The Four Governance Conditions That Determine Whether Urban AI Works

Artificial intelligence (AI) is entering the operating systems of cities, influencing how governments manage electricity, traffic, climate risks, environmental permits and public spending. However, the most consequential question is whether public institutions can govern those decisions in ways that are lawful, transparent, adaptable and fair, argues a new study published in the journal Sustainability.

Titled "AI Governance, Urban Sustainability Transitions, and the Institutional Determinants of Smart City Transformation: A Comparative Cross-City Framework," the study by Diego Navarra of EU Business School and StudioNavarra is based on 80 primary policy documents from Singapore, Amsterdam, Barcelona, Seoul and Toronto, alongside a broader evidence base of 170 sources. It concludes that governance design, not technological capability, is the primary determinant of whether urban AI produces sustainable transformation.

The research identifies four conditions that cities need to convert AI into public value: regulatory coherence, meaningful participation, transparency and auditability, and the institutional capacity to modify or suspend systems when they fail. Its warning is clear: cities can automate quickly while becoming no more sustainable, accountable or inclusive.

The Smart-City Race Is Measuring the Wrong Kind of Intelligence

Cities have powerful reasons to use AI. Urban governments must manage congestion, pollution, energy demand, infrastructure stress, extreme weather and rapidly expanding populations, often with limited budgets and fragmented administrative systems. AI can help balance electricity demand, optimise transport routes, analyse environmental data and accelerate administrative processes. But improved technical performance does not automatically produce better governance.

A traffic algorithm may shorten average journey times while redirecting vehicles and pollution into poorer neighbourhoods. A climate-risk model may identify flood-prone areas but overlook informal settlements that are weakly represented in official records. An automated procurement platform may reduce prices while ignoring environmental commitments that were never built into its decision criteria.

The study rejects the idea that AI acts independently of institutions. Algorithms operate through laws, administrative structures, political priorities and existing inequalities. The same technology may generate public value in one city and reinforce exclusion or fragmentation in another.

Navarra's analysis began with about 200 candidate documents published between 2015 and 2024. After screening for relevance, institutional authority and date, 80 were retained: 20 from Singapore and 15 each from Amsterdam, Barcelona, Seoul and Toronto. They included government strategies, regulatory guidance, programme reports, audits and civil-society assessments.

The thematic analysis generated 847 initial codes, reduced to 23 candidate themes and then 11 substantive themes. These were organised into four institutional conditions. An independent coder reviewed 20% of the material, producing a Cohen's kappa score of 0.81, indicating strong agreement.

The methodology cannot establish that a particular governance arrangement directly caused lower emissions or better services. It does, however, expose a consistent institutional pattern: advanced technology produces limited public value when cities lack the rules, participation mechanisms and oversight capacity needed to govern it.

Four Tests Separate Real Transformation from Digital Theatre

The first test is regulatory coherence. AI systems must be aligned with environmental law, data protection, procurement obligations and sector-specific rules before deployment. This condition appeared across all five cities and in 83% of governance strategy documents. Singapore provided one of the clearest models, treating legal and environmental requirements as design constraints rather than compliance issues to be addressed after implementation.

When regulatory alignment is missing, AI outputs may become unusable. Systems may rely on data that cannot legally support public decisions. Procurement algorithms may overlook sustainability standards. Environmental predictions may lack the legal recognition needed to trigger government action.

The second test is multi-stakeholder participation. The study distinguishes between consultations that collect views and participation that genuinely changes system design. Barcelona's community engagement processes helped identify neighbourhood-level energy poverty and air-quality concerns that top-down analysis had missed. Seoul created multi-stakeholder boards with authority to approve or suspend some AI applications.

Participation is especially important because urban datasets are incomplete. Informal communities, digitally excluded residents and poorly monitored districts may be absent from the information used to allocate resources. Without their involvement, AI risks optimising services for people who are easiest to measure rather than those most vulnerable to policy failure.

The third test is transparency and auditability. The study found these concerns in 91% of civil-society and regulatory evaluation documents but only 47% of government strategy documents, revealing a gap between public expectations and official planning. Transparency can take several forms: publishing design criteria before deployment, explaining individual decisions and allowing independent audits. Seoul and Singapore showed stronger pre-deployment transparency, while Amsterdam developed more extensive post-decision explanations and external review.

