Government technology leaders have been through this cycle before.
Cloud computing promised agility, scalability and efficiency. Those benefits were real, but so were the operational consequences. As agencies migrated applications and monolithic systems, they inherited new layers of complexity across infrastructure, services, vendors and costs. One lesson became clear: Adopting a new technology model without the visibility and governance to manage it replaces one form of technical debt with another.
Artificial intelligence is creating a similar challenge, one layer higher in the stack. As agencies move from pilots into production, they are no longer managing only applications and infrastructure. They must also understand models, prompts, tokens, agents, frameworks and the tools those agents invoke. The technology is different; the operational challenge should feel familiar.
That is an advantage. Agencies do not have to invent a new operating model for AI. They can apply what they learned during cloud modernization much earlier in the adoption cycle.
AI is an operational layer, not a separate domain
Traditional software is largely deterministic. Engineers establish the paths software can take, and when something goes wrong, teams investigate known systems and dependencies. Agentic AI changes that. Agents reason through a task, invoke tools, retry failed actions and choose different paths toward an outcome. That flexibility is what makes the technology valuable, but it means teams need visibility into more than the final output.
An analysis from Datadog of anonymized production telemetry from thousands of organizations running large language models (LLMs) shows how fast this complexity is arriving. More than 70% of organizations use three or more models, and the share using more than six has nearly doubled in a year. Adoption of agent frameworks nearly doubled in that time, and the number of services using them more than doubled.
AI environments are becoming distributed and interconnected faster than many organizations can build the security, governance and operational practices to manage them.
Reliability problems are not always model problems
An AI system can fail while the underlying model performs exactly as expected. Datadog’s analysis found that 5% of the LLM calls in that dataset returned errors, 60% of which came from exceeding rate limits.
In an agentic environment, a failed action can trigger repeated retries. Without safeguards, an agent may keep calling the same service, exhausting capacity and turning a localized failure into a sustained outage. What looks like an AI problem is often a distributed systems problem involving retries, capacity and dependencies.
Technology teams already know how to handle this: backoff mechanisms, usage budgets, fallback strategies, capacity planning and end-to-end monitoring. AI does not make those disciplines obsolete. It makes them more important.
The cloud cost lesson
Agencies have a vested interest in applying the hard-learned lessons of early cloud economics. The cloud delivered elastic infrastructure, but elasticity without visibility did not produce efficiency. Resources proliferated faster than teams could understand or optimize them.
Agencies lived this. In June, the Government Accountability Office found that 17 of 24 federal agencies said controlling cloud costs required changes to their IT management approaches. Agencies also cited multivendor environments, interoperability, procurement guidance and workforce capacity. Cloud modernization proved that adopting new technology models requires agencies to change how they manage cost, performance, risk and complexity
AI is showing similar warning signs. Over the past year, average tokens per request more than doubled for typical users and quadrupled for heavy users. Sixty-nine percent of input tokens were associated with system prompts, while fewer than one-third of eligible LLM calls used cached context.
The issue is not that tokens cost money. It is that agencies need to know why resources are being consumed and whether that consumption serves the mission. Token counts, model usage and the number of AI tools deployed are not measures of success.
Intuition here is unreliable. In a randomized controlled trial run by METR, a nonprofit that conducts empirical AI evaluations, 16 experienced developers completed 246 tasks in codebases they had worked in for an average of five years. When AI tools were allowed, completion time was 19% longer. Afterward, those same developers estimated AI had made them 20% faster. The trial measured the early-2025 model frontier, so it’s possible that newer tools may perform differently, but the gap between perceived and measured results is the durable finding. Self-reported benefits are not evidence. Agencies need to understand whether AI improves an outcome and have the telemetry to determine what happened when it does not.
From “human in the loop” to “human in the lead”
Responsible AI discussions focus on keeping a human in the loop. That is essential for high-impact decisions, but it should not mean a person approves every individual action a system takes. At scale, that becomes a bottleneck.
A more sustainable model is layered, risk-based oversight. Humans define purpose, operating boundaries, performance thresholds and escalation conditions. Automated controls and continuous telemetry enforce those boundaries, surface exceptions and alert people when intervention is needed. Humans remain in the lead even when they are not approving every step.
Federal policy reflects this. Office of Management and Budget Memorandum M-25-21 calls for AI infrastructure spanning development, testing, deployment and continuous monitoring. For high-impact uses, it requires ongoing testing, periodic human review and appropriate oversight and accountability. The National Institute of Standards and Technology’s Generative AI Profile recommends real-time monitoring that can trigger human intervention, processes for deactivating systems operating outside defined tolerances and post-deployment plans for incident response and change management.
Several agencies are already working this way.
- Customs and Border Protection established an AI Center of Innovation in 2022. Its models redact sensitive information in images, detect activity in streaming video and assist in screening cargo, while officers retain operational judgment.
- The State Department built a machine-learning system to support declassification review of diplomatic cables. Humans set decision thresholds, reviewed all uncertain recommendations and spot-checked confident ones, reducing manual review volume by 80% while preserving oversight of consequential decisions.
- A 2026 GAO report found that at the Centers for Medicare and Medicaid Services, analytics and machine-learning models flag anomalous billing patterns and generate investigative leads, while investigators review records and gather evidence. Across its broader program-integrity efforts, which extend beyond AI, CMS estimates that administrative actions prevented $11.9 billion in potentially fraudulent payments from fiscal years 2022 through 2024.
The distinction here is between human involvement and human leadership. AI processes volume, identifies patterns and prioritizes exceptions. Humans set the rules, evaluate uncertainty, weigh evidence and decide when a system should be corrected, overridden or stopped.
Observability is AI governance
The cloud era taught technology leaders they cannot manage what they cannot see. For AI, visibility has to extend beyond infrastructure health and model output to prompts, retrieved context, policies, tool calls, latency, errors and costs across the full path of a request.
Those signals answer the questions that matter: Why did an agent take that action? Which service did it call? What changed before performance degraded? Why did costs increase? Did the system stay within its policies?
Agencies operate at enormous scale, under strict accountability requirements, with finite resources. Manual oversight will not scale with AI adoption. But agencies enter this era with something the industry lacked at the start of the cloud transition: experience. The terminology is changing. Infrastructure becomes models, application traffic becomes prompts and tokens, distributed services become agents and tool calls — but the operating principles hold.
The next phase of government AI will be defined by whether agencies can apply those principles before complexity becomes technical debt. That starts with observability.
Chris Arroyo is a regional director for U.S. public sector at Datadog.
Copyright © 2026 Federal News Network. All rights reserved. This website is not intended for users located within the European Economic Area.