알림 목록

Analysis Area: Busan Metropolitan City

Core Area: Busan Metropolitan City Headquarters, 16 Districts and Counties, and Affiliated Public Institutions

Agenda: Verify whether AI and data have gone beyond administrative assistance and actually changed policy alternatives, budgets, implementation, and performance evaluations

Golden Time Type: Administrative AI Expansion + Decision Evidence GAP

Reference Date: 2026.08.28

Version: Regional AX Golden Time Intelligence v3.2 Enhancement Standard
 

At the end of 2025, Busan City became the first local government to launch generative AI administrative services based on small language models and search augmentation, and by early 2026, it had applied 22 types of services—including Q&A, document drafting, and administrative data search—to all departments. The scope of AI application in administration has shifted from individual employees' use of external generative AI to organizational-level services that combine internal administrative data with security systems. However, public decision logs showing that service usage led to policy alternative comparisons, budget changes, project suspensions, target area adjustments, or changes in implementation methods have not been verified. If the three-year enhancement plan from 2026 to 2028 proceeds primarily based on usage volume and the number of generated documents, administrative AI will become stagnant as merely an improvement in work speed. While Busan's AI administration has reached the stage of use across all departments, there is a lack of evidence to conclude that it has reached the stage of changing decision-making processes.

Actual policy questions, such as changes in sales following the conversion of mandatory hypermarket closure days to weekdays, credit card sales before and after the "15-Minute City" project, festival visitor numbers and commercial district sales, and parking shortages in Yeongdo-gu, have been registered as analysis tasks in Busan's Big Data Wave. Data utilization is shifting from statistical inquiries to verifying outcomes before and after policy implementation. However, publicly available follow-up records regarding which alternatives were excluded and which decisions, budgets, or executions were altered are limited. If analysis requests and policy decisions remain in a separate system for the next two to three years, the volume of analysis may increase, but the impact of decision-making cannot be measured. In  Busan's data-driven administration, while the creation of questions and the execution of analyses are confirmed, the traceability of decision changes is deemed unconfirmed.

The 2026 AI Administration Promotion Plan proposed the widespread adoption of AI across the entire administration, while the work plan for the data sector included expanded analysis, employee training, and the utilization of pseudonymized information to address municipal challenges. The digitalization of administration has shifted from partial pilot projects to a plan for widespread adoption across all sectors. Conversely, a public baseline linking departmental data quality, AI response accuracy, user verification behavior, error occurrence, and citizen outcomes into a single performance framework has not been identified. If services expand rapidly between 2026 and 2028, low-quality raw data and unverified responses will be repeatedly inserted into work procedures.  The time-sensitive risk for Busan's administrative AX lies in the fact that the speed of AI adoption is outpacing the speed of evidence quality and the establishment of accountability systems.

Decision support projects across various sectors, such as water supply network simulations, Smart Ocean Villages, and AI-based audit administration, are also underway. The application of AI is expanding from common document tasks to expert judgments in water supply, marine environment, auditing, and safety. However, there is no publicly available comparative data confirming how much predictive warnings have changed actual inspection priorities, budget allocations, and regulatory and audit judgments, or whether results have improved. If sector-specific systems are independently advanced over the next two to three years, Busan City will not possess a unified administrative learning system, even if it holds multiple AI services.  Currently, the "Golden Time" is assessed not as the period for expanding AI services, but as the two to three years leading up to the first connection between decisions changed by AI and their outcomes.

Golden Time Thesis — The unit of judgment for Busan Administrative AX is not the number of AI services or the number of generated documents. The unit of judgment is the Decision Runtime, which extends to whether AI and data created a choice different from existing decisions and whether that choice improved the outcomes for citizens, businesses, and the region.

Busan City's administration manages transportation, welfare, safety, industry, environment, tourism, and urban planning through the main office, business units, 16 districts and counties, and public corporations, agencies, and invested or contributed institutions. Administrative data has a distributed structure, generated by multiple organizations and entrusted agencies rather than a single entity. The extent to which the same citizens, businesses, spaces, or projects are managed by various departments using different IDs and update cycles cannot be fully verified in publicly available data. As departmental AI services expand on this distributed data between 2026 and 2028, different responses and risk assessments may arise regarding the same issues.  The scale risk of Busan's administration is determined by multiple records of the same administrative subject rather than a lack of data.

Big-DataWave provides public and private data, statistics, analysis consulting, and data utilization cases. The data foundation of Busan's administration has expanded from periodic statistics to policy analysis utilizing credit card sales, telecommunications population mobility, spatial grids, and surveys. However, publicly available operational indicators linking the total number of registered analysis requests, average processing time, policy reflection rate, and reuse rate are not available. While waiting times and duplication of individual analyses accumulate as request volumes increase over two to three years, the organization's overall analysis capacity remains unassessed. Although  Busan's data administration possesses an analysis infrastructure, its processing capacity relative to demand remains unmeasured.