The fourth test is adaptive governance capacity - the authority and ability to retrain, modify or suspend systems as conditions change. Amsterdam requires periodic audits and incorporates citizen feedback into model retraining. Of 23 public-service AI applications assessed in its first audit cycle, seven required significant modification and two were suspended.

These findings show why one-time approval is inadequate. Data quality, public priorities, climate science and algorithmic behaviour evolve. Urban AI therefore needs continuous oversight rather than permanent institutional trust.

AI Works Best Where Goals Are Clear and Worst Where Justice Is at Stake

The study compares governance maturity across energy management, mobility, climate adaptation and green public procurement. Energy and mobility perform best because they have longer regulatory histories, clearer performance indicators and stronger operational infrastructure.

Documents reviewed in the study reported energy-efficiency improvements of between 8% and 23% in AI-supported demand-management programmes. Journey times reportedly fell by 11% to 18% in Seoul and Singapore, while public-transport punctuality improved by 9% to 14% in Amsterdam and Singapore.

These are programme results reported in the analysed documents, not causal estimates independently produced by the study. Nevertheless, they indicate why AI advances faster in sectors where objectives can be clearly measured and responsibility is more firmly assigned.

Climate adaptation is less mature. Community participation remains weak, model uncertainty is poorly communicated and regulatory frameworks often lag behind predictive technologies. This weakness has distributional consequences. People most exposed to floods, heatwaves and infrastructure failure may also be least visible in official data. An AI system trained without meaningful community input can reproduce existing vulnerability in the very tools intended to reduce it.

Green public procurement is the most serious gap. The study rates all four governance conditions as weak in this domain. AI-supported procurement represented an estimated 2% to 7% of procurement volume across the five cities, yet sustainability criteria appeared in fewer than 15% of automated procurement processes.

Governments can use procurement to support low-carbon materials, circular production, cleaner transport and responsible labour practices. If automated systems prioritise cost and speed while excluding sustainability criteria, AI may make procurement more efficient while quietly weakening environmental policy.

The implications are particularly important for developing countries. Many Global South cities face severe climate exposure, fast informal expansion, incomplete datasets and limited administrative capacity. Yet these cities are under-represented in the study, making it uncertain whether the framework will operate in the same way across lower-income and less institutionally resourced settings.

Accountability Must Be Built In Before the Algorithm Goes Live

Advanced AI systems often cannot explain their internal reasoning in clear human terms. This creates an audit problem for public decisions affecting rights, services, resources and environmental outcomes. The study considers whether blockchain and distributed ledger technologies can strengthen accountability. They cannot explain why an opaque AI reached a decision, but they can preserve a tamper-resistant record of the data received, output produced and time of action.

South Tyrol's environmental-permitting system combined blockchain records with AI-supported document analysis and compliance checking. Reported outcomes included a reduction in processing time from 120 to 45 days, an 85% decline in documentation errors and a 40% reduction in administrative costs. Citizen satisfaction reportedly increased from 65% to 90%.

More importantly, the audit record helped officials identify systematic biases in automated environmental assessments and provided evidence for retraining the system.

Estonia offers a broader example. Its X-Road platform connects 99% of public institutions and processes more than two billion transactions annually. Its blockchain-based integrity layer timestamps government data and AI outputs, enabling retrospective verification of whether the underlying information had been altered.

However, blockchain is not a substitute for governance. An immutable record has little value if no institution has the authority, expertise or willingness to examine it and act on the findings. Cities therefore need accountability requirements at the design stage. Governments should specify which data and decisions must be recorded, who can inspect them, how long records must be retained and what remedies are available to affected citizens.

The authors acknowledge several limitations. Its five-city sample is purposive rather than representative, and cities in Africa, Latin America and lower-income Asia are largely absent. Policy documents may overstate actual implementation, while the main AI-blockchain cases are European. The qualitative design also cannot prove that the four governance conditions directly cause sustainability improvements.

Regardless, the study exposes the central flaw in technology-led smart-city strategies. More sophisticated algorithms cannot compensate for fragmented regulation, weak participation or absent accountability.