Public Sector Accelerates Agentic AI Deployment Despite Oversight and Data Gaps
A Capgemini survey of 600 senior officials shows more than half of public sector organizations are exploring or rolling out agentic AI, even as governance and data readiness struggle to keep pace. Experts urge agencies to focus on low-risk internal workflows to safeguard public trust.
Governments worldwide are shifting from experimental AI initiatives to full-scale deployments, driven by growing public expectations for faster services and measurable results. However, institutional readiness and oversight continue to lag behind technical ambition, according to recent findings from the Capgemini Research Institute.
Based on a survey of 600 senior public-sector executives, the study revealed that just over half of public-sector organizations are currently exploring, piloting, or deploying agentic AI. At the same time, fewer than a quarter of respondents reported that their responsible AI practices are keeping pace with their deployment speed.
Legacy Infrastructure and Data Bottlenecks
Industry experts warn that the gap stems from longstanding technical deficits within government administrations. Marc Reinhardt, global industry leader for the public sector at Capgemini, noted that public agencies have struggled for years with outdated infrastructure, legacy systems, interoperability, and governance. With the rise of agentic AI—which requires broad access to diverse databases and systems—modernizing these foundations has become urgent.
Reinhardt identified the absence of solid data foundations as the single biggest obstacle to scaling agentic AI. Furthermore, many legacy systems were never designed to be accessed via APIs, preventing autonomous software agents from interacting with them seamlessly.
Laura Gilbert, head of the AI for Government program at the London-based Tony Blair Institute for Global Change, reinforced that underlying systems, rather than the core models, are the primary bottleneck. She also emphasized a widening "capability and confidence" gap across public bodies, noting that heavy reliance on outsourced technical expertise leaves civil servants ill-equipped to evaluate safety, assess feasibility, or establish accountability when software agents take autonomous actions.
Mitigating Risk with "Two-Way Door" Workflows
To bridge the gap between innovation and safety, experts advise governments to establish clear guardrails detailing what decisions an AI agent can execute independently, what data it may access, and when it must hand control back to a human operator.
Early implementations are best targeted at repetitive, high-volume internal tasks where mistakes carry minimal consequences for the public. Reinhardt pointed to administrative workflows, document processing, case management, routine approvals, and internal IT processes as sensible starting points.
Gilbert similarly advocated for low-risk environments such as triaging incoming cases, managing correspondence, and summarizing evidence. These tasks offer visible time savings while functioning as "two-way doors," where errors can be easily undone without damaging public trust or directly altering outcomes for citizens.
Beyond operational guardrails, experts agree that agencies must develop small, highly capable in-house technical teams positioned close to decision-makers to test, challenge, and govern AI systems as deployments widen.
