Focusing AI on High-Value Work
The initiative reflects a broader AI strategy centered on solving business problems instead of deploying technology for its own sake. Before launching projects, the agency established four guiding principles: align AI initiatives with business goals, integrate them into existing processes, rely on trusted commercial AI models rather than building foundation models internally and require human verification of AI-generated output.
The unemployment adjudication project became one of Ohio’s first production AI use cases because it offered a measurable opportunity to improve operational efficiency. Using Amazon Bedrock, the agency created a streamlined version of its adjudication manual by extracting relevant statutes and administrative rules, identifying inconsistencies and gaps, and drafting plain-language content that can also support public-facing information.
Manoharan said the project reduced staff research time by roughly 30%, while the agency projects it could reduce unemployment claim processing times by about 5%.
Ohio also applied generative AI to post-call quality assurance in its unemployment contact center. Previously, supervisors could review only a small sample of recorded calls because the process was labor intensive.
The proof of concept used AI to analyze approximately 150 calls and is projected to increase the number of calls reviewed by 50% while reducing the time required for analysis by half, Manoharan said.
READ MORE: Here is a guide to AI governance for state and local governments.
Building Around Existing Processes
Matt Duncan, a solutions architect with Amazon Web Services who worked alongside Ohio on the projects, said the technical implementation focused as much on simplicity as on AI.
“We wanted to get the benefit of generative AI, but we also didn’t want to increase operational overhead to maintain those things,” Duncan said.
To accomplish that, the team relied on serverless and event-driven architectures that allowed individual components to be modified without affecting the entire application. Duncan said the team also wanted the AI models to rely primarily on Ohio’s own policies and data rather than their general knowledge.
“We wanted those models to rely upon the data more than what it knew,” he said.
The project also reinforced the importance of understanding business processes before designing AI workflows. Duncan said engineers initially believed they understood how adjudication reviews worked, but further conversations with agency staff revealed additional business knowledge that improved the ultimate solution.
DIVE DEEPER: State and local governments must address technical debt in the AI era.
Lessons for Other Agencies
Duncan said agencies pursuing similar initiatives should establish architectural principles early, validate concepts before scaling and prepare data carefully.
“Don’t be afraid to iterate on your prompts,” he said. “Don’t be afraid to put some examples into your generative AI prompts, because that’s going to help the model understand what your expectations are.”
He also encouraged agencies to adopt serverless, event-driven architectures that make it easier to modify AI applications as requirements evolve, allowing organizations to test, refine and scale projects without adding unnecessary operational complexity.