Busan's generative AI administrative services provide 22 types of functions, including question-and-answer, data search, and document creation, to all departments. The scale of generative AI application has expanded from training and pilot use to common services in actual work environments. However, active users by department, usage frequency, work time savings, error correction time, and response acceptance/rejection rates have not been disclosed. If only the number of functions increases between 2026 and 2028, services with low actual usage and work improvement will remain listed as implementation achievements.  While the scale of Busan's generative AI is verified based on functions, the actual scale of usage is deemed unconfirmed.

Decision support utilizing sensors, models, and AI is being promoted in the fields of water supply networks, marine environments, auditing, and safety. Administrative AI has shifted to a multi-layered structure where general-purpose language models and sector-specific prediction and detection models operate simultaneously. However, publicly available data comparing a common model registry, data lineage, accuracy, error cases, and operational accountability across the entire city has not been confirmed. As the number of specialized models increases over the next two to three years, the cost of integratedly managing model-specific risks grows rapidly.  While Busan's administrative AI is in a phase of multi-sectoral expansion, city-wide Model Governance remains unconfirmed.

The 22 types of generative AI services, the Big Data Wave, and sector-specific smart platforms demonstrate that Busan possesses numerous AI and data assets. Administrative digital assets have evolved into a structure where general-purpose AI and specialized analysis systems coexist within a single statistical portal. However, a common runtime has not been confirmed where analysis results across platforms are linked to the same policy meetings, approvals, budgets, and performance evaluations. As the number of systems increases between 2026 and 2028, information retrieval may become faster, but discrepancies in judgment between organizations could widen.  What has been confirmed so far is the accumulation of AI and data tools, not the integration of the decision-making ecosystem.

Requests to analyze card sales before and after the "15-Minute City" project, sales at traditional markets and small businesses following changes to mandatory store closures for large supermarkets, and pedestrian traffic and card sales before and after festivals are asking about outcomes rather than inputs. Administrative inquiries have shifted from whether a project was implemented to whether the policy actually changed the local economy. However, no public records linking analysis results to cases where existing policies were maintained, modified, or discontinued are found. If analyses remain merely as retrospective reference material for two to three years, the expansion of outcome questions does not lead to changes in decision-making.  While Busan possesses a demand for outcome analysis, an ecosystem for outcome-based policy coordination is deemed unconfirmed.

The 2025 practical AI training covered strategic planning and technical practice for the city hall, districts, counties, and affiliated organizations, while the expansion of AI services across all sectors was forecast for 2026–2028. The structure of talent development expands from simple tool usage to service planning capabilities. However, during the 2025 administrative audit by the Busan City Council, it was pointed out that over 80% of some AI training courses were at a basic level. If the speed of service diffusion over the next two to three years outpaces planning and verification capabilities, employees will use AI results but be unable to identify the limitations and errors of the models.  It is assessed that a gap exists within Busan's AI administrative talent ecosystem between the expansion of users and advanced verification capabilities.

Smart Ocean Village is pursuing marine environment monitoring and AI-based decision support, while the water supply sector is implementing big data and AI-based pipe network simulations. Administration across these sectors is shifting from post-incident complaints and inspections to proactive detection and prediction. However, publicly available evidence comparing the concordance rate between predictions and actual events, response times following alerts, and accident reductions is limited. Even if the system is operational between 2026 and 2028, the platform will remain stagnant as a monitoring screen if prediction accuracy is not linked to on-site actions. While  sector-specific AI has secured prediction capabilities, ecosystem evidence demonstrating that predictions have actually changed administrative behavior is lacking.

Generative AI reduces information gathering costs by processing administrative data searches and document drafts in seconds. The bottleneck in administration shifts from the time spent searching for data to the time spent determining which evidence to trust and which alternative to select. While Busan's AI services provide search and drafting functions, decision-making capabilities that compare the effectiveness, costs, and risks of different alternatives are not found in publicly available data. If search speeds alone increase between 2026 and 2028, a structure capable of reproducing existing policy logic more rapidly could be formed. Although the structural change in administrative AX has shifted from document generation to judgment and verification, Busan's current function is assessed to be closer to the former.

RAG-based administrative services strengthen the basis of responses by searching internal administrative data. This structure represents a shift in generative AI from being centered on publicly available internet information to being centered on data held by Busan City. However, a public Audit Trail—which records the original text referenced in the response, the reference date, conflicting documents, discarded guidelines, confidence levels, and final reviewers—is not available. As administrative data accumulates over two to three years, the risk of incorrectly mixing recent, historical, and conflicting evidence increases.  The structural gap of Busan-style generative AI is evident in the determination of evidence validity rather than in the generation of responses.

Private card, telecommunications, and spatial data reveal changes in commercial, mobility, and living areas faster than administrative statistics. Policy evidence has shifted from annual statistics to runtime levels at the daily, hourly, and 50m grid levels. However, no disclosure system has been identified that manages the sampling bias, changes in service providers, formula modifications, costs, and reproducibility of private data alongside policy decision records. If reliance on private data increases between 2026 and 2028, changes in supply contracts could lead to a disruption of the policy baseline. While  Busan's data administration has secured speed, its long-term comparability and reproducibility are deemed unconfirmed.

