URBEXA

INJ Architects — Research edition 0.3 Concept and research direction: Ibrahim Joharji Evidence snapshot: 25 September 2026 Research design paper and documented evidence atlas.

URBEXA connects urban systems, employment and artificial intelligence through the relationship between livelihoods, households and property.

Abstract

Artificial intelligence may alter the relationship between work and place. A city can face a change in its economic role even when its population initially remains stable: firms may purchase fewer outsourced tasks, workers may change occupations, households may reduce spending, and commuting may change before residential relocation occurs. This study proposes an urban framework connecting occupational change to adaptation, household decisions and spatial consequences. Housing affordability and the use of residential and commercial property are its principal planning concerns. Its starting proposition is conditional: technologies reshape urban systems through institutions, markets and household constraints, with outcomes that can include growth, stagnation, redistribution or decline.

The study combines a structured review of 21 historical references with contemporary research and reproducible statistical extractions. Its evidence atlas covers 36 urban profiles: 18 European cities, six US metropolitan labour-market benchmarks, and 12 Middle East urban centres. It also includes a global country context and an occupational exposure reference. These components have different geographic and statistical meanings. They are linked through explicit identifiers and documentation, without converting national export statistics into local employment, or task exposure into a probability of displacement. The interactive model illustrates accounting relationships under user-defined assumptions. It does not yet estimate AI-attributable migration or rank the world’s most vulnerable cities.

1. The urban question

The central question is: under what conditions does AI-related change in work alter who lives in a city, how households earn and spend, and what urban space is used for? Employment is one part of this question. Housing, commuting, education, family composition, municipal services and the geography of opportunity are equally important outcomes.

Consider three hypothetical workers facing similar changes to their tasks. One remains employed and produces more. Another changes occupation locally after training. A third loses income but cannot afford to move. Their occupational exposure could be similar while their urban consequences differ substantially. An effective planning instrument must represent this divergence, rather than assigning one demographic destiny to an exposed profession.

A further distinction concerns employment location. A job recorded inside a metropolitan area may be performed by a commuter living elsewhere. A resident may work remotely for a foreign employer. Consequently, the loss of one workplace job cannot be counted as the departure of one resident. The database preserves administrative-city, metropolitan and morphologically defined urban-centre boundaries, and future household analysis must use residence-based information.

The research therefore treats cities as linked employment, residential and service systems. Its contribution would lie in connecting these systems transparently, especially in places specialized in internationally traded expertise. This remains a proposed contribution to be tested, not an established claim of unprecedented invention.

Why this is an architectural and real-estate study

The principal planning concern is the transmission from changing livelihoods to housing demand and the use of urban property. Income pressure can affect a household’s ability to meet rent, mortgage and running costs before it causes relocation. Responses may include drawing down savings, postponing household formation, sharing accommodation, choosing a smaller dwelling, changing neighbourhood or moving to another city. These are mechanisms to investigate, not observed AI effects established by this edition.

Housing is a substantial, recurring commitment for many households. Eurostat’s Housing in Europe 2025 reports that EU households devoted approximately 19% of disposable income to housing in 2024, with substantial differences across countries. OECD affordability evidence shows why low-income renters and mortgaged households deserve separate attention. A global claim that housing always takes the largest share of every worker’s salary, or that real estate is always the first affected sector, is not warranted by these aggregates. [38][39]

URBEXA therefore places housing affordability, residential demand and property use at the centre of the study. It distinguishes pressure on an individual household from changes in an entire housing market. Many households losing income could weaken effective demand, but prices and vacancy also depend on construction, existing shortages, credit conditions, inward migration and policy. Smaller or cheaper homes might face greater demand even when purchasing power declines. Receiving cities could face additional housing pressure. These potentially opposite effects are reasons to model housing segments and destinations separately.

The architectural questions are concrete: which dwelling sizes and tenures may be needed; where adaptive reuse or flexible workspace could help; how employment districts and everyday services might change; and where affordable housing and transport access could support households through occupational transition. The study concerns the spatial consequences of changing work, with housing as a principal transmission channel.

