信任要务5.0：依托实用型 AI 保障构建政府公信力 (2026)

Building Trust in Government Through PracticalAIAssurance 2026 Global Report ExecutiveSummary Artificial intelligence (AI) has the potential to helpgovernments improve service quality, achievebetter outcomes, raise public sector productivity,and build greater trust in government. Acrosscountries, governments have responded to advancesin AI capabilities by introducing governance andassurance frameworks, principles, risk assessmenttools, transparency requirements and registers, andprocurement controls. Together, these measures areintended to support responsible AI adoption whilehelping organizations manage risks associated withAI. But how effectively do they work in practice?What works well, what isn’t working, and whatlessons can be drawn from the experience of realpractitioners sofar? These challenges may grow even more acutebecause many current AI frameworks were designedbefore the emergence of today’s advanced frontiermodels and agentic AI use cases. These tools aremore capable and accurate, but they are also moreautonomous and interconnected, with sharedresponsibility across model providers, platforms,partners, and agencies. This situation leads tofriction across risk assessment, procurement,deployment, and oversight. It also shifts thegovernance question. Particularly for agenticsystems, the issue is not just whether a modelis technically safe, but what authority is beingdelegated, which responsibilities must remain witha human, and how accountability is maintained assystems act within definedlimits. Across the ten countries analyzed for this report,we found that many government organizations aregrappling with common challenges: how to classifyrisk, what evidence is sufficient, who owns thedecision, and what “safe enough” means in differentAI use cases. When the answers to these questionsare unclear, AI assurance processes often lead todelays, rework, andduplication. Rather than adding another layer of policycomplexity, governments should focus on makingexisting assurance more usable. The goal is tomaximize public value while staying disciplinedenough to manage risk and scale AI responsibly.The opportunity cost of delay is significant.BCG estimates that GenAI could unlock $1.75trillion in annual productivity benefitsby 2033for governmentsglobally, but realizing that valuedepends on adoption proceeding without high-profile failures that erode public trust and triggerregulatorybacklash. BCG research finds that while most governmentshave established foundational principles andframeworks for working with AI, putting them intopractice is far more challenging. Many governmentshave named accountabilities, tiered risk categories,practical transparency tools, and strong governancestructures. But risk thresholds are often too vagueto apply consistently. Accountable roles exist, butnot always with clear mandates or decision rights.Processes remain fragmented and duplicative, andinternal capabilities at the points of deploymentand usage can be uneven. As a result, many lower-risk productivity use cases are subject to overgovernance, while more complex questions involvingdeployment, platforms, and shared responsibilityremainunresolved. To navigate this transition effectively, governmentscan make risk triage more proportionate, clarifyaccountability, redesign assurance for generativeAI (GenAI) and agentic systems, embed assuranceinto delivery, create reusable artifacts, strengthencapability, and measure whether assurance ishelping manage both risk and opportunity. Withthis approach, assurance can become an enabler ofresponsible AI adoption atscale. Governments thatget this balance right will be better positioned toadopt AI confidently while enhancing public trust intheprocess. 01. Balancing the AI OpportunityandRisks Effective assurance is intended to help governmentscapture the value of AI while maintaining publictrust and accountability. In many government andpublic sector agencies, AI ethics principles have beenpublished, procurement guidance has been updated,and various forms of impact assessment, transparencyrecording, or risk review have been introduced. However,the public sector is lagging other industry sectors interms of AI adoption and AI use cases. Are governmentsbecoming overly conservative in their assessment ofrisk, limiting productivity gains and delaying service improvements? And are current assurance frameworksstriking the right balance between managing risk andenabling public value? Is this really a trade-off or canyou haveboth? This report examines the evolution of AI risk, assurance,and governance frameworks in government andwhether they are effective and efficient in practice. Italso highlights what is working well in AI assurance andwhere further refinement could improveoutcomes. Key Focus of this Report 01.02.03.04.05.What is the purpose of AI principles and assurance frameworks and are theyachieving their intended objectives?What has been the experience of people in government w