Every federal agency now has an AI strategy. Far fewer have data an AI system can trust. Since OMB memo M-25-21 directed agencies to accelerate AI adoption in April 2025, the pressure has shifted from choosing models to feeding them the data they need.
At a recent GIST 360 Breakfast Briefing held on November 10, 2026 at the National Press Club, government and industry IT leaders gathered to dissect this very issue and share two different approaches for building the data foundation needed to deliver on the promise of Federal AI.
Swish CTO Sean Applegate hosted Harley Stout of the War Data Platform in the Department of War’s Chief Digital and AI Office, Murari Selvakesavan, who leads enterprise information management services at the Centers for Medicare & Medicaid Services, and Jim Cosby, CTO for Public Sector and Partners at NetApp. Their lessons start with what happens when a platform tries to do everything.
The Department of War’s Advana analytics program began with too expansive of a purpose. The department had hundreds of business systems, each with its own schema and no central place to support a fiscal audit. Over time the platform grew into a monolith that tried to solve every problem at once. It also left little data maturity behind at the action officer level, so institutional knowledge evaporated with each contract and administration transition.
To address these issues, DoW restructured Advana into the War Data Platform (WDP) in January 2026 to speed AI and support a clean audit by fiscal 2028. The platform is now aligned to one mission, moving data from more than 750 source systems into warfighters’ hands as fast as possible. Optimizing for two or three outcomes, such as resiliency and latency, made investment decisions for the WDP far clearer than balancing hundreds of variables. Stout’s team retired antiquated technology and saw resiliency metrics climb within six months.
Focus clarifies the mission, but data still has to be usable when it arrives. “AI-ready data to me is what we used to call enterprise-ready data,” Stout said. That means tagging, standard sharing protocols, known latency and context. When data is tagged correctly, access can be approved automatically or by a single data steward. When it is not, sensitive data can take weeks to reach the people who need it. WDP applies one bronze, silver and gold standard across structured, unstructured and imagery data, so a promoted table or document means the same thing everywhere.
CMS has spent nearly 20 years building toward the same goal. Its Integrated Data Repository holds about two petabytes of data. A federated data fabric launched in September 2026 connects datasets that stay with the offices that govern them. A move to Apache’s Iceberg open format lets many tools work from one copy instead of duplicating data for each. Selvakesavan’s larger focus for CMS is building meaning and context into the agency’s data. An AI agent has to understand what a claim is, how a provider ties to an organization and who counts as a beneficiary. Without that context and the business rules behind it, an AI tool can answer questions but cannot reason.
Context makes data usable. Lineage makes it defensible. “Trust travels with the data,” Selvakesavan said. CMS tracks the source, owner, steward and every transformation, along with versions of each dataset. If data is poisoned, the team can trace the blast radius across data products and applications and revert to a known good state. The Justice Department, the inspector general and congressional analysts rely on CMS data to fight fraud and shape policy. The same discipline now extends to AI agents, which CMS manages through a registry that limits what each one can access and do.
The DoW applies a controls mindset for the same reason. Comptroller teams plan to use agents to help pass the fiscal 2028 audit, and those agents must clear the same IT and financial controls as any person. Stout measures them on speed and repeatability, because a fast answer that changes from run to run earns no confidence. Independent auditors recently found the platform’s financial controls sound, a clean result that lets the offices relying on its data skip repeat testing.
Trusted data still has to survive an attack, and agents add a new kind of risk. Cosby described an agentic tool that, while attempting routine cleanup, escalated its privileges and deleted a company’s production data and most of its backups in nine seconds. Ransomware crews work in a similar order, disabling logging, destroying backups and then encrypting primary data. His defense starts with immutable backups locked by a compliance clock, plus multi-administrator verification so no single person or agent can delete them. Snapshots then let an agency recover hundreds of terabytes to a point minutes before an attack.
A strong foundation pays off only when people keep coming back to it. Both agencies stopped counting what they build and started tracking what gets used. WDP once counted applications and use cases on its platform. It now measures monthly active users, how often each dataset is accessed and a stickiness ratio that shows whether people return, across more than 100,000 monthly users on three networks. CMS also measures usage. Every team that draws on shared, trusted data is one less office building its own pipeline, copy or pilot. The payoff is speed. Selvakesavan wants answers that take about two weeks today to be delivered in under an hour.
The panel closed with the moves leaders can make in the next 12 months.
Explore more sessions along with the GIST of Govt IT podcast at gist360.com and reach out to the team at Swish to assess agency data readiness for building an AI foundation for the future.
Thanks to NetApp for sponsoring the discussion.