AI-based audit administration expands the scope of audit target selection and risk detection from sample inspections to data-driven full search. The structure of audits shifts from post-hoc detection to the early detection of anomalies. However, the accumulation of public data to verify audit targets selected by the model, unselected targets, actual violations, false positives, and biases is still in its early stages. If AI recommendations are incorporated into audit priorities between 2026 and 2028, unexplained false positives will also transform into administrative burdens and reputational risks.  Audit AX is determined to exhibit both an expanded detection scope and a procedural liability gap simultaneously.

The 2026 AI Administration Promotion Plan presented a direction for systematically disseminating AI throughout the administration. Pilot implementations by individual departments have shifted to a city government-level portfolio. However, an AI Project Registry that discloses project-specific baselines, target outcomes, responsible departments, model risk levels, and discontinuation criteria in a common format has not been identified. If the expansion across all sectors proceeds first between 2026 and 2028, it will be difficult to distinguish between low-performance and high-risk projects using the same criteria.  Busan's AI response is currently in the portfolio formation stage, while the portfolio performance management stage is deemed unconfirmed.

Generative AI administrative services were applied to all departments in early 2026, and a schedule for advancement and expansion over the next three years was presented. The application of AI has expanded from limited user testing to the common work environment for all employees. However, a public baseline comparing work hours, complaint response accuracy, correction rates, rework, and security incidents before and after use has not been confirmed. Once usage becomes incorporated into basic work procedures over the next two to three years, it will become difficult to compare the effects with the period before AI was introduced.  Currently, while the expansion schedule for the three-year response is clear, the structure for measuring causal effects is deemed unconfirmed.

The 2026 data work plan includes expanding analysis to resolve municipal issues, employee training, and utilizing pseudonymized information. The scope of data utilization is expanding from public statistics to combination analysis that protects personal information. However, no data has been found that collectively discloses policy outcomes, re-identification risks, combination failures, deliberation periods, or actual policy changes related to pseudonymized combination analysis. If utilization and protection are managed as separate performance indicators for the 2026–2028 period, it becomes difficult to assess the risk between expanded analysis and public trust.  Busan's response to pseudonymized information is currently in the stage of expanding the utilization base, not yet at the stage of integrated outcome and trust evaluation.

Practical AI and data training have expanded their scope to include central offices, districts, counties, and public institutions. Competency policies have shifted from fostering a small number of data experts to utilizing AI among general administrative staff. However, actual analysis performance before and after training, changes in policy planning, error detection rates, and field application rates are not disclosed. If only the number of trainees accumulates over two to three years, it is impossible to distinguish between the expansion of training and the improvement in decision-making quality. While  the response to training has expanded in terms of participation scope, it is deemed unverified in terms of work outcomes.

Busan announced that it has launched sLLM and RAG-based generative AI administrative services across all departments, becoming the first local government to do so. Evidence of the early adoption of a common AI work environment is confirmed in comparisons with domestic local governments. However, comparisons of active usage rates, time savings, accuracy, and policy change rates using the same standards as other leading cities are not disclosed. If the designation of "initial adoption" in 2026–2028 replaces performance comparison, the difference between technological leadership and administrative performance is obscured. While  Busan shows leading signals at the time of adoption, it is judged to be incomparable in terms of decision-making outcomes.

AI Governance in the international public sector focuses on managing data lineage, human review, explainability, impact assessment, objections, and model monitoring. Busan's generative AI services were built on the premise of utilizing security systems and internal data. However, evidence regarding the disclosure of risk-level approvals, citizen impact assessments, error disclosures, and model change history at the city government level has not been confirmed. As AI expands in the fields of citizen engagement, auditing, and safety over the next two to three years, procedural accountability, rather than operational convenience, will determine comparative competitiveness. While  Busan's technology adoption readiness is high, its public governance readiness is judged to be unconfirmed.

Large local governments such as Seoul and Gyeonggi operate data analysis organizations, public data, digital twins, and AI services in a multi-layered manner. Busan is also narrowing the tool gap by establishing a Big Data Wave and sector-specific AI platforms. However, comparable metrics for analysis → decision change → outcome across cities have not been standardized domestically. Comparing only the number of platforms and budgets for 2026–2028 does not allow for an assessment of Busan’s actual administrative productivity and resident impact.  Busan’s domestic standing is assessed as upper-middle tier in tool ownership, but as having a common data gap in Decision Runtime.

The analysis of Busan's 15-minute city, festivals, commercial districts, transportation, and marine environment addresses local issues by combining them with private sector data. This represents a more advanced structure than the stage where only structured administrative statistics are utilized. On the other hand, evidence managed by common standards regarding analysis reproducibility, control groups, pre-baselines, and post-outcomes is limited. Even if the number of analysis cases increases over two to three years, they are not aggregated into the learning of the entire city if the comparative designs differ.  Busan's relative strength is determined by the realism of its analysis topics, while its weakness is the standardization of the analysis results into policy experiments.

