By Senior Industry Analyst/Section Manager Rachel Liao
Over the past decade, smart city development has focused largely on sensors, big-data platforms, and visual dashboards, enabling city authorities to monitor traffic, air quality, flood risks, infrastructure, and other urban conditions in real time.
Yet an important gap remains. Cities may identify problems quicker, but interpreting abnormalities, coordinating across agencies, dispatching resources, and acting still depend heavily on human intervention.
As AI becomes increasingly embedded in public-sector processes, this model is beginning to change. The next stage of smart city development will not simply be about helping officials see problems, but whether AI can participate in — or, under certain conditions, carry out — parts of public-service operations.
This shift will also reshape the smart city market, with demand likely to move from platform construction toward the long-term operation of AI-enabled governance capabilities.
Taking action
Borrowing from autonomous-driving level classifications, urban AI capabilities can be broadly divided into four stages. At L1, digital assistance, cities use real-time sensing and visual dashboards to understand urban conditions. At L2, predictive optimization, AI supports scenario simulation and provides recommendations for decision-making.
At L3, conditional autonomy, AI is allowed to execute specific public services within clearly defined areas of authorization. At L4, agentic governance, AI can integrate functions across different domains and operate as a city-level AI system with learning capabilities.
The real challenge emerges when cities move from decision support to actual execution.
At that point, governments must decide whether AI agents can be authorized to perform certain public-service tasks independently under specific conditions. Three questions become critical: how far AI authority should extend, where accountability and responsibility should lie when errors occur, and which scenarios are appropriate for automated pilots.
Without clear mechanisms in these areas, technological capability alone cannot be translated into effective governance.
Unlike commercial processes, where efficiency may be the primary concern, digital public services must also account for social fairness and the rule of law. Errors in AI-based decision-making — such as incorrect case classification, delayed dispatch, mistaken eligibility decisions or opaque automated approvals — can trigger public complaints, media scrutiny, and legislative questioning, undermining public trust in government.
Emerging urban AI approaches
International cases point to two broad approaches: controlled integration and authorized execution.
Boston provides an example of controlled integration. In October 2025, the city introduced the OpenContext initiative, based on the Model Context Protocol, or MCP. The project created a controlled and structured interface between AI agents and city open data.
Municipal staff can use natural language to query city data and receive analytical results, while AI systems are prevented from directly issuing external commands. The aim is to allow AI to operate in a controlled environment, build experience, and create auditable records before its operational role is gradually expanded.
The European Union has taken a different but related approach centered on standards and digital-twin modeling.
The CitiVERSE alliance, established in 2024, uses the OpenUSD standard to integrate GIS, BIM, and real-time IoT data. This allows cities to simulate situations such as traffic diversion, flood risks, and urban heat-island effects in a virtual environment before policies are implemented.
Examples of authorized execution are taking place in Hangzhou and Abu Dhabi where they have moved toward the L3 conditional-autonomy stage.
Hangzhou’s City Brain 3.0 embeds AI agents into frontline processes including policing, health insurance, tourism, and social welfare. AI can support online applications, retrieve data across government agencies, trigger automated warnings, and connect with subsequent response procedures.
Hangzhou has also established a municipal data resources authority to serve as a cross-agency authorization center and assume institutional responsibility for AI operations, reducing the need for individual agencies to determine their own risk limits independently.
Abu Dhabi’s TAMM AutoGov follows another model built around an integrated government-service platform. Once users provide one-time authorization, the system can automatically handle routine administrative services such as license renewal, fee payment, and regular medical scheduling.
Abu Dhabi’s Department of Government Enablement serves as a single authorization window, while each government agency has a chief data and AI officer, extending the authorization framework into individual implementation units.
Together, these cases suggest that a clearly designated authorizing body is an important institutional condition for AI to operate across agencies and move into the execution stage.
Taiwan and urban AI
For Taiwan, the priority should not be to pursue comprehensive autonomous decision-making too quickly. A more practical path would be to build on existing smart city infrastructure, establish common rules for connecting AI agents to city data, and begin with low-risk, high-frequency scenarios in which data sources are clear and governance outcomes can be measured.
Taiwanese cities already have substantial digital assets, including 1999 citizen-service systems, disaster-prevention platforms, traffic-control centers, environmental sanitation systems, and public-facility maintenance systems. While these systems are not yet fully intelligent, they provide starting points for AI integration.
Priority could therefore be given to public services involving limited administrative discretion and clearly defined operational processes.
Disaster prevention is one possible area. AI could integrate hydrological information, water levels, flooding hotspots, historical damage records, and real-time reports to support risk classification, warning prioritization, and recommendations for resource dispatch.
Public-facility management is another. By combining 1999 citizen reports with image-recognition technologies, AI could automatically classify cases such as potholes or streetlight failures and help determine repair priorities.
Through these applications, cities could gradually establish cross-agency authorization structures, define the boundaries of AI execution and clarify exception-handling rules, providing an auditable foundation for future conditional automated execution.
Shift toward AI governance
The development of urban AI will also reshape competition in the smart city market. In the past, business opportunities were concentrated mainly on sensing hardware, communications networks, data platforms and system integration.
As AI becomes more deeply embedded in public governance processes, market demand is expected to shift from building platforms to continuously operating AI governance capabilities.
For companies, three areas of opportunity stand out.
The first is data governance and AI integration, including data standards, access control, access records, traceability of model outputs, and personal-data protection.
The second is scenario-specific AI agents and customized workflow modules, such as disaster-warning systems, public-facility dispatch, intelligent classification of 1999 cases, environmental sanitation scheduling, and traffic-congestion prediction.
The third is AI governance auditing and long-term operation, including model-risk assessment, cybersecurity audits, bias detection, and model adjustment.
Future competitive smart city suppliers will therefore need to take on a broader role. Rather than focusing only on individual systems or platforms, they will need to help governments connect data, models, operational processes, authorization mechanisms, and auditing requirements into sustainable AI governance capabilities.
The next stage of smart city development is therefore not simply about deploying more AI. It is about creating the institutional and operational conditions that allow AI to move carefully from observation and recommendation towards clearly authorized forms of execution.
