Artificial intelligence is widely recognized as a linchpin of modernization. Yet, as federal agencies increasingly look to integrate AI into their operations, the stumbling blocks are not technological. It's failures in procurement practices that, if not avoided, undermine the potential of AI.
Two high-profile government automation projects offer cautionary tales. Each reveal how even well-funded or successful efforts can falter—not because the technology is lacking, but because of the acquisition processes in place.
Project Maven: Ethics Overwhelm Innovation
In 2017, the U.S. Defense Department launched Project Maven, aiming to harness computer vision for object detection in drone video feeds. The project reached operational use within the year. But in June 2018, Google, a key contractor, announced it would not renew its contract after its employees protested the use of Google's AI in military applications. The program's momentum was halted not by technical limitations, but by an ethical debate that procurement processes were ill-equipped to anticipate or resolve.
IRS Modernization: Legacy Systems and Contractual Inertia
In March of this year, after spending $1.5 billion on modernization projects for the IRS, the U.S. Treasury Department paused the initiative so it could reassess technology choices, including AI. The IRS is now developing a new modernization framework due by early 2026 that focuses on unifying fragmented data systems, leveraging tech for taxpayer service and fraud detection, and reducing tech contractors.
The need for a pause highlighted a common problem: legacy platforms and broad, multi-year contract vehicles can stifle adaptation when new tools emerge.
Recurring Weaknesses in AI Procurement
These cases point to three recurring risks in AI acquisition: continuity of operations, data rights, and sustainment activities.
First, continuity of operations is often overlooked. When a contractor withdraws, as Google did with Maven, agencies can be left scrambling to maintain critical capabilities. Without explicit contract provisions for continuity, agencies risk the sudden loss of capability.
Without explicit contract provisions for continuity, agencies risk the sudden loss of capability.
Second, data rights are a persistent challenge. Federal Acquisition Regulation (FAR) and Defense Federal Acquisition Regulation Supplement (DFARS) default to the government maintaining only limited rights in software and technical data. Without explicit licensing agreements for privately developed model weights and training data, agencies could find themselves unable to modify, share, or fully control essential AI system components.
Finally, sustainment activities are frequently underappreciated. AI models are not static; they degrade over time unless continuously retrained and monitored. The National Institute of Standards and Technology calls for ongoing evaluation throughout the AI lifecycle, and the Office of Management and Budget require agencies to manage rights and safety risks. Yet, procurement often treats AI as a one-time deliverable, ignoring the need for ongoing maintenance and future budgetary investments.
Regulatory Reforms: Necessary but Not Sufficient
Under Executive Order 14275 and OMB M-25-26 (PDF), the government is beginning to rewrite acquisition rules more broadly, aiming for greater flexibility. This could also be an opportunity to think through the nuances of AI procurement. Without explicit contract language, agencies will continue to face vendor lock-in, unmanaged model decay, and transition challenges.
Federal rules already require modular contracting for software, and the GAO's guides on Agile assessment and portfolio management set out incremental practices that could be applied to AI acquisition. But these best practices are not yet standard in AI procurement.
AI doesn't create new acquisition problems; it amplifies existing ones. To build AI systems that work—and keep working—agencies should be mindful to:
- Define desired outcomes before purchasing AI tools.
- Use pilot programs that can evolve as technology changes.
- Clarify data ownership and rights in contracts.
- Build internal expertise to manage and sustain AI systems.
Setting clear standards for AI procurement practices now could help avoid pitfalls for years to come.
