Federal agency leaders say trustworthy data, clear AI governance and risk assessment must underpin the deployment of agentic AI across government, according to a November 10 report by GovCIO Media & Research. 

Speaking at GovCIO Media & Research’s AI Summit on November 6 in Tysons, Virginia, officials from the Defense Health Agency (DHA) and Commerce Department discussed how agencies should assess the risks of systems that can act and make decisions with limited human intervention.

Jesus Caban, chief data and analytics officer at the DHA, said the reliability of agentic AI depends on the data and governance frameworks behind it. “Everything starts with the data,” he said, emphasizing that even advanced AI models cannot be trusted without dependable underlying information.

The comments offer a practical perspective for chief data officers (CDOs) responsible for preparing agency data environments for AI adoption. They also highlight the importance of establishing authoritative data sources, maintaining auditability and assessing the potential consequences of automated decisions before introducing agents into operational workflows.

Authoritative Data Sources Form the Basis for AI Trust

Caban said the DHA initially focused its data strategy on indexing and publishing authoritative datasets so AI systems could identify which sources to use and trust. The agency is taking a staged approach to agentic AI adoption, building on data quality, data strategy and governance before expanding the technology across the enterprise.

He said the DHA evaluates potential agentic AI use cases against three criteria: whether an agent’s decision can be reversed, whether there is a single auditable source of truth for its actions and whether its use could cause patient harm.

Use cases that raise concerns under these criteria undergo further governance review before the agency begins a pilot. The approach is particularly relevant to healthcare environments, where automated decisions can have direct consequences for patients and where the provenance and reliability of information are critical.

Commerce Department Senior Advisor for Technology and Strategy Matheus Passos also emphasized trust as a condition for AI adoption. He said employees would be reluctant to use AI tools if they could not trust that their data was safe, secure and protected.

Clear Objectives Must Guide Agentic AI Deployment

The summit discussion also addressed the need to define what agencies expect AI systems to accomplish before deploying them. Alexis Bonnell, head of AI adoption and deployment at OpenAI for Government, said agencies should establish clear measures of success and identify the workflows where AI can deliver meaningful improvements.

Bonnell recommended starting with workflow efficiency, determining which data AI systems should access and identifying processes that may need to change before introducing agents. This helps agencies establish whether a use case is suitable for agentic AI and what outcomes they should measure.

Together, the recommendations point to a governance process that begins before an agent is deployed. 

Agencies need to understand the data it will access, establish clear boundaries for its actions and determine how those actions will be monitored and evaluated. So, making data quality, authoritative sources and auditability part of the groundwork for agentic AI adoption is imperative.

Data Leaders Must Establish the Conditions for Trustworthy AI

The discussion at the summit highlights several responsibilities for CDOs as agencies move toward agentic AI. They must ensure that datasets are authoritative, current and properly documented, with clear ownership and access controls. 

Agents also need to operate within defined permissions, and their actions must be traceable so agencies can establish what data informed a decision and how the system reached it.

The DHA’s emphasis on reversibility and a single auditable source of truth offers a practical starting point. Before approving an agentic AI pilot, CDOs should establish whether its actions can be reviewed or reversed, whether the information behind its decisions can be verified and whether potential harm has been assessed. These safeguards also need to extend to ongoing monitoring once an agent enters operational use.