Airys Intelligence Lab · Case Study
Beyond Frontier Models
Federal Agency Specialization
Federal AI procurement, rather than consolidating into a handful of frontier vendors, is fragmenting into narrow, agency-specific workflows underneath them.
General-purpose deployments like Claude at HHS or Copilot across agencies are just the on-ramp. The real procurement activity is happening one workflow at a time, underneath them.
Small business isn't optional
Structural requirement
FAR Part 19 requires agencies give small businesses "maximum practicable opportunity" in federal acquisitions — enforced via subcontracting plans, set-asides, and SBA eligibility pathways.
The goal: keep a pipeline of new entrants able to compete in emerging technical domains.
VA Case Study
What deployment data actually shows
The public narrative centers on frontier tools — ChatGPT-style assistants at HHS, Claude at FDA, Copilot everywhere. The VA's 2025 AI use-case inventory — 155 systems — tells a different story:
VA AI use-case inventory · 2025
- In-house, bounded tools
- General / other systems
Examples of the in-house 32%: REACH VET (suicide-risk prediction) and ambient clinical documentation pilots — bounded tools for bounded environments, not repurposed general models.
General-purpose = discovery infrastructure
Sequencing, not substitution
Through GSA OneGov agreements, agencies get frontier models at near-zero marginal cost — value flows to a few infrastructure companies, not a broad vendor base.
HHS's 2025 AI strategy is explicit about the order of operations: build shared infrastructure first, then layer specialized tools on top.
Constrained > general
Workflow-specific demand
VA's clinical ambient-documentation initiative shows what "specialized" buys in practice — real-time transcription conforming to VA-specific documentation standards, at accuracy and latency levels general LLM interfaces don't reliably hit.
Why bounded tools win the workflow
Every meaningful contract is government-specific
The structural condition
Every civilian AI contract awarded at meaningful scope — outside general infrastructure managed by large IT primes — is definitionally government-specific.
CMS
Fraud detection against the claim schema
VA
Documentation against record standards
Treasury
Modernization against legacy systems
The Takeaway
The market isn't the frontier model. It's the workflow underneath it.
Frontier deployments are the on-ramp, not the destination. The tractable opportunities sit in bounded, agency-specific systems — where domain fit beats general capability, and small businesses have a statutory lane to compete.