Connecting the 22 types of generative AI services with their application across all departments confirms their functionality and deployment scope. However, since active users, query counts, adoption rates, modification rates, time savings, and policy change rates are not disclosed, usability performance cannot be calculated. The number of services multiplied by the number of departments shows the scale of diffusion but does not reflect administrative performance. If the number of functions increases between 2026 and 2028, success indicators without a denominator could become larger.  The current figures for Busan's generative AI represent the scale of implementation, while decision-making productivity is determined by the Data Gap.

Analysis requests for Big-DataWave include actual policy issues such as the "15-minute city" concept, mandatory store closures for hypermarkets, festivals, parking, pedestrian traffic, and consumer sales. The number of requests and the "completed response" indicators demonstrate that the analysis supply is functioning. However, subsequent figures regarding the progression from "completed response" to "approval reflection," "project modification," and "outcome change" are not disclosed. If the number of completed cases accumulates over two to three years, the performance of analysis processing and the performance of policy changes may be perceived as identical.  The analysis completion rate is determined to be an administrative service indicator, not a policy effectiveness indicator.

The analysis of card sales before and after the 15-Minute City project was requested to compare changes in 2023 and 2024, specifically before and after completion. While the structure of a time-based comparison exists, evidence that control regions, external economic conditions, prices, and seasonal factors have been separated cannot be confirmed solely from the requested information. Simple increases and decreases before and after mix policy effects with simultaneous external changes. If similar analyses are repeated between 2026 and 2028, erroneous causal determinations could accumulate, leading to budget expansions or reductions.  The numerical gap in Busan data analysis lies between the measurement of change and the determination of causality.

The analysis of shifting mandatory store closures for hypermarkets to weekdays is a policy question that compares the sales of hypermarkets, traditional markets, and small business owners together. It confirms a structural shift aimed at examining the distribution of benefits among stakeholders rather than the sales of a single industry. However, open outcomes combining consumer benefits, workers' working conditions, the shift to online consumption, and long-term effects are not identified. If conclusions become fixed based solely on short-term credit card sales over a period of two to three years, the costs and benefits of the policy become limited to specific groups. While  Busan's data-driven judgment expanded to multi-group comparisons, it is determined that it did not reach the total social outcome.

While the number of participants and courses in AI education demonstrates the input of capabilities, a city council audit pointed out that over 80% of some courses are at a basic level. This serves as evidence that an increase in training volume does not equate to the acquisition of advanced capabilities. Training has not been linked to on-the-job analysis quality, model validation, or policy changes. If the expansion centered on basic education continues from 2026 to 2028, a structure will remain where the number of AI users increases but there is a shortage of verifiable decision-makers.  The number of training sessions indicates expansion, but the organization's judgment capabilities remain unassessed.

The first gap is the absence of a Decision Log . Analysis requests, AI responses, administrative documents, approvals, budgets, and execution results exist individually. However, public records linked to the same policy ID to identify which evidence discarded existing alternatives and changed the final decision are not available. As the use of AI becomes commonplace between 2026 and 2028, it will become even more difficult to track the actual impact of AI in the decision-making process.  The greatest gap in Busan's administrative AX is determined to be the traceability of decision changes, rather than analytical capabilities.

The second gap is the absence of an Outcome Baseline . While the 22 types of generative AI services present their functionalities, common baseline values ​​for pre-implementation work hours, error rates, complaint quality, and policy processing periods have not been disclosed. This structure provides weak standards for comparison even if changes occur after implementation. If AI becomes entrenched as the standard work environment within two to three years, it becomes virtually impossible to compare the pre-implementation state with a non-user control group.  The Busan generative AI is assessed to have lower effectiveness verification readiness than deployment readiness.

The third gap lies in the unverified management of data lineage and reference dates . While RAG utilizes internal data, administrative records contain pre-revision guidelines, duplicate documents, and conflicting statistics. There is no confirmed structure for publicly tracking which data and reference dates were used by AI responses. As the volume of documents increases between 2026 and 2028, the risk of recency errors being repeated in the form of accurate sentences grows.  The gap in evidence-based administration lies not in data access, but in the ability to determine which data is valid.

The fourth gap is the disconnect between analysis and the budget . Even if data analysis verifies policy effectiveness, public records linking it to increases, decreases, or terminations in the following year's budget are limited. The structure is such that analysis is managed as reference material, while the budget is treated as separate review material. If low-performing projects are repeatedly included in the budget for two to three years regardless of analysis results, the financial benefits of data-driven administration disappear.  Busan's Decision Readiness is judged as unconfirmed in budget reallocation rather than in policy analysis.

