Integrating AI into Government Protocols
Federal agencies deploy AI faster than they can govern it safely.

The scale of AI adoption across the federal government has grown dramatically in a short period. OMB's Federal Agency AI Use Case Inventory recorded 3,611 use cases across 56 agencies in 2025, up from 2,133 in 2024 and roughly 700 in 2023, a nearly 70% year-over-year increase. This trajectory signals not gradual creep but accelerating institutional commitment. More than three-quarters of CFO Act agencies deployed at least one major AI chatbot to at least 10,000 employees in 2025, and nearly 90% of federal agencies report adopting AI in some form, though adoption varies enormously in depth and maturity. Concentration is particularly notable among large mission agencies: HHS leads by use-case count, while the Department of Veterans Affairs stands out for high-impact AI specifically, reporting more than 200 such use cases. Yet raw use-case counts tell us scope but not whether these systems are well-governed, effective, or trusted, and that gap defines the central challenge explored throughout this piece.
What AI-Government integration means in practice
Integrating AI into government is not a flip of a switch; agencies move through identifiable phases. The process begins with inventory and use-case identification, where agencies catalogue where AI can or already does touch their workflows. This is followed by a policy and governance framework phase, in which agencies establish who is responsible, what rules apply, and what oversight looks like. Procurement and deployment come next, involving the selection of systems, contracting with vendors, and standing up infrastructure. Finally, ongoing monitoring and accountability ensures that systems continue to perform safely and fairly over time.
The OMB inventory process itself is one mechanism that drives this progression, because requiring agencies to report use cases forces internal discovery before governance can follow. According to a Google Public Sector survey of 250 federal IT leaders, the most common current applications are document and data processing (54%), workflow and process automation (40%), and decision support systems (34%), offering a useful window into where the process actually stands for most agencies. Critically, there is a meaningful distinction between deploying a tool and embedding it into a protocol: the latter requires institutional change, not just technical installation. Readers who understand these phases can locate any specific news item, such as a new executive order or a procurement announcement, within a coherent map of where the process stands.
The policy architecture agencies are required to work within
A cascade of executive and OMB actions between January 2025 and early 2026 reshaped what agencies must do. EO 14179, issued in January 2025, revoked the prior administration's AI order and directed agencies to accelerate AI use in mission-critical workflows. In April 2025, OMB M-25-21 and M-25-22 established mandatory human oversight, AI impact assessments, and continuous monitoring for "High-Impact AI," which is defined as systems whose outputs serve as a principal basis for decisions with significant legal, material, or safety effect, and required agencies to designate Chief AI Officers.
The USA's AI Action Plan, released in July 2025, identified more than 90 federal policy actions spanning innovation, infrastructure, and international diplomacy. OMB M-26-04, issued in December 2025, introduced new procurement and transparency requirements specifically for generative AI and large language models. That same month, EO 14365 established a national policy framework aimed at preempting inconsistent state-level AI laws and directed the Attorney General to form an AI Litigation Task Force within 30 days to challenge state laws inconsistent with federal policy.
The "High-Impact AI" category is the sharpest edge of M-25-21: it is where mandatory oversight requirements bite hardest, and it is where most governance controversy will concentrate. The state-level pressure driving EO 14365 is substantial, as more than 1,000 AI-related bills were introduced across U.S. states and territories in 2025, creating a patchwork that federal policy is now explicitly trying to resolve. This architecture gives agencies a compliance map, but it leaves open significant discretion in how they implement requirements, which produces wide variation in rigor across agencies.
How the Chief AI Officer role is meant to hold the framework together
OMB M-25-21 mandated that agencies designate Chief AI Officers responsible for governance, deployment oversight, and compliance, making the CAIO role a structural mechanism rather than just a title. The role is designed to accomplish several things simultaneously. CAIOs are expected to own their agency's AI use-case inventory and ensure it remains accurate and current, oversee impact assessments for high-stakes systems, serve as the internal accountability point when things go wrong, and bridge between technical staff and agency leadership on risk decisions.
The workforce reality underneath this structure complicates its effectiveness. Only 37% of public sector employees, compared to 53% in the private sector, say their organization has a clear AI strategy, according to Gallup's Q4 2025 data, meaning the CAIO structure exists in many agencies before the strategic clarity to guide it does. This gap matters because a governance role without strategic direction can produce compliance theater rather than genuine risk management. The role's ultimate impact is too early to measure at scale, and readers should treat it as a watch area as implementation matures.
Defense integration as the most structured & highest-stakes instance of the process
For the first time in history, the FY2026 defense budget created a dedicated budget line for AI and autonomous systems, totaling $13.4 billion. Unmanned and remotely operated aerial vehicles account for $9.4 billion of that figure, while maritime autonomous systems account for $1.7 billion and underwater capabilities for $734 million. Supporting software and cross-domain integration represents $1.2 billion, and autonomous ground vehicles account for the remaining $210 million. The DOD's AI Acceleration Strategy, released in January 2026, signals a further shift toward agentic AI capabilities in military operations. On the procurement side, the Chief Digital and AI Office awarded $200 million in 2025 to commercial providers including Anthropic, OpenAI, Google, and xAI, indicating that the government is accessing cutting-edge private-sector capability rather than building from scratch.
The competitive pressure shaping these decisions is significant. China's approach to AI governance in weapons development is permissive by design, creating an asymmetry that U.S. and allied governments are actively managing, a challenge that is both strategic and ethical. Defense is not just the largest instance of the integration process; it is also the most explicit about what accountability looks like when AI touches life-or-safety decisions.
