$300,000 a year per vehicle just for AI licensing: the GAO's lessons
5 min read
The US Congress asked the Government Accountability Office — the independent body that checks how the federal government spends public money — to examine how federal agencies buy artificial intelligence. Its answer, published on 13 April 2026, is report GAO-26-107859: one concrete case after another that anyone buying an AI system today — in a ministry, in a large company, in a hospital — should read before signing their next contract. Not because it only concerns the United States, but because it documents, with names and figures, the same mistakes that recur everywhere when an organisation buys a technology it does not yet have the tools to evaluate.
The context explains the urgency: according to the federal Chief Information Officer, AI use across US agencies more than doubled from 2023 to 2024, and Congress has appropriated around $1.7 billion for artificial intelligence initiatives across government. The GAO analysed 44 contracts and agreements signed between September 2018 and February 2025, examining 13 of them in depth across four agencies — Defense (DOD), Homeland Security (DHS), the General Services Administration (GSA) and Veterans Affairs (VA). The central finding does not concern a single supplier or a single contract: none of the four agencies had a policy requiring the systematic collection of lessons learned from its AI acquisitions. Every office starts from zero, every time.
The half-a-billion-dollar-a-year mistake
The starkest example comes from the Army, in the same field that combines mechanics and software as defence programmes. The XM-30, the reduced-crew combat vehicle meant to replace the Bradley, includes an AI-assisted targeting system. For the software licence alone, programme officials received proposals of around $300,000 per vehicle per year — a figure that, applied across the planned fleet, would have meant over $500 million a year in licensing fees alone, before even counting the cost of the vehicles themselves. Officials themselves call it “exorbitant”. The point is not whether it truly is: it is that nobody, when evaluating the proposal, had a benchmark to compare it against. Without a structured record of prices and terms used elsewhere in the same government, every office negotiates blind — and suppliers know it.
The system that vanishes without a trace
The Department of Veterans Affairs had launched an operational trial to detect signs of suicide risk in veterans’ messages, run by supplier SoKAT. In January 2023 it withdrew it: officials said it did not improve enough on existing solutions to justify the extra cost. A reasonable decision. The problem is what did not happen next: nobody documented why the tool did not work, which requirements were wrong, what to look for next time. The GAO calls it a missed opportunity, because the same department runs other AI programmes on veteran suicide prevention that could have used exactly that experience. The knowledge disappeared along with the contract.
What actually works
The report also cites two positive examples, and it is instructive that they share the same principle. The Pentagon’s Maven programme works with an agile method — weekly planning, 90-day cycles — adding requirements as the system’s real-world capability proves itself, instead of writing them all up front for a system that is still unknown. USAi, the GSA platform that gives fifteen agencies access to models from Google, Meta, Anthropic and OpenAI to test before adopting at scale, is moving from a free model to a pay-per-use one: you pay for what you actually use, not a fixed licence negotiated before knowing whether the system is needed. In both cases the principle is the same one we apply to every operational trial on a real use case: verify before committing at scale, and pay against a measured result, not a promise.
Four recommendations, and a deadline this month
The GAO recommended that Defense, DHS, GSA and Veterans Affairs update their internal policies to require the systematic collection of lessons learned and their sharing in a common repository run by the GSA. All four agencies agreed. DHS has committed to updating its AI procurement guidance by July 2026 — this very month. It is the same gap found in every regulated sector between announcing a principle and making it verifiable: an accepted recommendation is not yet a policy in force, and a policy in force is not yet a repository populated with real cases.
What to do before signing your next AI contract
The lesson is not American, nor purely public-sector. It applies to an Italian public body as much as to the procurement office of a large company evaluating an AI supplier for the first time:
- Set a cost benchmark before negotiating: what others pay for a comparable capability, and how cost is measured over time, not just at signature.
- Test on a real case, at small scale, before signing a multi-year contract: that is the difference between XM-30 and USAi.
- Document the acquisitions that do not work, not only the successful ones: the lesson from a failure is worth as much as one from a success, if someone writes it down.
- Get independent technical expertise to evaluate an AI proposal before signing, not after the contract is already running.
- Write an exit route into the contract: data portability, reversibility, no technological lock-in — the same principle that Italy’s AI procurement guidelines are about to make mandatory for public administration.
Anyone who has already been through the same learning curve — buying critical technology from a supplier that can change terms mid-course — knows the cost of not having these answers ready before the crisis, not during it.
Want to structure your next AI system purchase around a verifiable operational trial, with measured costs and data that stays yours, instead of a supplier’s promise? Let’s talk in a thirty-minute session.