2. What history supports—and what it does not

Agricultural change: direction matters

Agricultural innovation is not a universal recipe for rural population loss. McGowan and Vasilakis examine hybrid corn and provide evidence that productivity improvements can increase agricultural labour demand and restrain urbanization. Bustos, Caprettini and Ponticelli study Brazilian agricultural technology and find labour-saving change in soy production associated with industrial growth. Together, these studies make the factor bias of innovation central: whether technology saves labour, complements it or changes production scale affects the direction of structural adjustment. [3][22]

Putterman’s work on the diffusion of agriculture addresses long-run development differences. Its geographic scale and historical horizon differ from a study of displaced workers relocating between contemporary cities. It is useful for recognizing historical path dependence, but it cannot supply a migration coefficient for the AI economy. [2]

Studies of Southern US agriculture place machinery inside a wider institutional setting. Fligstein examines agricultural transformation and Black and white migration in the 1930s. Heinicke and Grove examine the later diffusion of the mechanical cotton picker. Their histories suggest that diffusion timing, farm organization, policy and the conditions under which machinery becomes economical matter alongside technical possibility. They do not justify reducing the Great Migration to a single machine. [11][12]

Tolnay’s demographic review and Alexander’s comparative treatment further support studying migration as a process with heterogeneous populations and routes. Perkinson and Hoover offer a different kind of evidence: potential out-migration following tobacco mechanization. An anticipated response must be distinguished from a measured population flow. URBEXA keeps that distinction in its evidence registry. [13][14][15]

Industrial change: skills, transport and organization

The transition to steam affected which skills were valuable. De Pleijt, Nuvolari and Weisdorf connect steam-engine adoption with changes in human capital formation; working skills and elementary education need not respond identically. Chin, Juhn and Thompson examine the movement from sail to steam in the merchant marine. These are useful precedents for occupational recomposition, including the creation of new technical roles alongside the contraction of older ones. Neither study is treated here as a complete demographic model of a city. [4][5]

Devine’s account of electrification emphasizes changes in production organization beyond substituting one energy source for another. Pietrykowski’s analysis of Fordism similarly directs attention to how firms organize labour and space. A contemporary analogy worth testing is that AI could change the organization and location of service production without immediately eliminating entire occupations. This is an inference guiding research design, not an effect established by those historical studies. [6][7]

Atack, Margo and Rhode connect nineteenth-century industrialization and urbanization. Kim examines the division of labour and the rise of cities. Together they motivate including transport access, production scale and occupational matching in a spatial account of technological change. The historical observation that industry and cities expanded together does not isolate the effect of technology from infrastructure and market development. [16][17]

Computerization, robotics and uneven cities

Berger and Frey provide a particularly close precedent for the project’s broad premise. Their research relates adaptation to the computer revolution to differences in urban population, human capital and wages. Cities with historically cognitive specializations benefited differently. Therefore, the general claim that technological change can reorder urban fortunes already has substantial empirical precedent. The defensible question is which additional mechanisms URBEXA measures and connects. [20]

Jerbashian examines automation and job polarization in Europe. Guidetti and Leoncini synthesize technological change, wages and employment. These contributions suggest that occupation counts alone can conceal distributional effects: a city may retain employment while the wage structure changes. Southall’s short publication is a book review and is preserved as intellectual context, not counted as a separate empirical test. [8][9][10]

Green Leigh and Kraft distinguish regional patterns in robotics-related activity. A location supplying a technology and a location adopting it occupy different positions in its economic geography. Acemoglu and Restrepo’s study of industrial robots provides evidence of local labour-market disruption, but its industrial setting and technology are not interchangeable with generative AI. Coefficients from robot adoption will not be inserted into an AI migration calculator. [21][23]

Rieniets and Beauregard show why urban decline must also be understood historically and geographically. A shrinking administrative municipality may coexist with suburban expansion; other cities decline because of changes unrelated to automation. Population loss is an outcome to explain, not evidence that a proposed technological mechanism is correct. Didier’s synthesis provides broader framing across technological transitions, while the empirical mechanism must be evaluated case by case. [1][18][19]

The historical proposition

History supports a conditional hypothesis: when technology changes the demand for tasks, the resulting urban trajectory depends on local specialization, adaptation opportunities, household constraints and the capacity of other places to receive people and investment. It does not support a universal sequence in which every technological era necessarily displaces workers and empties cities.