A cascade of executive and OMB actions between January 2025 and early 2026 reshaped what agencies must do.
What agency-level use cases reveal about where the process works & where it strains
Examining specific agency deployments reveals a clear pattern in what separates successful integration from strained governance. The USPTO released an enterprise AI strategy in January 2025, then launched an Automated Search Pilot in October 2025 that surfaces up to ten ranked prior-art references per application before it reaches a human examiner, representing a contained, high-value use case with clear success criteria and human review built in. The NRC's AI Strategic Plan commits to deployment across knowledge management, operating-experience analysis, and regulatory decision support through FY2027, and explicitly anticipates license applications that themselves incorporate AI, meaning the NRC will need to assess AI it did not build. A large federal social services agency deployed predictive analytics for benefit fraud identification and eligibility processing, reducing case backlog by more than 40% and freeing caseworkers for complex claims. That outcome represents a strong efficiency gain, but it also falls into the category of use case where algorithmic bias risk is highest.
International examples offer additional perspective. Portugal's Gov.pt portal introduced a generative AI assistant in 2025 covering more than 2,300 services, enabling multilingual guidance and appointment scheduling and demonstrating what citizen-facing integration looks like at scale. Following Japan's Noto earthquake in 2024, AI analyzed social media and environmental data for real-time crisis response, illustrating an emergency management use case where speed and accuracy trade-offs are at their starkest. Across all of these cases, the deployments that hold up best share a clearly bounded task, a measurable outcome, and a human in the loop for consequential decisions. The ones that strain governance tend to be broader, more autonomous, or to touch civil rights directly.
The accountability gaps that current frameworks have not yet closed
Despite the progress in policy architecture, several accountability gaps remain open. On agentic AI performance, GAO's science and technology arm has noted that even the best-performing AI agents complete only about 30% of complex tasks autonomously without error, meaning the remaining 70% lands on a human who may lack the context or training to catch the error before it propagates. This is not an argument against AI deployment; it is an argument for designing human-oversight protocols around actual error rates rather than assumed ones.
Algorithmic bias in high-stakes services presents a separate and serious concern. The potential for unfair outcomes is particularly sensitive in government services, as child protection algorithms have already generated public outcry and benefit fraud models carry similar risk. The civil rights dimension is not hypothetical. Security remains the dominant structural blocker, with adversarial risk affecting 48% of agencies as a barrier to adoption and reliability concerns affecting 35%.
The public trust deficit is equally significant. Seventy-two percent of U.S. adults expressed concerns about AI in a 2025 Heartland survey, citing privacy, cybersecurity, lack of transparency, and algorithmic bias. Only 6% of government agencies are in the "ideal" state of both high internal confidence and demonstrably trustworthy AI, which is the lowest share of any industry surveyed. Compounding this, 38% of government organizations are simultaneously underutilizing trustworthy AI safeguards and overrelying on AI, which represents precisely the failure mode the governance frameworks are supposed to prevent. OMB M-26-04's transparency requirements for generative AI are a direct response to some of these gaps, but the requirements are new and compliance remains uneven.
Where the federal-state boundary sits & why it matters for how AI protocols develop
States are not waiting for federal leadership on AI governance. More than 1,000 AI-related bills were introduced across U.S. states and territories in 2025, reflecting active legislative activity at the state level. EO 14365, issued in December 2025, directly responds to this by framing state-level AI laws as a threat to federal AI dominance and directing the Attorney General's AI Litigation Task Force to challenge state laws inconsistent with federal policy. The underlying tension is genuine: states have historically served as laboratories for governance innovation, and preemption could suppress experimentation that the federal framework has not caught up to, or it could resolve a genuinely unworkable patchwork. For agencies operating across jurisdictions, and for regulated industries dealing with both federal AI-using agencies and state regulators, the federal-state boundary is not abstract because it determines which rules they must actually follow. This question remains unresolved as of mid-2026 and will likely be shaped by early outcomes of the litigation task force.
What the current state of integration implies about where government AI protocols are headed
Several dynamics point toward how government AI protocols will evolve from here. The inventory-to-governance loop will tighten: as use-case counts grow, the pressure on Chief AI Officers and OMB oversight mechanisms to keep pace will increase, and the 3,611 figure is best understood as a measure of how much is now in-scope for governance, not just deployment. Agentic AI represents the next frontier that current frameworks are not yet built for, as the State Department's move toward agentic systems and the DOD's Acceleration Strategy both point in this direction. The existing oversight models, which assume a human reviews AI output, will need to be redesigned for systems that act with less human review.
The trust gap is structural, not just perceptual. The 6% "ideal state" finding means that most agencies face a legitimacy challenge alongside a technical one, and closing it will require transparency and demonstrated accountability rather than simply better AI. International precedents are beginning to accumulate, with Portugal's 2,300-service assistant and Japan's disaster-response AI showing what mature citizen-facing integration can look like. U.S. agencies can benchmark against these examples, but will need to build the public trust foundation first. The through-line for readers is that integration is not a destination but a continuing process of policy, deployment, and accountability refinement, and understanding the structure of that process is what allows observers, practitioners, and policymakers to engage with it usefully rather than reactively.
Sources
- https://www.hklaw.com/en/insights/publications/2026/03/white-house-releases-a-national-policy-framework-for-artificial
Provided details on EO 14365's national policy framework and the Attorney General's AI Litigation Task Force directive.
- https://www.softwareimprovementgroup.com/blog/us-ai-legislation-overview/
Provided context on the volume of state-level AI bills and the federal-state regulatory tension.