The fifth gap is the lack of organizational learning regarding AI errors . Generative AI and predictive models can generate incorrect answers, omissions, biases, and false positives. No registry has been found that integrates and publicly discloses error cases, correction times, recurrence rates, and version changes by model for Busan City. As the number of services increases between 2026 and 2028, the same error could spread to documents and decisions across multiple departments.  The risk to Busan's AI administration is assessed to lie not in the existence of errors, but in a structure where errors are not retained in organizational memory.

Big Data Wave and Generative AI administrative services provide analytical data and business support infrastructure, respectively. Busan City currently possesses both data inquiry and analysis capabilities, as well as document and search AI. However, no public evidence has been confirmed linking the two infrastructures to the same policy task ID. If separate upgrades continue from 2026 to 2028, data analysis results may not be reflected in the Generative AI responses, or the reference dates may change.  It is determined that multiple technical infrastructures have been secured, while policy-level interoperability remains unconfirmed.

AI and data training was conducted for the headquarters, districts, counties, and affiliated organizations, and practical planning training was also implemented. The target audience for training was expanded from select IT personnel to general policy staff. On the other hand, the lack of advanced application and convergence capabilities was pointed out during the 2025 administrative audit. If AI spreads across all departments over the next two to three years, a bottleneck may arise where a single verification personnel handles errors from multiple users.  The talent infrastructure is assessed as being in the user expansion phase, while the advanced verification personnel are judged to be in a shortage phase.

The utilization infrastructure for card sales, telecommunications flow, spatial grids, administrative statistics, and pseudonymized information observes policy phenomena from various angles. Data infrastructure has expanded from structured public data to public-private integration. However, data that collectively manages the impact of sample bias, changes in formulas, missing values, and the termination of data acquisition on analysis results has not been identified. As reliance on private data increases between 2026 and 2028, changes in contracts and formulas will disrupt the time series of policy outcomes. While  data readiness has expanded in scope, it is judged to be unconfirmed regarding quality and continuity.

Busan City’s AI administration promotion plan and personal information and security system serve as the institutional foundation for AI utilization. The introduction of AI has shifted from autonomous use by employees to the formal administrative system. However, no integrated system has been identified that publicly discloses human final judgment in high-risk decision-making, citizen notification, objections, prohibited automation areas, and model discontinuation criteria. If the application expands across civil complaints, welfare, auditing, and safety sectors over the next two to three years, procedural gaps will transform into risks to citizens' rights.  It is determined that while the system has reached approval for adoption, it has not yet reached liability for impact.

Generative AI administrative services aim to reduce employee workload and public response time by supporting administrative data searches and drafting of civil complaint responses. The structure involves internal operational technology extending to the quality of responses at the point of contact with residents. However, comparisons of civil complaint processing time, re-inquiry rates, error rates, and citizen satisfaction before and after implementation have not been disclosed. Even if employee usage increases between 2026 and 2028, if the speed and accuracy experienced by residents do not improve, administrative AX will remain merely an internal efficiency measure. While  resident adoption was presented as an expected effect, actual outcomes are deemed unconfirmed.

Data analysis regarding mandatory store closures for large supermarkets, the "15-minute city" initiative, festivals, and commercial districts targets policies directly related to the lives of small business owners and residents. Data administration is shifting from the creation of internal reports to the analysis of regional economic outcomes. However, follow-up evidence demonstrating that analysis results have altered business environments, facility layouts, budgets, and event designs is limited. If an analysis is concluded after two to three years, local businesses remain merely targets for data provision and do not become beneficiaries of policy changes.  While analysis targeting businesses and residents has expanded, the tangible impact of changes in decisions remains unconfirmed.

Public data and the Big Data Wave provide access to data for small and medium-sized enterprises (SMEs) and citizens. The scope of data utilization has expanded from internal administration to private sector use. However, regular data linking the performance of open data—leading to corporate products, services, sales, and employment—with data quality issues is not being verified. If only the number of datasets increases between 2026 and 2028, data with low actual utility value will remain as an achievement of openness. While  diffusion into the private sector is confirmed in terms of accessibility, it is deemed unverified in terms of industrial outcomes.

AI-based audit, safety, water supply, and marine environment systems aim for a structure that reduces citizen damage by detecting accidents, corruption, and leaks at an early stage. Resident outcomes are manifested not in the implementation of the system, but in accident reduction, recovery time, water quality, and safety. Changes in citizen damage and false positive costs before and after the introduction of these systems are limited in publicly available evidence. If only the number of alerts and screens increases over two to three years, the resident benefits are obscured by the track record of administrative technology implementation.  The resident diffusion of AI by sector is confirmed at the level of detection capabilities, while the outcome of damage reduction is deemed unconfirmed.

