A new GAO report says agencies aren’t systematically capturing what they learn when they buy AI. That sounds boring, which is how you know it’s important.
The government is buying AI like it’s 2003 search engines again, except the invoices meter your ‘productivity’ and the exit costs come later.
What happened
The Government Accountability Office reviewed federal AI acquisitions and found that four major agencies (DoD, DHS, GSA, VA) were not systematically collecting lessons learned from AI procurements, even though OMB guidance expects knowledge-sharing via a GSA repository (GAO; Nextgov; FedScoop).
GAO’s fix is brutally unsexy: update agency policies so acquisition teams must collect and submit lessons learned. All four agencies concurred.
The non-obvious angle: “pricing” is becoming a governance problem, not a finance problem
Private-sector AI sellers are already shifting from selling seats to selling work. Goldman Sachs described the pitch as “units of labor” or “units of productivity,” and companies like Salesforce and Workday are packaging usage that way (Business Insider).
Now combine that with what GAO flagged as one of the hardest issues for agencies: AI pricing and overall cost uncertainty, including long-term infrastructure and licensing assumptions (GAO).
In other words, procurement is being asked to buy a moving target with a meter attached. If you don’t capture lessons learned, you don’t just overpay. You lock in bad meters.
The six failure modes GAO keeps seeing
- Missing experts: agencies struggle to access AI and cybersecurity specialists to evaluate proposals.
- Data rights and IP: FEMA couldn’t share certain model outputs with partners because it didn’t secure the right data rights at award (GAO).
- Timeframes that don’t match reality: officials warned traditional multi-year acquisition cycles don’t fit fast-moving AI.
- Requirements + contract terms: vague requirements make it hard to hold vendors accountable.
- Testing + continuous evaluation: AI systems vary widely, and universal tests don’t exist yet, so agencies need structured evaluation over time.
- Cost shocks: GAO cites an Army example where proposed AI licensing fees were around $300,000 per vehicle per year, implying more than $500 million annually just for licensing (GAO).
Why it matters
“Lessons learned” sounds like a corporate training slide. In practice, it is the difference between a government that can bargain and one that gets bargain-hunted.
If agencies don’t retain institutional memory, every new AI procurement becomes a fresh start, which is exactly how vendors win: discounts up front, proprietary workflows in the middle, and the hostage note at renewal.
The Singularity Soup Take
The AI arms race is being decided by contract clauses, not demos. The winning strategy is not “buy faster,” it’s “buy in a way you can exit.” If you can’t explain your data rights, testing regime, and meter design in plain language, you’re not buying AI, you’re buying a future budget surprise.
What to Watch
- Repository reality: do agencies actually submit usable terms and lessons, or does it become a compliance graveyard.
- Meter standardization: do procurement templates start demanding standardized reporting, audit logs, and “work unit” definitions.
- OneGov lock-in: GAO notes government-wide agreements with major model vendors. Watch whether “convenience” becomes de-facto standardization.
Sources
GAO — "Artificial Intelligence Acquisitions: Agencies Should Collect and Apply Lessons Learned to Improve Future Procurements (GAO-26-107859)"
Nextgov/FCW — "Agencies are missing a step to share information on better AI acquisition, GAO finds"
FedScoop — "Agencies fall short on documenting AI acquisition best practices, GAO says"
Business Insider — "AI firms rethink pricing, shift from users to work done"