⚡ Quick Summary
Published by the World Economic Forum in collaboration with Capgemini, this April 2026 insight report sets out a readiness framework for agentic AI in government. It argues that the right unit of analysis is the function – the recurring workflows that cut across organizational boundaries, such as “cybersecurity monitoring”, “eligibility assessment” or “fraud detection” – rather than departments, tools or isolated use cases.
The framework has four building blocks. It catalogues 70 core government functions in nine categories; scores them on agentic AI potential (three sub-criteria) and implementation complexity (six sub-criteria, split between implementation challenges and risk and ethical impact) on a one-to-four scale; calculates a readiness score as potential minus complexity; and maps the results into three areas with thresholds above 1.5 (high), 1.0 or higher (medium) and below 1.0 (low). Ten functions fall in the high-readiness area, 25 in medium and 35 in low, and the report states that 50% of functions combine significant agentic AI potential with manageable implementation complexity.
It concludes with public-sector use cases from Ukraine, Germany and the United Arab Emirates, and a six-step path from global topography to local roadmap: establish local baselines, develop risk-mitigation strategies, reassess function-level scores, sequence implementation, validate through testing and iterate.
🧩 What’s Covered
The report moves from the agentic opportunity to a scored topography and local roadmaps.
- The agentic opportunity (Section 1): Defines agentic AI as the coordinated use of agents with bounded autonomy, contrasts it with robotic process automation and generative AI across focus, output, autonomy, learning, interactivity and capability, and cites a Capgemini survey in which 90% of 350 public-sector organizations plan to explore or deploy agentic AI within two to three years.
- Function-based assessment lens (2.1): Argues that the unit of analysis should be a function – a recurring, outcome-oriented activity spanning one or more end-to-end workflows – not a department, tool or use case, and describes the inventory of 70 functions in nine categories, built on criteria such as recurring execution and a legal mandate.
- Assessment of potential against complexity (2.2): Sets out the scoring model, with three potential sub-criteria (potential for automation, agent requirement, volume and impact) and six complexity sub-criteria covering implementation challenges and risk and ethical impact, each scored one to four and aggregated into composite ratings.
- Topography of government readiness (2.3): Presents readiness as potential minus complexity with three areas, lists the ten high-, 25 medium- and 35 low-readiness functions, and draws five analytical patterns, including prioritizing “public services” functions and targeting fully automatable ones.
- From global topography to regional roadmap (2.4): Lists nine baseline factors – digital maturity, data infrastructure, cloud readiness, workforce capabilities, funding, political mandate, legal clarity, citizen acceptance and the govtech ecosystem – risk guidance on sustainability, workforce readiness, value alignment, infrastructure resilience and explainability, and six steps from baseline to iteration.
- Learning from successful deployments (Section 3): Summarizes four public-sector use cases, including Ukraine’s Diia.AI assistant, Germany’s AI-based construction permit system, the United Arab Emirates’ HR AI agent and the German Federal Employment Agency’s Jira ticket agent.
- Appendices (A1, A2): Describe the methodology (literature review, expert consultation, coding scheme, expert scoring) and reproduce Table 2 with potential, implementation challenge and risk and ethical impact scores for all 70 functions.
💡 Why it matters?
The framework gives institutions a defensible way to decide where agentic AI should be deployed first. By scoring functions on potential and complexity rather than departmental enthusiasm, it turns “where do we start?” into a comparable assessment that supports sequencing, capability building and risk mitigation. Its guidance on bounded autonomy, human escalation, explainability and audit trails, and its insistence on reassessing global scores against local conditions, connects directly to the accountability expectations that oversight, audit and assurance functions must meet. The report also relates its approach to COFOG, O*NET, the GovTech Maturity Index and the ESOAR process-transformation methodology.
❓ What’s Missing
The scores are meso-level global estimates; the report states that its findings have limited direct applicability to specific local contexts and that individual countries will score higher or lower. It provides no country-level results, cost model, procurement guidance or legal analysis for any jurisdiction, and stays at the level of prioritization rather than implementation detail. The use cases are short descriptive summaries without an evaluation methodology. The authors acknowledge that the inventory deliberately covers core recurring workflows rather than all government activity, and that functions scoring low may still contain automatable subprocesses.
👥 Best For
Best for strategy, digital-transformation and AI governance teams in public administration deciding where to pilot agentic AI; for policy advisers building national roadmaps from the six-step process; and for auditors and oversight bodies wanting a function-level view of where automation potential, complexity and ethical risk combine.
📄 Source Details
Making Agentic AI Work for Government: A Readiness Framework is an insight report published by the World Economic Forum in collaboration with Capgemini, dated April 2026 and running to 40 pages, with a foreword, four numbered sections, a conclusion, appendices A1 and A2 and endnotes. Contributors are listed from the World Economic Forum, the Global Government Technology Centre Berlin and Capgemini. No ISBN, DOI or URL for the report itself is printed. The supplied text extraction covers all 40 pages, though figure labels are only partially legible.