For URBEXA, historical evidence has three functions. It identifies plausible mechanisms, supplies counterexamples to simple narratives, and helps specify variables. It does not validate the numerical settings of the demonstrator or prove that present-day AI will reproduce a particular historical migration episode.

3. Contemporary evidence and the originality boundary

The contemporary literature already measures occupational and spatial exposure to AI. Eloundou and colleagues assess task exposure to large language models, while the ILO’s 2025 framework classifies occupational exposure using an updated task-based approach. These are measures of technical potential under defined rubrics. They do not directly observe employers replacing people or households moving. [24][25]

The OECD’s 2024 analysis reports average exposure of 32% of workers in urban areas versus 21% in non-urban regions under its definition. Brookings’ metropolitan analysis reports a different comparison: roughly 43% of the San Jose workforce versus 31% in Las Vegas in occupations where at least half of tasks could be affected under its method. London’s 2026 analysis estimates that at least 46% of workers are in exposed roles, compared with 38% across the UK, using the ILO framework. These figures demonstrate existing spatial analysis; they are not a common league table. Definitions, thresholds, geographies and source years differ. [26][27][28]

Augmentation also matters. The published study Generative AI at Work reports evidence from 5,172 customer-support agents and an average productivity improvement of about 15%. This workplace result does not estimate a city-wide employment or population response. It nevertheless makes a model with only displacement structurally incomplete. Productivity improvements, new demand and changing tasks must be possible outcomes. [29]

This edition is a structured evidence review, not a registered systematic review with an exhaustive screening flow. The 21 historical DOI records were checked against publisher-deposited metadata, and review depth is recorded for each item. Many historical claims were checked at abstract or institutional-summary level. Full-text verification of every study, forward and backward citation searching, and a formal search log remain necessary before publishing a categorical novelty claim. [30]

The proposed contribution is narrower and more useful than “the first AI-and-cities study”: a documented system connecting traded-service specialization, occupation-level change, local adaptation, household mobility constraints and urban consequences. Whether a comparable integrated model already exists must remain open to further review.

4. Selecting places through economic mechanisms

G20 or G7 membership is not the sampling rule. A place can be highly dependent on internationally traded expertise without belonging to either group. The initial country screen uses ICT service exports in dollars and as a share of service exports, alongside GDP. This is an accessible starting measure of economic structure, with an important coverage limit: ICT services do not encompass all business-process outsourcing or all digitally delivered professional services. The WTO’s broader digitally delivered services dataset is an identified expansion source, but its values are not included in this extraction. [31][32]

An additional housing layer contains 1,131 published EU-SILC observations for Poland, Czechia and Romania, covering 2019–2025. It distinguishes national household-type and income-group statistics from national urbanization-class overburden rates. Neither is substituted for a named-city observation. [40]

The database covers 217 separately identified economies in the World Bank country catalog, excluding regional aggregates. Not every economy reports every indicator and year. Missing observations are retained explicitly. A country appearing in the catalog does not imply that all its cities have been collected.

Selection proceeds through three distinct screens. First, identify national specialization and exposure to external service demand. Second, locate the urban employment clusters carrying that specialization using subnational evidence. Third, assess the availability and consistency of local demographic, housing and mobility data. National export intensity alone cannot select a “most affected city” or establish that its residents work in the exporting industry.

The 18 European pilot cities are Warsaw, Krakow, Prague, Brno, Bucharest, Cluj-Napoca, Sofia, Plovdiv, Ljubljana, Bratislava, Tallinn, Riga, Vilnius, Kaunas, Lisbon, Porto, Budapest and Zagreb. Their role is to establish comparable city-statistics workflows in a set of service-economy locations. Selection reflects data availability and a service-economy research interest, not an established ranking of export dependence or AI exposure. Six US metropolitan areas—Austin, Las Vegas, New York, San Francisco, San Jose and Seattle—provide a methodological benchmark where detailed occupation and wage estimates can be linked to an exposure reference. They are not substitutes for collecting city data in India, the Philippines, Ukraine, Vietnam or other intended service-export cases.

Future expansion should document both positive and negative selection decisions. An economy may be economically relevant but lack usable city microdata. A city may have excellent data but be a weak test of dependence on exported expertise. These are different reasons for inclusion or exclusion and should be visible to readers.

