The Social Security Administration (SSA) believes agentic artificial intelligence (AI) is ready for mission use, but agencies will only realize its potential if they pair the technology with rigorous governance, continuous testing, and modernized data infrastructure, the agency’s chief AI officer said on June 26.
Speaking at ACT-IAC’s AI Forum, Brian Peltier, SSA’s head of technology, customer data, and AI and chief AI officer, said the agency is approaching a fundamental shift in how it delivers services as AI capabilities mature alongside years of technology modernization.
“We’re on the cusp of really changing the way we operate, how we work, and how we service our customers,” Peltier said.
Agentic AI is ready to support agency missions not only because the technology has advanced, but also because agencies have spent years modernizing their underlying systems, Peltier said. Even so, he cautioned that successful deployment depends on governance, security, and operational safeguards.
“[Agentic AI] is nondeterministic. It can do things that you may not expect,” he said. “So, what you really need to do is constantly look at and constantly test from the beginning.”
First, agencies should first determine whether AI is the right solution for a problem before moving forward. Once a use case is identified, they should establish clear performance metrics and desired outcomes, and then bring together security professionals, attorneys, developers, and data specialists throughout development and deployment, Peltier said.
“Monitor that all the way through into production to make sure that it is doing exactly what you want, how you want,” he said. “You must be testing; you must be sampling in production … looking for is it doing what I expect … or … where people are trying to then do injections … if you’re doing something public facing and they’re trying to hack your system.”
Peltier also stressed that AI modernization should strengthen service delivery across all customer channels rather than shift resources exclusively toward digital interactions. By encouraging customers who are comfortable using online services to do so, he said, SSA can devote more staff and resources to people who require in-person assistance.
As an example of AI already improving agency operations, Peltier highlighted an internal retrieval-augmented generation (RAG) tool that combines SSA policy and compliance documents with a large language model.
Before deploying RAG, employees often struggled to locate policy guidance in legacy systems and sometimes turned to Google because the agency’s internal search capabilities were less effective, he said.
SSA ingested policy documents into a searchable vector database, paired that information with the large language model, and built a chatbot that allows employees to ask questions in natural language. Answers providing direct links to the underlying policy documents.
“They can interact in more natural ways … and refer back to the exact policy with a link. I would say it exploded in use,” Peltier said.
According to Peltier, frontline employees now use the tool daily to help serve customers, while information security teams and inspectors general have expressed interest in adopting similar capabilities. SSA has also made the capability available through an application programming interface so it can be integrated into other agency applications.
The tool has also improved collaboration between policy and user experience teams by making authoritative guidance easier to access.
“Now folks can use the tool … quickly search to see what the policy actually says and go back intelligently and debate … what should be on screen,” Peltier said.