While the Busan Metropolitan City Hall possesses common Generative AI services and specialized data departments, the data personnel, budgets, and analysis experience of the districts and counties are not identical. Although Administrative AX aims for expansion across all departments and districts, actual implementation capabilities may vary depending on organizational size. Data comparing active usage rates, analysis requests, internal analysis, and policy reflection rates by district and county is unavailable. If these capability gaps accumulate between 2026 and 2028, the accuracy of data-based policies could differ even among Busan citizens based on their addresses. It  is determined that within Busan’s Administrative AX, the equilibrium of service access does not align with the equilibrium of utilization capabilities.

Big-Data Wave analysis requests include region-specific issues such as pedestrian traffic in Dong-gu, parking in Yeongdo-gu, commercial districts in Busanjin-gu, and the Gwangalli Festival. While data demand arises across Busan, the types of problems and the difficulty of securing data vary. If analysis support is operated on a request basis, organizations with strong application capabilities and project design skills are more likely to secure support. If this structure is maintained for two to three years, regions that failed to draft effective analysis requests will be excluded before regions lacking data.  It is determined that spatial disparities are more likely to stem from the ability to design administrative questions than from data possession.

50m grid pedestrian traffic and card sales capture lifestyle changes at the neighborhood level and below. The unit of administrative analysis is shifting from district and county averages to micro-living areas. However, evidence has not been confirmed that the same spatial resolution and update cycle are applied to all policy fields. If only certain areas are precisely observed between 2026 and 2028, unobserved areas could be misidentified as having no demand.  While Busan's spatial intelligence has been refined, the uniformity of observation across the entire region is deemed unconfirmed.

The original downtown, West Busan, and East Busan differ in population, industry, transportation, and commercial structures. Even if the same AI model is applied across the entire city, prediction errors vary by region if the regional composition of the training data differs. There is no publicly available data comparing model performance by district, county, age, income, or business size. If only average accuracy over two to three years is managed, the low accuracy of vulnerable areas is obscured by the overall average.  The spatial risk of Busan's administrative AI is determined by the invisibility of regional error rates rather than service deployment.

Busan-style generative AI administrative services were applied to all departments in early 2026 and will be further enhanced and expanded over the three years from 2026 to 2028. The next two to three years will be a period during which initial functions, data, and usage patterns become fixed as standard business procedures. The baseline prior to introduction and the AI ​​impact decision log have not yet been disclosed. Once basic procedures become entrenched, it becomes difficult to compare work with that without AI, and incorrect usage practices will remain as organizational norms.  Currently, the "Golden Time" is determined not to be before the introduction of AI, but before the initial three-year operational standards solidify.

AI and data analysis are rapidly expanding into areas such as the 15-minute city, commercial districts, festivals, transportation, marine affairs, water supply, and auditing. The scope of application has shifted from general tasks to expert judgments affecting citizens' rights and safety. However, no public disclosure system has been identified that integrates the management of risk levels, errors, objections, and suspension criteria. If automation expands to high-risk sectors between 2026 and 2028, it will be difficult to apply post-hoc regulatory supplements to already accumulated decisions and complaints. The Governance Golden Time is determined to be before the application of high-risk measures becomes entrenched as part of daily operations.

The utilization of private credit card and telecommunications data, as well as pseudonymized information, reveals policy trends faster than administrative statistics. The speed of data judgment in Busan is accelerating from annual statistics to monthly, daily, and hourly units. However, the public disclosure system for managing changes in private data formulas, samples, and contracts, as well as long-term comparisons, is limited. If policies relying on the same data accumulate over two to three years, the reproducibility of past judgments disappears after a change in providers.  The "Data Golden Time" is determined to be before the continuity of the baseline is severed, rather than the expansion of usage volume.

Although employee training has expanded, the 2025 administrative audit pointed out an overemphasis on basic courses. This serves as evidence that the diffusion speed of AI services differs from the growth speed of senior planning and verification personnel. Data measuring advanced capabilities and actual error detection has not been confirmed. As the use of AI becomes routine for all employees between 2026 and 2028, it will become difficult for a small number of experts to verify all models, data, and responses. The  "Golden Time" for talent is determined to be before the number of users structurally surpasses the number of verifiers.

12-1. Golden Time Application Case in Basic Local Governments ① — Busanjin-gu

A request was made to Big-Data Wave for an analysis of changes in card sales before and after the 15-Minute City project, which includes the Dangam and Gaegeum living areas. The evaluation of Busanjin-gu's living area policies has shifted from facility completion to the actual sales outcomes of surrounding commercial districts. However, no public analysis results separating the effects of prices, seasonality, economic conditions, and control areas from changes before and after completion, nor any subsequent policy changes, have been confirmed. If the 15-Minute City project is expanded using the same method between 2026 and 2028, errors in the initial causal determination could be replicated to other living areas.  Although Busanjin-gu is a region where data-based outcome questions were raised, it is determined that it has not yet reached the stage of confirming policy effects.