The Middle East is an explicit part of the research frame

The regional expansion includes Dubai, Abu Dhabi, Amman, Irbid, Cairo, Alexandria, Tehran, Riyadh, Jeddah, Doha, Beirut and Kuwait City. Their inclusion responds to the planning question and the availability of a common spatial baseline; it is not a finding that these are the most threatened cities. Jordan, Egypt, Iran and the Gulf economies require investigation of their actual occupational composition, firm adoption and household constraints before city-level vulnerability can be assessed.

Each profile contains 52 distinct indicators drawn from the JRC Urban Centre Database: population and growth estimates; built-up surface, volume, height and age; estimated age and sex shares; urban-centre area; and network-test measures. This is an urban and digital baseline. It does not contain 52 occupations or a validated measure of AI displacement. The layer uses fixed 2025 urban-centre boundaries, whereas the European and US layers use different source geographies. [41][42]

Geographic scope changes the meaning of the numbers. The source's Dubai centre includes Sharjah and Ajman; Amman includes Zarqa, Sahab and Al-Ruseifa; Cairo includes Giza and several adjoining cities. Abu Dhabi refers to the delineated urban core, not the whole emirate. Displaying these figures as municipal populations would be misleading. Each profile therefore names the constituent places recorded by the source.

Not all attributes share one statistical population. JRC warns that the WorldPop-based age and sex proportions should not be multiplied by GHSL population totals. Network performance reflects voluntary tests rather than universal coverage or the number of digitally employed workers. Built-up surface measures a footprint, not floor area, housing capacity or market value. These distinctions constrain which urban mechanisms the available data can support. [42]

Regional AI research also provides reasons to examine opportunity alongside risk. The ILO's 2025 Gulf-and-Levant report describes substantial scope for augmentation while warning about unequal outcomes; its regional scope excludes Egypt and Iran. An IMF study of Qatar similarly distinguishes occupational exposure from complementarity and productivity gains. Neither is a city-specific forecast, and neither supports assigning a common loss rate to Middle East cities. [44][45]

5. A matrix of people, livelihoods and urban systems

The unit of an observation is a defined indicator, geographic entity, reference period and population. The unit of a future causal analysis is more demanding: a city, occupation and worker or household group under a specified adoption scenario. Keeping these units distinct prevents a large spreadsheet from masquerading as a linked microdataset.

The analytical matrix is organized around ten domains: population and age; sex composition; household structure; employment and occupations; task characteristics; earnings and income; skills and adaptation; housing; transport; and mobility. The design catalog contains 70 candidate fields across these domains. It is a specification, not a claim that all 70 have already been observed in every city.

In the European profiles, the latest available value is shown once per indicator. Historical observations remain available separately. This stops ten years of the same variable from being presented as ten different dimensions. The current profiles contain 78–226 distinct published indicators each, but many are age or sex breakdowns. The number of populated rows is therefore shown alongside thematic coverage and gaps, rather than presented as proof of analytical completeness. [34]

The global occupational reference contains 427 ISCO-08 unit groups and preserves the source’s occupational means, dispersion and exposure classes. Its downloaded task table contains 3,265 records. This is distinct from the much larger task universe and survey samples described in the ILO methodology. Global occupation scores are not measurements of how many people in a particular city hold those jobs. [25][35]

For the US benchmarks, 4,003 detailed occupation-by-metropolitan-area rows provide employment, wages and employment-estimate precision where reported. The database excludes major-group totals from the detailed occupation matrix to avoid summing parent groups together with their children. Suppressed employment and wage values remain missing; they are not zero. [36]

The source exposure file for the US benchmark uses O*NET-SOC occupations. A match is accepted only when the complete seven-character SOC occupation code, including its hyphen, and normalized occupation title both match. Ambiguous and unmatched records remain unscored. No ISCO-to-SOC equivalence is invented. The share of metropolitan employment covered by accepted matches is reported, and unmatched employment is not assigned zero exposure. [37]

A physical-versus-cognitive dimension requires task evidence. Occupations combine manual, interpersonal and analytical activities; a nurse, architect or technician cannot be classified reliably from a single occupational title. This release leaves that analytical field unpopulated pending a documented coding rubric or validated task dataset. A future classification should permit mixed work and separate physical presence from technical automatability.

