Case study: AI-powered carbon accounting & MRV — a city or utility pilot and the results so far

In September 2024, the Seoul Metropolitan Government launched the Seoul Carbon Intelligence Platform (SCIP), a city-wide AI-powered carbon accounting and measurement, reporting, and verification (MRV) system designed to provide near real-time emissions tracking across the metropolitan area's 25 autonomous districts. The pilot represented one of the most ambitious municipal deployments of AI-driven emissions monitoring in the Asia-Pacific region, covering approximately 9.7 million residents and 680,000 commercial and industrial entities. Eighteen months into operation, the platform has generated sufficient data to assess what works, what fell short of expectations, and what other cities and utilities can learn from Seoul's experience.

Why It Matters

Municipal carbon accounting has historically relied on annual greenhouse gas inventories compiled from utility billing records, transportation surveys, and industrial reporting. These retrospective approaches produce data that is typically 12-18 months old by the time it reaches decision-makers, rendering it effectively useless for operational emissions management. The Global Covenant of Mayors for Climate and Energy reports that only 23% of signatory cities in the Asia-Pacific region produce emissions inventories more frequently than every two years. The gap between reporting cadence and policy requirements is widening as national governments implement increasingly stringent climate disclosure mandates.

South Korea's Carbon Neutrality and Green Growth Framework Act, enacted in 2021 and amended in 2025, requires all metropolitan cities and provinces to maintain annual carbon budgets and report quarterly progress. The amendment introduced mandatory sectoral emissions tracking at the district level, creating a compliance requirement that traditional manual inventory methods cannot practically satisfy. Japan's revised Act on Promotion of Global Warming Countermeasures imposes similar quarterly reporting for designated cities. Singapore's carbon tax framework, expanded in 2024, requires facility-level emissions verification for installations exceeding 2,000 tonnes of CO2 equivalent annually.

The financial implications are substantial. The Asian Development Bank estimates that Asia-Pacific cities will need to invest $26 trillion in climate infrastructure by 2030, with credible emissions data serving as the prerequisite for accessing green bond markets and concessional climate finance. Cities that cannot demonstrate verifiable emissions reductions face exclusion from the rapidly growing market for green municipal bonds, which reached $47 billion in issuance across Asia-Pacific in 2025.

Design Choices and Architecture

Data Integration Layer

SCIP integrates data from 14 distinct source systems across six municipal departments. The electricity consumption data feeds from Korea Electric Power Corporation (KEPCO) cover 3.2 million metering points, transmitted at hourly intervals. Seoul City Gas provides natural gas consumption data for 1.8 million connections. The Seoul Transport Operation and Information Service (TOPIS) contributes real-time traffic flow data from 18,000 sensors, 42,000 bus GPS units, and 430 subway station monitoring points. Industrial emissions data comes from the Korea Environment Corporation's stack monitoring telemetry, covering 2,340 facilities with continuous emissions monitoring systems (CEMS).

The integration challenge was not data availability but harmonisation. KEPCO data uses a billing-period structure that does not align with calendar months. Gas consumption data arrives with 72-hour latency. Traffic data streams at five-minute intervals but requires aggregation to meaningful emissions factors. The platform engineering team, led by Korea Advanced Institute of Science and Technology (KAIST) researchers in collaboration with SK Telecom's AI division, spent seven months building the data harmonisation layer before any AI modelling could begin.

AI Modelling Approach

The platform employs a hybrid architecture combining bottom-up facility-level models with top-down atmospheric verification. The bottom-up component uses gradient-boosting machine learning models trained on three years of historical facility data to estimate emissions from electricity and gas consumption, applying time-varying grid emission factors that account for South Korea's fluctuating renewable generation share. For the transportation sector, a graph neural network models vehicle emissions across Seoul's road network using traffic flow, speed, and fleet composition data.

The atmospheric verification layer, funded through a separate Korea Meteorological Administration grant, ingests data from 12 ground-based atmospheric monitoring stations and two satellite data streams (Japan's GOSAT-GW and the European Copernicus Atmosphere Monitoring Service). Machine learning reconciliation algorithms compare bottom-up emissions estimates against atmospheric concentration measurements, flagging discrepancies exceeding 15% for investigation.