The Dangam and Gaegeum areas utilize actual living zones—larger than dong units but smaller than district units—as policy units. Spatial judgment has shifted from the average of administrative dongs to living zones that combine resident mobility and consumption. However, there is no confirmed public evidence linking floating population, card sales, facility accessibility, resident satisfaction, and commercial survival into a single living zone runtime. If the representative living zone projects are completed within two to three years, only facility-specific performance results may remain without a baseline.  Busanjin-gu's Golden Time is not the existence of the living zone analysis itself, but the period during which it can be recorded whether the analysis influenced subsequent facility placements and budgets.

12-2. Golden Time Application Cases in Basic Local Governments ② — Yeongdo-gu

Yeongdo-gu applied for data analysis consulting regarding its parking shortage after the analysis of Saeol consultation complaints confirmed that parking-related complaints were the most frequent. This marks a case where resident complaint texts shifted beyond the aggregation of recurring complaints to questions for policy analysis. However, there are no publicly disclosed follow-up records confirming which decisions—such as parking space allocation, enforcement times, public transportation, or tourist routes—were changed based on the analysis results. If parking demand changes due to tourism, housing, and population aging between 2026 and 2028, past complaint analyses become one-time data with a different reference point.  While Yeongdo-gu is a case where data identified a policy problem, it remains unconfirmed as a case where data changed the solution alternative.

Yeongdo-gu's parking problem is a complex issue where the hourly demands of residents, tourists, merchants, and the mobility-impaired clash within the same space. The number of complaints alone cannot distinguish which group's demand or which time period constitutes a bottleneck. Furthermore, a public runtime linking complaints, parking turnover, illegal parking, pedestrian safety, and commercial revenue cannot be verified. If judgments based primarily on the volume of complaints are maintained for two to three years, areas with high reporting rates and areas with actual high risks become intermingled.  Yeongdo-gu's "Golden Time" is assessed as the period during which the causes of complaints and the differences between policy alternatives can be determined, moving beyond the stage of merely identifying high-complaint hotspots.

If data on usage, errors, corrections, and adoption from the initial stages of AI application across all departments in 2026 is not retained, the baseline for comparing subsequent administrative performance will disappear. AI work support is structured to expand over three years, with non-use tasks rapidly decreasing. Currently disclosed outcomes before and after implementation are limited. After 2026–2028, it becomes difficult to distinguish whether AI has improved work processes or merely absorbed an increase in workload.  The first loss will be determined by causal evidence from the early stages of AI adoption, rather than the technology budget.

If analysis results are not linked to approvals, budgets, and execution, analyses of the same issues are repeated across departments and years. While numerous real-world policy questions are identified in the Big Data Wave, the history of reuse and decision changes is not disclosed. If personnel transfers and organizational restructuring occur over two to three years, the premises, limitations, and subsequent decisions of the analysis become separated from organizational memory.  The second loss is determined not by individual reports, but by administrative memory that does not repeat policy failures.

If errors in AI responses and prediction models are not accumulated in a registry, errors will be repeated using the same data, prompts, and models. While services are expanding to all departments and specialized fields, a city-wide error learning system has not been verified. As multiple systems become more advanced between 2026 and 2028, versions and responsible parties will change, making it difficult to reproduce past errors.  The third loss is determined not by accuracy figures, but by the learning time required to incorporate errors into subsequent decisions.

If differences in data personnel and question design capabilities accumulate among districts and counties, analysis support may be concentrated in organizations with higher application capabilities rather than in regions with high demand. Currently, request-based analysis systems and regional tasks are identified, but unapplied demand is not observed. If this structure becomes entrenched for two to three years, the administrative AX gap will transform into a regional service gap.  The fourth loss is determined not by technological equality, but by equity in policy judgments that are not swayed by address or organizational capacity.

Busan already possesses generative AI services across all departments, Big Data Wave, the utilization of pseudonymized information, and sector-specific AI platforms and practical training. The components of Administrative AX are not entirely absent, but rather dispersed across different stages. However, evidence linked by Policy IDs and Decision Logs has not been verified. There is still time to compare existing assets based on the same policy outcomes before the system structure becomes fixed between 2026 and 2028.  Currently, the opportunity is assessed as a period to establish the decision-making connectivity of dispersed assets rather than a new system.

Requests for analysis regarding the "15-minute city," large supermarkets, festivals, and parking demonstrate that the administration is already posing specific questions regarding policy effects. Busan's gap lies not in the absence of questions, but in the lack of decision records following the responses. A public chain of analysis results and policy changes has not been identified. If analysis requests continue to increase over two to three years, the ability to connect follow-up decisions will become the criterion determining administrative productivity.  The current opportunity lies not in expanding the volume of analysis, but in a phase where realistic policy questions can be preserved as repeatable decision evidence.

The RAG-based generative AI utilizes internal administrative data from Busan City. This serves as a foundation for reusing the organization's past policy rationale while reducing search times. However, the structure for searching for data validity, conflicting grounds, and decision results together has not been disclosed. If the rationale and outcome are not linked before administrative data increases between 2026 and 2028, the AI ​​will be able to quickly locate past documents but will be unable to distinguish between the success and failure of past decisions.  Currently, the Opportunity is assessed as the initial operational period during which administrative documents can be established as organizational learning data.

