GAO says agencies aren’t systematically collecting lessons learned from AI acquisitions. Which sounds like a paperwork problem until you notice the paperwork is where the lock-in lives.
The U.S. government is buying AI fast, but not learning fast. That is how you end up paying $300,000 per vehicle per year for ‘productivity.’
What happened
A new Government Accountability Office report (GAO-26-107859) found that four major agencies (DoD, DHS, GSA, VA) were not yet systematically collecting lessons learned from AI acquisitions, which blocks cross-agency knowledge-sharing that the Office of Management and Budget expects via a GSA-managed repository.
GAO’s recommendation is simple and humiliating: update department policies so officials must collect lessons learned (including useful contract clauses) and submit them to that repository. All four agencies agreed (GAO; Nextgov/FCW).
The non-obvious angle: pricing is becoming governance
In private markets, AI vendors are increasingly selling “work done,” not seats. In government, GAO flags one of the hardest problems as pricing and overall cost uncertainty, including long-term assumptions about infrastructure and licensing.
Combine those two things and you get the real risk: procurement teams are being asked to buy a moving target with a meter attached. If you don’t capture lessons learned, you do not just overpay. You lock in a bad meter, and then pay to learn what it measures.
The six failure modes GAO keeps seeing
- Expert bottlenecks: difficulty accessing AI technical experts (like data scientists) to evaluate proposals.
- Data and IP rights: contracts that fail to secure usable rights can block sharing outputs with partners.
- Acquisition timeframes: traditional cycles don’t match fast-changing AI systems.
- Requirements and terms: vague requirements make accountability and performance measurement harder.
- Testing and continuous evaluation: AI systems need ongoing evaluation, not one-time acceptance.
- Cost shocks: GAO cites an Army XM-30 AI licensing proposal priced around $300,000 per vehicle per year, implying more than $500 million annually for licensing alone.
The Singularity Soup Take
“Buy faster” is not a strategy. If you cannot explain your data rights, your evaluation regime, and your meter design in plain English, you are not buying AI. You are buying a future budget surprise with a nice demo video.
What to Watch
- Repository reality: does the GSA repository become a reusable library of clauses and failure modes, or a compliance graveyard.
- Meter standards: do procurement templates start demanding standardized reporting, audit logs, and clear “work unit” definitions.
- Discount lock-in: whether low entry pricing through government-wide agreements becomes de-facto standardization without matching oversight capacity.