Likewise, separate tables of women, age groups and occupations do not establish their joint distribution. The number of women in a city and its count of software developers cannot reveal the number of female developers. Household income, household size and occupational dependence require appropriate joint or microdata before family-level effects can be estimated.

6. From evidence to scenarios

A local exposure measure can be specified as the employment-weighted average of occupational scores, using a specified source and a common reference period. Its coverage must be reported separately. For example, the mean among matched jobs uses matched employment in the denominator; it must never be labelled a complete metropolitan exposure score when some jobs are unmatched. A technical exposure score is not multiplied by population to estimate departures.

The demonstrator instead starts with a visibly hypothetical worker cohort, N. Users select a displacement fraction d, a local re-employment fraction r among displaced workers, and an outward-mobility fraction m among those not re-employed locally. These are editable assumptions, not parameters estimated by this study.

The accounting relationships are:

• Displaced workers: D = N × d.
• Workers continuing in existing employment: K = N − D.
• Displaced workers re-employed locally: R = D × r.
• Potential outward-moving workers: M = (D − R) × m.
• Workers remaining without local re-employment in this simplified scenario: U = D − R − M.
• Conservation check: K + R + M + U = N.

The equations describe one simplified allocation over a user-defined adjustment episode; they contain no calibrated calendar horizon. The labels deliberately stop short of estimating unemployment duration, future population or the probability of reaching a destination. A worker who moves might remain unemployed, and a worker who stays might leave the labour force. A fuller model will need additional states and observed transition rates.

The interface can explore an optional household assumption: each moving worker represents a distinct moving household, with an entered mean household size. Under that explicit condition, people associated with outward-moving households equal M multiplied by household size. This is a gross scenario quantity, not net migration. It excludes inward flows, births, deaths, shared households among workers and differential household migration. It is disabled until the assumption is acknowledged.

No formula in this edition estimates rent changes from worker departures. That connection requires household tenure, vacancy, supply response, incomes and prices. Similarly, office demand requires firm location and workplace arrangements, while transport demand requires origin–destination and commuting data. Listing these required inputs makes the urban mechanism testable without inventing numerical elasticities.

A separate household housing-budget stress test

A second demonstrator makes the housing mechanism explicit. Let Y be monthly disposable household income and C the entered monthly housing cash cost, both in the same currency. Let q be an assumed proportional reduction in total household income, with housing cash costs held constant. Baseline budget pressure is C / Y; after the income change it is C / [Y × (1 − q)]. For an illustrative income of 5,000, housing cost of 1,500 and income reduction of 20%, the ratio rises from 30% to 37.5%. This follows from arithmetic, not from an estimated AI effect.

The tool also shows disposable income remaining after housing. The chosen cash-cost basket may include payments that official affordability measures handle differently, including mortgage principal or housing allowances. This simple budget ratio is therefore not labelled an official Eurostat overburden statistic. It neither estimates default probability nor forecasts property prices. Source-compatible local rent, disposable-income, tenure and household data are required before replacing illustrative inputs with city-calibrated ones.

7. An index should follow validation

A useful planning index should identify interpretable dimensions before collapsing them into one number. The proposed dimensions are occupational exposure; dependence on external service demand; local transition capacity; household financial constraints; and urban absorptive capacity. The present release populates parts of this structure and leaves unobserved components explicitly unavailable.

Weights are substantive assumptions. Equal weighting is not automatically neutral, while statistical weighting does not automatically create causal meaning. Before producing a public ranking, the research must specify definitions, minimum data coverage, normalization, geographic comparability and sensitivity to alternative weights. It should also show whether a city changes rank when data years or exposure sources change.

A missing household measure must not raise a city’s apparent resilience merely because the missing field was converted to zero. A high exposure score must not by itself imply vulnerability if adoption increases demand and local workers adapt successfully. The preferred initial output is a multidimensional profile with visible evidence quality, followed by a validated composite only if it improves decisions.

