Maryland, Office of Community Engagement and Constituent Services receives hundreds of resident concerns by phone, online, and even in handwritten letters. These aren’t routine inquiries but complex problems, such as zoning disputes, permitting questions, and other issues that residents have escalated to the highest levels of county government.
Staff want to help, but first they must summarize voicemails, transcribe letters, and manually categorize each case. That work can consume 45 minutes or more before they can begin addressing the problem. Using Amazon Web Services (AWS), the county’s app development team integrated generative artificial intelligence (AI) into their existing case management system to improve the process. Now, voicemails, emails, and letters are automatically summarized, categorized, and routed for support, giving staff more time to focus on helping residents.
Focusing AI on staff to help support citizens
While the county initially considered implementing AI assistants for citizen use, they decided to focus on improving internal staff productivity to enhance response times for the county’s more complicated citizen needs. It would also allow county staff to maintain their personal touch, which is a cornerstone of the county’s culture.
The idea was to look at employees’ daily activities and identify bottlenecks. Constituent services staff spend much of their time in the field, meeting with residents and putting eyes on problems firsthand. They didn’t have hours for tedious data entry. “If we’re going to use generative AI for something, we should use it in a way that we’re actually helping employees,” said Thomas Mann, software developer and project manager for the county.
County leadership fully supported this internal-first strategy. “The county administration pushed us to set up responsible AI policies, establish test environments, and figure out how to do this right,” said Jack Martin, chief information officer (CIO) for Anne Arundel County. That meant starting with clear goals and privacy protections, using Amazon Bedrock Guardrails to set boundaries for the AI. The team also committed to a core principle: AI would make suggestions in their Case Manager tool, but staff would have the final say.
Integrating AI into existing workflows
Rather than build something new, the team enhanced what already worked. Case Manager, their in-house proprietary ticketing system for tracking resident concerns from intake through resolution, was the foundation. The development team integrated generative AI capabilities directly into this familiar workflow via middleware they built themselves, Generative Intelligence Summaries and Tagging (GIST). GIST connects Case Manager to:
- Amazon Bedrock for accessing large language models (LLMs)
- Amazon Transcribe for converting speech to text
- Amazon Textract for extracting text from scanned documents
- FFmpeg to convert incoming files into formats compatible with AWS services
The county chose Amazon Bedrock for its flexibility. “Amazon Bedrock allows us to switch out foundation models while still maintaining the same format for our requests,” said Paul Ng, full-stack developer for the county. This model-agnostic approach has proved valuable. The county started with Claude by Anthropic in Amazon Bedrock and has since moved to testing Amazon Nova to evaluate performance improvements and potential cost efficiencies, swapping models without rewriting code or disrupting operations. Additionally, GIST uses AWS Cloud Development Kit (AWS CDK) for automatic deployment, and the team uses AWS CodePipeline and AWS CodeBuild for continuous integration.
Throughout the process, weekly meetings with their AWS account team provided a sounding board for technical decisions. When questions required deeper expertise, AWS connected them with the specialists they needed. “You always know you have the support of the AWS team,” Mann said.
Keeping staff in control
While the workflow looks familiar to staff, there are key differences. When a voicemail, scanned letter, or email arrives, they drag and drop the file into Case Manager. Within seconds, the system:
- Transcribes audio or extracts text
- Generates a summary
- Applies categories and creates relevant tags
- Extracts contact information
- Drafts a response using existing templates
Staff review all AI-generated content and decide how to proceed. They can edit summaries, adjust categories, modify responses, or ignore the AI suggestions entirely.
“It’s a human support system, rather than a replacement product,” Mann said.
This approach extends to how responses reach residents. The system outlines replies using templates created by the constituent services team but includes placeholders for case-specific details that staff must complete to personalize each response. Before any text reaches Amazon Bedrock, algorithms mask sequences of numbers that could represent sensitive data, such as addresses or phone numbers. Amazon Bedrock Guardrails provide additional filtering to detect and block inappropriate content, prevent hallucinations, and more.
Smarter categorization for better reporting
The system also helps staff understand what they’re really dealing with. Case categories drive the county’s reporting on everything from noise complaints to zoning issues, but rigid categories often force staff to label complex issues as “general,” undermining metrics. AI-generated tags solve this by adding relevant keywords based on case content, helping to capture each issue’s layered complexity. For example, a zoning complaint might also involve school access, or a noise complaint might connect to a larger neighborhood event. These details help staff grasp the full picture before they respond.
Accelerating response times
That understanding now comes faster than ever. Multiple paragraphs of resident correspondence are now distilled into concise overviews that help staff prioritize and assign cases to the right people. Audio files representing voicemails are processed in under 20 seconds for summarization and categorization, and handwritten letters can be scanned and summarized just as easily. In every case, relevant information is parsed and used as case information. Altogether, menial elements of the process that once took close to an hour now require almost no manual effort.
County staff recognized the value immediately. Within weeks of launch, team members proactively requested additional features. “The AI summaries are very helpful to get a snapshot of what the case is about,” said Stacey Fitzgerald, the county’s constituent services coordinator. “For processing voicemails, the time has significantly improved since there is less data entry needed and the need to relisten to voicemails to capture all of the details accurately.” For residents, the result is faster, more informed responses to their cases.
Building toward deeper insights
With the foundation in place, the county is already exploring what else AI can do. One initiative uses Amazon Bedrock via GIST for sentiment analysis. Not every message carries the same urgency. Understanding whether a resident is frustrated after months of unresolved issues or simply has a quick question helps staff triage accordingly.
The team is also working with AWS on Amazon Neptune, a graph database service, to identify connections between cases that might otherwise go unnoticed. Multiple residents might report what appear to be separate issues, like traffic complaints, noise concerns, and questions about a public event, that all stem from the same incident.
“Through the use of AI and Neptune, we could assign multiple related cases to one person to handle with less repeated effort across an already overloaded staff,” Mann said.
Other opportunities are emerging as well. Permitting accounts for more than 30% of county website traffic and involves a tangle of federal, state, and county regulations, an area where AI-assisted tools could eventually help residents navigate intricate processes. SeeClickFix from CivicPlus, the tool residents use to report potholes, graffiti, and other issues, generates its own stream of cases that could one day benefit from the same AI-powered workflow.
Advice for other local governments
For other local governments considering similar projects, Mann’s advice reflects what worked in Anne Arundel County: “Look for where repetitive tasks slow your team down and investigate how generative AI can save time,” he said. “Citizens will appreciate the improved responsiveness without knowing what specific technology is used behind the scenes.”
That captures the county’s philosophy. The goal was never to deploy AI for its own sake, but to help staff do their jobs better and deliver quicker, more responsive service to residents, and strengthen the connection between county government and its community. “By focusing on practical problems, building on existing systems, and keeping humans in the loop, we created something that works,” said Martin. “Now AI enables our staff to spend more time interacting with citizens, which is the work that matters most.”
To learn how AWS can help your organization build and scale AI solutions, reach out to the AWS Public Sector Team.