Share with your COO
Federal AI deployments are failing procurement, not technology, according to this RAND commentary on AI acquisition failures. Two cases frame the argument: Project Maven lost Google as a contractor in 2018 after employee protests, leaving DoD without a key capability mid-flight; the IRS paused a $1.5 billion modernization effort in 2025 to reassess AI choices, exposing how multi-year contracts lock agencies into decisions before the technology landscape settles. The recurring gaps are contract continuity provisions, data rights ownership, and sustainment budgets for models that degrade without retraining.
What this means for your business
The failure pattern here isn’t unique to government. Any large organization that treats an AI deployment as a one-time procurement, signs a broad multi-year contract, and assumes the vendor relationship is stable enough to skip contingency planning is running the same risk. The IRS situation, specifically, has a private-sector mirror in every enterprise that bought a “platform” to solve a data problem and discovered three years later that the platform owns the data model weights, not the company that paid for it.
Data rights is the clause most procurement teams skip because it feels theoretical at signing. It stops feeling theoretical the moment a vendor is acquired, pivots its pricing model, or simply stops supporting the version your operations run on. Model weights, the trained parameters that encode everything the AI system learned from your data, are the core asset in any deployed AI system. FAR and DFARS defaults favor vendors, and private-sector contracts are often worse. If your current AI vendor agreements don’t explicitly address who owns the weights, who can retrain them, and under what conditions you can transfer them, that’s a gap that compounds with every month of production use.
The second-order problem is sustainment. AI systems aren’t software in the traditional sense where a shipped version holds its value. A fraud detection model trained on 2023 transaction patterns performs worse against 2025 fraud patterns, quietly, without throwing an error. Most enterprise AI budgets fund the build; almost none fund the ongoing retraining, monitoring, and evaluation that NIST’s AI Risk Management Framework treats as mandatory. If you’re renewing a vendor contract this cycle, the question to bring is not “what does the new version include” but “what does continuous model maintenance cost, who owns it, and what happens to performance metrics if we stop paying for it.”
Concept deep-dive: Model decay
Model decay is the gradual degradation in an AI system’s accuracy as the real-world conditions it was trained on drift away from current reality. Think of it like a map printed in 2020: still mostly accurate, but wrong in ways that matter and getting worse. Unlike broken software, a decaying model keeps running and producing outputs, just increasingly unreliable ones. The business risk is that performance erosion is invisible until it surfaces in a decision, a fraud miss, a wrong classification, or a failed prediction.
Based on reporting from AI Won’t Outrun Bad Procurement, originally published 2025-09-29 03:00:00.