The Busan Metropolitan Council's administrative audit publicly questions the level of AI education and the effectiveness of projects. The structure involves the administration's dissemination of AI and the council's external verification existing simultaneously. However, connecting data linking audit findings, municipal administration, model/education changes, and subsequent outcomes is not being verified. If council verification remains as individual inquiries for two to three years, AI governance becomes fragmented among budget reviews and project reports.  Currently, the opportunity is determined to be the period during which the internal operation and external verification of AI administration can be linked within a single evidence chain.

Administrative AX Runtime

Policy Question → Baseline → Data & AI Analysis → Alternative Comparison → Decision Change → Budget → Execution → Citizen Outcome → GAP → Audit & Council Verification → Correction → Next Decision

Administrative AI Expansion + Decision Evidence GAP

Busan City's AI and data administration has moved beyond the introduction phase and is now in the expansion phase across all departments.

Task search, document creation, and individual policy analysis are confirmed.

There is no connection between what AI and data changed regarding existing decisions and how the results changed.

The current danger is not administration that does not use AI, but administration that uses AI but cannot determine whether it made better decisions.

The final ruling is Administrative AI Expansion + Decision Evidence GAP.

1. Generative AI active users by department
2. AI usage frequency per employee
3. Usage rates by 22 types of services
4. AI Answer Adoption, Modification, and Rejection Rates
5. Average correction rate of AI-generated documents
6. Reducing administrative data search time
7. Reducing time spent drafting civil complaints
8. Re-inquiry rate of civil complaints before and after AI usage
9. Number of AI response errors, omissions, and biases
10. Time from error discovery to correction
11. Recurrence rate of the same error
12. Analysis request reception, completion, and waiting period
13. Approval reflection rate of analysis results
14. Policy alternative change rate after analysis
15. Change rate of project target and region after analysis
16. Budget increase, decrease, and completion rates after analysis
17. Baseline disclosure rates prior to implementation by policy
18. Analysis ratio using control region/control group
19. Private data formula and sample change history
20. Accuracy, False Positives, and Missed Detections by Model
21. Gap in AI Usage and Internal Analysis by District/County
22. On-the-job application rate after completing training
23. Number of advanced AI and data verification personnel
24. Actual correction rate following issues raised by the council and audit
25. Improvement rate of citizen and business outcomes after policy change

Data GAP: In the public evidence, the Busan City Decision Runtime connecting Policy Inquiry → Data/AI Analysis → Comparison of Existing Proposal and Alternatives → Approval/Budget Change → Execution → Citizen/Business Outcome → Audit/Council Verification → Follow-up Correction under the same Policy ID is not confirmed.

Runtime Chain

Data → Analysis → Alternatives → Change of Decision → Budget → Implementation → Outcome → Errors/GAP → Audit/Council → Correction → Next Decision

Source · Verification

• Busan Metropolitan City Launches Nation's First Generative AI Administrative Service
• Busan Metropolitan City, 2026 AI Administration Implementation Plan
• Busan Metropolitan City, 2026 Data-Based Administrative Major Work Plan
• Busan Big-Data Wave, Status of Policy Data Analysis and Consulting
• Busan Big-Data Wave, Statistical Explanatory Materials
• Busan Metropolitan City, AI-based practical training for future administration
• Busan Metropolitan Council, 2025 Administrative Audit of the Korea Information Industry Promotion Agency
• Busan Metropolitan City, Data-based Intelligent Marine Environment Management Support Platform
• Busan Metropolitan City to Support AI and Big Data-Based Decision Making for Water Supply Networks by 2026
• Busan Metropolitan City, AI-based Smart Audit Administration Support System

Structural insights remaining from this analysis

Administrative decision-making retains only the final proposal written in documents and erases unchosen alternatives. Even if AI presents more evidence and alternatives, if the differences between the original plan, alternatives, and the final plan are not recorded, it cannot be verified whether the AI ​​changed its judgment. It is a structure where usage remains, but the impact vanishes.

In this structure, the difference between good AI and bad AI is not determined solely by accuracy. If accurate analysis fails to change approvals and budgets, the administrative effect is close to zero; conversely, if incomplete analysis alters high-risk decisions, even small errors escalate into citizen outcomes. The administrative value of AI is formed by the product of model performance and decision impact.

Therefore, there is only one structural insight into Busan Administration AX. AI becomes an administrative capability not when it replaces decisions, but when it records why decisions changed.  Without this record, Busan City can search faster and generate more documents, but it cannot determine whether it made better decisions than in the past.

Golden Time Thesis — The period from 2026 to 2028 for Busan City’s administrative AX is not a time for the diffusion of generative AI, but rather a period for the first time to record the difference between Evidence → Alternative → Decision Change → Outcome in the administrative records. Without this connection, AI can increase the speed of administration, but it cannot increase the speed of detecting and correcting policy failures.