I was handed an AI-built pricing sheet last week with rates wrong enough to break a contract, and a hundred AI-written questions meant for someone else. Same root cause. The model is now the cheap part. The expert whose judgment makes it safe is the part you cannot buy, and the part the market is busy trying to replace.
Six Army captains audited a fleet of dozer maintenance records, found more than 82% of the entries useless, and published it under their own names. A firm priced a job with real AI productivity and got told the price was too low to be credible. Two cases this spring, one about data and one about price, and the system marked both honest parties down. Here is why the machinery underneath speed, innovation, and small business still pays out to the old model, and what it means for the military health enterprise that has the most at stake.
On May 29, 2026, VA posted RFI 36C10B26Q0485, market research for an enterprise AI buy meant to move a 540,000-person workforce from assistive tools to autonomous agents acting on veteran health data. The same document sends governance out of scope. Here is how VA earns the leadership it claims: sequence the agents by risk, put governance on the quarterly clock the price already runs on, and buy all four parts, the capability and the three preconditions that make it safe.
On April 20, 2026, the Defense Health Agency replaced thirty years of how it buys medical capability. A new portfolio-based acquisition model, a requirements process built to kill 'bring me a rock,' and an FY2027 budget that already voted on where the money goes. Here is what changed, who runs it, and how it shows up in live contracts.
No code. No exploit kit. Plain English. A security researcher pulled 60 pages of hidden instructions out of an AI doctor, rewrote them, made it triple a drug dose. The Defense Health Agency is fielding the same architecture in military exam rooms right now.
A credentialed third party generated evidence before the read. Twenty-one years later, that is the architecture CMS is shutting other modalities down for not having. The 2:47 a.m. stroke scene that proves the primitive, the OpenAI/MCP pattern radiology operationalized two decades early, and the federal procurement vehicle that has not yet been built.
The strongest detail in the HealthSplash case is a physical-exam test, documented as performed, on a patient the clinician had never met. The workflow was the fraud. How American healthcare keeps designing the same disaster, and the procurement language that closes the gap.