As Agentic AI Emerges, NIST Rethinks How Standards Keep Up

The National Institute of Standards and Technology (NIST) is working to address barriers many sectors face in adopting artificial intelligence tools and standards, officials said at the 2026 AI+Expo in Washington, D.C.

“We need to figure out how to equip those who are adopting and deploying these systems so that they can evaluate and understand which should I adopt, which is best for my use cases,” said Stevie Bergman, head of applied systems at NIST’s Center for AI Standards and Innovation (CAISI).

One CAISI project aims to reduce barriers to agentic AI adoption in the healthcare, financial services and education sectors through a community engagement tool designed to support interoperability across the digital ecosystem. Bergman said NIST’s long-standing and diverse community network is helping identify which standards are most relevant to specific sectors versus the broader ecosystem.

“We get a lot of information on what those experts in their realm need in order to understand if an AI system is fit for their purpose. NIST can support them through pilots, evaluations and understand what we need to gain confidence,” said Bergman. “From those learnings we can put out like papers and guidelines, we can really leverage our position to support the wider field. This is crucial across NIST, including at CAISI.”

Leveraging the Standards Community

An April 30 report from NIST partner Center for Democracy and Technology (CDT) noted the need for broader engagement across the standards community to help translate contextual evaluations into real-world applications.

Miranda Bogen, chief technologist and director of CDT’s AI Governance Lab, said involving a wide range of professionals has historically been difficult. Conversations often became too focused on technical processes and practices, making it harder for others to understand how their work fits into the broader ecosystem.

“We’re trying to highlight places where different actors across the ecosystem could play more of a role, like developers. Even if they don’t know all of the eventual end uses, they probably do know how to measure systems reasonably well, compared to a lot of the deployers,” said Bogen. “We’re just trying to kind of connect the dots across that space.”

Ensuring Standards Keep Pace with AI Innovation

Researchers are also examining how to accelerate the standards development process to keep pace with AI’s rapid growth. Jesse Dunietz, AI standards, policy and international engagement lead at NIST, said the standards community is still developing frameworks for predictive machine learning systems as agentic AI adoption accelerates.

“Old AI systems have not gone away. We have generative and agentic AI, but in practice everyone’s still deploying predictive machine learning systems,” said Dunietz. “The standards community barely got our hands around what assurance, measurement and evaluation looks like for those early systems.”

Across NIST, researchers are taking a broader approach to evaluating whether systems are fit for purpose. Dunietz said researchers are also examining how evaluation systems themselves should be assessed. He pointed to NIST’s Information Technology Laboratory (ITL) generative AI project, which tested synthetic content detection capabilities.

ITL researchers evaluated the persuasiveness of deepfakes and monitored how generative AI systems responded when users asked them to reveal they were deepfakes. Dunietz said this type of research – examining the long-term process of standards evaluation – provides the building blocks for developing standards for future generations of AI.

“We can assess how good the deepfake detectors and generators perform and pit them against each other to build up both the measurement ability and see how systems are performing in both places,” said Dunietz.