Verification and Quality Assurance

The MRV framework follows GHG Protocol City Standard (GPC) methodology, with AI-assisted verification replacing portions of the traditional manual audit process. The system generates automated anomaly detection alerts when facility-level emissions deviate by more than two standard deviations from historical baselines or sector averages. Human auditors from the Korea Environment Corporation review flagged anomalies and conduct physical inspections for a stratified random sample of 5% of reporting entities quarterly.

Measured Outcomes

Reporting Speed and Coverage

The most immediate measurable benefit has been the transformation of reporting cadence. Seoul now produces district-level emissions estimates within 14 days of month-end, compared to the previous 14-month lag for annual inventories. Coverage expanded from approximately 60% of actual emissions (estimated from sampled data) under the old methodology to 89% under SCIP, with the remaining 11% consisting primarily of fugitive emissions and small-scale combustion sources below monitoring thresholds.

The quarterly reporting requirement under the amended Carbon Neutrality Act was met on schedule for Q3 and Q4 2025, making Seoul the first Korean metropolitan city to achieve compliance. The reporting process required approximately 120 person-hours per quarter, compared to an estimated 2,800 person-hours for equivalent manual compilation.

Emissions Identification and Reduction

SCIP identified 847 anomalous emissions patterns in its first 12 months of operation. Of these, 312 were confirmed as genuine emissions exceedances requiring intervention: 186 related to commercial building HVAC systems operating outside design parameters, 74 to industrial process deviations, and 52 to transportation infrastructure (primarily idling patterns at freight logistics hubs). The remaining 535 were attributed to data quality issues, sensor malfunctions, or seasonal variations within normal operating ranges.

The confirmed anomalies, once communicated to facility operators and district authorities, resulted in an estimated 340,000 tonnes of CO2 equivalent in annualised emissions reductions. This figure represents approximately 0.8% of Seoul's total annual emissions of 43.7 million tonnes. While modest in percentage terms, the reductions were achieved at near-zero marginal abatement cost, as they involved correcting operational inefficiencies rather than requiring capital investment.

Cost Analysis

Total implementation cost for SCIP was KRW 28.7 billion (approximately $21.5 million), comprising KRW 11.2 billion for platform development and data integration, KRW 8.9 billion for AI model development and training, KRW 5.1 billion for sensor infrastructure upgrades, and KRW 3.5 billion for project management and stakeholder engagement. Ongoing annual operating costs are approximately KRW 4.2 billion ($3.1 million), covering cloud computing infrastructure, data licensing, model maintenance, and a platform operations team of 18 staff.

The cost per tonne of identified emissions reduction in year one was approximately $63, comparable to many conventional abatement measures. However, this metric understates the platform's value, as the primary benefit is not direct emissions reduction but the enabling of informed policy decisions, green finance access, and regulatory compliance that would be impossible without granular, timely emissions data.

What Did Not Work as Expected

Industrial Sector Coverage Gaps

SCIP's industrial emissions modelling performed significantly below expectations for small and medium enterprises (SMEs), which constitute 94% of Seoul's commercial and industrial entities but account for only 31% of total emissions. SMEs lack the metering infrastructure, operational data systems, and technical capacity to participate in automated data collection. The platform relies on statistical estimation for SME emissions, applying sector-average intensity factors that introduce substantial uncertainty. Validation studies found that SME emissions estimates carried error margins of 35-50%, compared to 8-12% for large facilities with CEMS.

The Seoul Metropolitan Government has allocated an additional KRW 6.5 billion for a phased smart metering programme targeting 15,000 high-priority SME facilities, but full coverage remains 3-4 years away. Other cities considering similar platforms should anticipate that the "long tail" of small emitters will require pragmatic approaches that accept lower accuracy rather than attempting comprehensive coverage from the outset.

Atmospheric Verification Limitations

The top-down atmospheric verification layer, initially promoted as the system's key innovation, has proven less operationally useful than projected. Seoul's dense urban environment creates complex atmospheric mixing patterns that confound source attribution. The 12 monitoring stations provide insufficient spatial resolution to attribute concentration anomalies to specific districts or facilities. Satellite data, while improving, provides column-integrated measurements that struggle to distinguish surface-level urban emissions from background concentrations.