8. Research tests and urban decisions

The first hypothesis is that cities concentrated in externally traded, highly exposed tasks face different labour-demand adjustments from cities with diversified demand, conditional on firm adoption. A test needs city-sector exports or employer-client evidence, occupational employment, and a credible measure of AI adoption. National ICT exports can inform the sampling frame but cannot substitute for that local treatment measure.

The second hypothesis is that local occupational adjacency and retraining opportunities moderate displacement and outward mobility. Testing it requires observed transitions or linked worker records, training access and subsequent employment outcomes. Similar occupation names are not enough to establish that workers can move between them at low cost.

The third hypothesis is that household resources condition mobility. Lower income may encourage a search for opportunities elsewhere while simultaneously making relocation unaffordable. Age, partners’ employment, dependants, housing tenure and legal mobility barriers should therefore be examined jointly rather than treated as independent arithmetic multipliers.

The fourth hypothesis concerns receiving cities. New arrivals may increase demand for housing and services, but effects depend on available supply, vacancy, infrastructure and local labour demand. A receiving-city model needs observed destinations and capacity constraints. Drawing an arrow between two economically similar cities is not evidence of a migration corridor.

Candidate empirical designs include longitudinal city panels, event studies around independently measured adoption, and worker transition models where access permits. Any causal design must address concurrent changes in trade, interest rates, remote work, conflict, education and national policy. Timing alone—an outcome occurring after the arrival of AI—does not establish attribution.

For an urban consultancy, the resulting questions are practical. Where should training provision be located? Which employment districts need flexible premises? Which household groups face income stress even without migration? Which receiving areas could need housing and transport capacity? The purpose is to make planning choices conditional on evidence, rather than present a dramatic map of inevitable decline.

9. Release limits and next empirical gates

This research edition provides a documented data foundation and an interactive scenario instrument. It is not a global city census, a demographic forecast, a completed causal study or a validated vulnerability index. Its strongest current elements are traceable source retrieval, distinct geographic units, occupational reference data, a benchmark labour matrix and explicit scenario accounting.

Before a city enters a published predictive comparison, it should have harmonized detailed occupational employment, relevant household structure and income data, consistent boundaries, observed mobility information, and a documented source-to-model transformation. Appropriate uncertainty, validation against withheld observations and comparison with a simple baseline must precede claims of predictive performance.

The 21 historical references remain available in full bibliographic form, together with the review depth and intended use of each. Their presence does not mean all studies provide equally strong evidence or that their effect sizes are transferable. The contemporary sources establish both the relevance of the question and the substantial prior work that a credible consulting study must acknowledge.

The project’s urban identity follows from what it explains: the relationship between changing livelihoods and the organization of urban life. Its strongest form is a testable connection between workers, households and places, supported by a database in which every observed value has a source and every simulated result has a visible assumption.

10. Source quality and reproducibility

The European layer contains 18,880 indicator-year observations and 3,001 latest city-indicator pairs. The Middle East layer adds 2,460 source values and 624 latest city-indicator pairs. Repeated years are never counted as new indicators. The separate US table contains 4,003 metro-occupation records. These are different evidence structures; their row counts are not a measure of comparable predictive completeness.

Zagreb's published population totals differ from the published male-plus-female sum by −6 people in 2018 and +1 in 2019, 2020, 2021 and 2024. Original values are retained and the discrepancies are disclosed. Bratislava, Budapest, Lisbon and Porto use source boundaries labelled “greater city”.

The Middle East source is pinned to JRC UCDB R2024A version 1.2. Each extracted value retains its original field, reference year, source identifier and definition note. Missing source values remain missing. The 2030 source projections are excluded from the displayed matrix; values for 2025 can themselves be source projections and must not be described as a municipal census.

The map uses Natural Earth land geometry and geographic coordinates. Middle East marker coordinates are transformed from the UCDB's declared World Mollweide coordinate reference system to longitude and latitude. They identify source urban-centre centroids, not exact city halls or administrative boundaries. No connecting line is interpreted as a migration flow. [41][42][43]

For review, the research text, evidence register, city tables and collection specifications remain separate components. Reproducibility requires the stated source versions, complete code mappings and an auditable record of transformations. A later update must record changes in boundaries and source definitions before interpreting differences as urban change.

References and evidence register

The 21 historical references are retained as [1]–[21], with the review depth and intended use recorded for each source.