Rather than deploying AI everywhere, agencies will become more selective, particularly in the FedCiv space. Through trial and refinement, federal organizations will identify high-value, mission-aligned use cases—such as benefits processing, regulatory review, logistics optimization, or IT operations—and double down on those areas. Expect fewer pilots, fewer one-off tools, and a stronger focus on repeatable, scalable AI applications that directly improve outcomes.

As agencies work to operationalize AI at scale, interoperability will emerge as a foundational requirement. Fragmented data systems will increasingly limit the effectiveness of AI models, forcing agencies to prioritize data integration, standardization, and secure sharing across programs and bureaus. Agencies that can’t connect data across legacy and modern platforms will struggle to extract meaningful insights from AI investments.

For its part, the Department of Defense’s (DoD) AI Acceleration Strategy, released in January, makes hay of reducing the bureaucratic barriers that stand in the way of accessing data for AI models (which appears to include for purposes of either training or interaction with, say, chatbots).

AI ambitions will put pressure on aging infrastructure, pushing federal IT modernization into overdrive. Agencies will prioritize cloud migration, data platform upgrades, and application modernization to support AI workloads. Modernization efforts will be less about technical debt alone and more about building the foundational capabilities—compute, data, security, and governance—required to make AI operational at scale.

As AI becomes more deeply embedded in federal systems, it will also expand the attack surface for adversaries. Agencies will face growing risks from data poisoning, model manipulation, and AI-enabled cyberattacks. LLM vendors Google and Anthropic have begun noting in recent months that the same capabilities that this form of AI leverages in cybersecurity can be repurposed for adversarial uses. In response, cybersecurity strategies will evolve to include AI-specific protections, such as model integrity monitoring, secure data pipelines, and continuous threat detection—making AI security inseparable from broader zero-trust initiatives.

Talk over whether the private AI industry is operating in a bubble increased markedly in 2025. One reason may be the (perceived) mismatch between model capabilities and model impacts. Another may be the trajectory of commercial firms model development, like OpenAI’s GPT-5, and the apparent shortfall from Artificial General Intelligence.

2026 will likely be an inflection point for this “bubble” talk. Federal agency efforts involved in a litany of AI-related efforts — from the types of models most prominently built and offered in the private sector to infrastructure like data centers and the increased reliance on them — will be shaped by this outcome.

In June and July of 2025, the DoD’s Chief Digital and Artificial Intelligence Office (CDAO) awarded contracts to the major commercial LLM vendors: OpenAI, Google, Anthropic, and xAI. The contracts specified that the companies develop “prototype frontier AI capabilities” in “warfighting and enterprise domains.” With a ceiling value of $200M each, the work in all four cases is to be completed by July of this year.

It is unclear what specifically will result from these contracts, given that the focus of the contracts is both “warfighting and enterprise domains” (i.e., combat and organizational workflows), but the applications that are developed may signal the direction that U.S. defense agencies find the most direct value in this class of “generative” AI models.

Mission-critical defense scenarios — where the risk of failures must be minimized — hinge on the technology deployed for operation working reliably and predictably with defense forces and personnel. Recent years have seen a sharp uptick in DoD interest in AI models that can be applied to mission-critical scenarios with this reliability. As many in-house R&D efforts note, however, the reliability of machine learning models — which currently dominate the commercial AI scene — is often unacceptably low for trustworthy deployments in mission-critical scenarios.

Hybrid techniques in AI that do not rely solely or primarily on artificial neural networks (such as the neural networks underpinning generative LLMs) will continue to be fleshed out within DoD R&D. It is likely that 2026 will see an increasing prominence of these techniques, including the lesser-hyped “neuro-symbolic AI,” alongside the building out of more mainstream techniques by firms including OpenAI, Google, and so forth. 

This will occur in parallel with the DoD’s continuing focus on “autonomy,” or systems that can execute human-given inputs over a potentially shifting range of circumstances.