The verification layer successfully identifies city-wide emissions trends and seasonal patterns, but its ability to validate district-level estimates remains limited to detecting discrepancies exceeding 20-25%, which is too coarse for regulatory verification purposes. The KAIST research team is developing higher-resolution inverse modelling approaches, but operational deployment is not expected before 2028.

Data Governance and Privacy Friction

Facility-level emissions data intersects with commercially sensitive information about production volumes, operating hours, and process efficiency. Several industry associations challenged SCIP's data collection authority under South Korea's Personal Information Protection Act, arguing that granular energy consumption data constitutes business intelligence. The Seoul Metropolitan Government secured a regulatory determination that aggregated emissions data published at the district level does not constitute personally identifiable information, but facility-level data remains restricted to regulatory use only.

This constraint limits the platform's utility for supply chain emissions reporting, where corporate buyers increasingly require facility-level Scope 3 data from their Korean suppliers. The tension between municipal emissions transparency and corporate data protection remains unresolved and represents a significant barrier for cities seeking to extend AI MRV platforms into supply chain applications.

Transferable Lessons

Lesson 1: Data Integration Consumes Most of the Budget and Timeline

SCIP's experience confirms that 55-60% of implementation effort and cost goes to data acquisition, harmonisation, and quality assurance rather than AI model development. Cities planning AI MRV deployments should allocate resources accordingly and resist the temptation to begin model development before data pipelines are stable and validated. The seven-month data integration phase, initially seen as a delay, proved essential for model reliability.

Lesson 2: Start with High-Confidence Sectors

Electricity and natural gas consumption, where metering data is comprehensive and emission factors are well-established, provided the foundation for SCIP's credibility. Transport emissions modelling, while less precise, added policy-relevant insights about spatial and temporal patterns. Industrial process emissions and waste sector modelling, which require facility-specific data that is often unavailable, should be treated as second-phase priorities rather than launch requirements.

Lesson 3: Human Verification Remains Essential

Despite AI's ability to process data at scale, the platform's credibility depends on human auditors who investigate anomalies, conduct physical inspections, and validate model outputs against ground truth. The 5% quarterly physical inspection rate provides statistical confidence that AI-generated estimates are within acceptable error bounds. Fully automated MRV without human oversight is not yet credible for regulatory compliance purposes.

Lesson 4: Set Realistic Accuracy Expectations by Sector

SCIP's experience suggests that AI-powered MRV can achieve 8-12% accuracy for large, well-metered facilities; 15-25% accuracy for the transportation sector; and 35-50% accuracy for small commercial and industrial entities without dedicated metering. Cities should communicate these accuracy tiers transparently rather than claiming uniform precision that the technology cannot yet deliver.

Action Checklist

• Assess existing municipal data infrastructure before scoping an AI MRV platform, focusing on metering coverage, data latency, and format standardisation
• Allocate 55-60% of implementation budget to data integration and quality assurance rather than AI model development
• Begin with electricity and gas sectors where metering data is comprehensive and emission factors are reliable
• Establish data governance frameworks that address commercial sensitivity before collecting facility-level consumption data
• Maintain human verification processes alongside AI-generated estimates, targeting 5% physical inspection rates for regulatory credibility
• Set sector-specific accuracy targets rather than claiming uniform precision across all emissions categories
• Plan for 3-5 year phased deployment rather than attempting comprehensive coverage at launch
• Engage industry associations and data providers early to resolve data-sharing agreements before platform development begins

Sources

• Seoul Metropolitan Government. (2025). Seoul Carbon Intelligence Platform: First Year Operational Report. Seoul: SMG Climate Policy Division.
• Asian Development Bank. (2025). Climate Finance in Asia-Pacific Cities: Data Requirements and Investment Flows. Manila: ADB.
• Korea Advanced Institute of Science and Technology. (2025). Hybrid AI Architecture for Urban Emissions Estimation: Technical Validation Study. Daejeon: KAIST.
• Global Covenant of Mayors for Climate and Energy. (2025). Asia-Pacific City Emissions Reporting: Compliance and Capacity Assessment. Brussels: GCoM Secretariat.
• Korea Environment Corporation. (2025). Continuous Emissions Monitoring and AI-Assisted Verification: Seoul Pilot Results. Incheon: KECO.
• World Resources Institute. (2025). GHG Protocol for Cities: AI-Enhanced Implementation Guidance. Washington, DC: WRI.