Datadog Proposes Observability Framework for Government AI Deployment
Chris Arroyo argues that deep visibility into probabilistic AI is essential for security, cost management, and taxpayer accountability.
Government agencies are increasingly integrating large language models into citizen services and back-office operations, but the shift from traditional software to probabilistic AI is creating significant hurdles for security and oversight. Chris Arroyo, a regional director at Datadog, warns that the statistical nature of AI requires a fundamental change in how agencies monitor and authorize their technology.
Unlike typical software, which is deterministic and produces predictable results, AI is probabilistic. This distinction makes it difficult for authorizing officials to guarantee secure deployments, as outputs can vary. To address this, Arroyo proposes a "crawl, walk, run" framework designed to transition AI tools from isolated sandbox environments into mission-critical services. This approach emphasizes the need for deep visibility into both inputs and outputs to ensure accuracy and security.
The Infrastructure Gap
The transition is further complicated by technical debt. Datadog's "State of AI Engineering" report reveals that many teams struggle to build modern AI tools on top of outdated technology infrastructure. This gap often leads to inefficiencies in how models are deployed and managed. Arroyo suggests that agencies must be strategic about model selection, avoiding the use of "beefy" models for simple tasks like document analysis and continuously updating models to ensure they remain fit for their specific purpose.
Linking Costs to Mission Outcomes
Beyond security, observability serves as a critical tool for financial accountability. The high computational demands of AI can lead to sudden surges in infrastructure costs, such as increased cloud bills for powerful databases. Observability platforms allow agencies to tie these expenses directly to specific mission outcomes.
"You’re able to tell a story across a number of different constituencies and make them speak a common language so that they can identify that extra expense as being worth it to support the mission," Arroyo said.
The Path to Accountability
Without the ability to "look inside" a model to understand how it reaches a conclusion, government officials face a significant barrier to safe deployment. Establishing a data-driven path to confidence is necessary to prevent hallucinations and model drift, ensuring that AI remains a reliable tool for public service.
Ultimately, Arroyo emphasizes that technology cannot replace human oversight in the public sector. "People will always be in the lead; they will never be out of the loop because ultimately we’re responsible to the taxpayer," he stated. Moving forward, the focus for agencies will be on implementing these observability frameworks to maintain transparency while scaling AI capabilities.