Industry witnesses urge DoD to prioritize AI-ready data, enterprise infrastructure and acquisition reform
Get AI-powered insights, summaries, and transcripts
SubscribeSummary
Representatives of Scale AI and Cohere told senators the Department of Defense must build AI-ready data systems, adopt an "implementation-first" mindset, and reform procurement to field AI tools faster and avoid vendor lock-in.
Dan Tadros, head of public sector at Scale AI, and David Farris, global head of public sector at Cohere, told the Cyber Subcommittee that three near-term actions would help the Department of Defense adopt AI at scale: create AI-ready data requirements and enterprise-wide data infrastructure, shift to an "implementation-first" mindset, and reform acquisition rules that slow AI deployment.
Tadros said the U.S. government is the "world's leader" in the quantity and diversity of data but that much of that data is unused. "If The US wants to turn our data into an advantage, this must change," he said, recommending enterprise AI data infrastructure and AI-ready data standards.
Both witnesses urged experiment-driven deployment. Tadros advocated setting a "North Star" for robust AI implementation within five years, focused on agentic applications, and said Scale is working on deployments through the Defense Innovation Unit's Thunder Forge effort.
On procurement, Farris suggested using existing authorities and raising simple-acquisition thresholds for urgent AI operational requirements so lower-level contracting officers could acquire capabilities without complex processes. Tadros and Farris both recommended modernizing procurement to prevent vendor lock-in and favor agility and performance over firm size.
Why it matters: Witnesses argued that data handling, testing benchmarks and acquisition practices are the bottlenecks slowing DoD AI use. Cohere recommended custom, domain-specific benchmarks and human evaluations to avoid overfitting to public tests.
Supporting details: Tadros said China had spent "at least $1,200,000,000 on data labeling alone" last year while the U.S. spent "under $100,000,000," a discrepancy he used to illustrate different national approaches to data investment. Farris described building "custom benchmarks" that models had not seen to better assess performance in operational contexts.
Ending: Senators and witnesses agreed on the need for sustained public-private engagement; witnesses called for clearer pathways for industry to bring solutions to the department and for Congress to consider acquisition and budget authorities to speed fielding.
