What AI Job Studies Show About Task Change for Urban Planners

What do AI job studies actually measure?

When a headline says an occupation is “exposed” to AI, it describes a modeled relationship between technology and tasks, not how many workers have adopted a tool or will lose a job. For urban and regional planners, a task may look technically assistable on paper while remaining unsuitable, unapproved, or too costly in a public planning workflow. Exposure models estimate task applicability; usage datasets classify interactions on particular AI services; case databases record examples; labor statistics describe employment and duties. None alone estimates a planner’s probability of losing a job.

A useful reading rule is to separate six links. Technical capability means what a system can do under specified conditions. Task exposure estimates whether that capability could apply to described work. Observed use means a recorded interaction with a tool. Employer adoption means an organization has incorporated it into a workflow, with rules, staff, and resources. Labor demand concerns employment, openings, and hiring conditions. Displacement means jobs or hours have changed because of technology. Evidence for one link does not establish the next: a model’s ability to draft text does not prove that a planning agency uses it, that planners’ total work falls, or that staffing changes as a result.

The International Labour Organization’s 2025 update is an exposure study, not a survey of planning departments. It refined a global assessment using task-level data, expert input, and AI predictions across nearly 30,000 tasks, evaluated at detailed occupational levels in the ISCO-08 classification. The brief describes four exposure gradients and reports that one in four workers worldwide is in an occupation with some degree of generative-AI exposure. It says most jobs are more likely to be transformed than made redundant because human input remains necessary. These findings describe modeled potential globally; they do not isolate U.S. urban and regional planners, measure adoption, or count jobs lost.

The ILO’s 2026 explanation says exposure indicators rely on static descriptions of current tasks, reflect assumptions, and do not account for economic feasibility or adoption constraints. They are an early signal of where work may change, to be considered alongside labor-market developments and institutional conditions. A planner’s actual work also depends on jurisdiction, specialization, public-process duties, and agency policy, details an occupation-level model may not capture.

When reading an AI-and-jobs claim, ask: What is counted—tasks, interactions, occupations, jobs, or hours? Which population and place does it cover? What period does it describe? What outcome was measured? A task signal can justify inspecting a recurring duty; it cannot establish whether a particular department is changing staffing. That would require evidence connecting capability to use, organizational adoption, changes in work or demand, and employment outcomes. This distinction lets a planner examine a workflow without mistaking a reason to investigate for a prediction about a career.

Sources: Generative AI and Jobs: A 2025 Update; New ILO brief explains what AI exposure indicators reveal about jobs

Why does the planner job title hide different exposure patterns?

Urban and regional planning is a bundle of information, analysis, communication, place-based, and decision-support tasks. Some recurring document and data work may be candidates for AI assistance, but the surrounding checks, local context, regulation, and public responsibility shape what assistance can safely mean. A job title is too broad to tell an individual planner which parts of a working week may change. The U.S. Bureau of Labor Statistics (BLS), in “Urban and Regional Planners,” describes planners as gathering and analyzing data, assessing development site plans, staying current on zoning and other rules, meeting with officials, developers, and the public, and recommending whether plans should be approved or denied. It also notes that planners use statistical, data visualization, financial-management, and GIS software. These duties combine information handling with interpretation and decisions that affect places and communities. The same page says planners may specialize in transportation, community development, historic preservation, or other areas. The U.S. Department of Labor’s O*NET profile for “Urban and Regional Planners” supplies a standardized occupational description and task statements. That distinction matters when considering document search, organizing data, drafting report language, maintaining map or database records, or grouping comments as possible assistance points. A first draft, summary, or organized record may still need comparison with source material, correction, and approval. Other parts of planning work depend on circumstances that a generic description cannot fully capture. A planner may need to visit a proposed site, connect records to conditions on the ground, interpret rules that apply in a particular jurisdiction, explain trade-offs to stakeholders, or make and defend a recommendation. BLS specifically lists site visits as part of the work environment and stakeholder meetings and site-plan recommendations among typical duties. It also describes communication, decision-making, and interpersonal skills as important. This does not mean that these activities are beyond technological assistance; it means the job includes accountability and context around information tasks. A tool’s ability to produce text or sort records does not transfer responsibility for the resulting planning judgment. The same title can also hide different task mixes and different costs of error. A transportation planner might spend recurring time with mobility data and network plans; a permit reviewer may focus on applications, applicable rules, and site-plan details; a comprehensive-plan engagement lead may spend more time organizing input and explaining how it informs policy. These are illustrations of how specializations named by BLS can differ, not measured comparisons of their hours or exposure. An error in a draft internal note, a missed constraint in a permit review, and a distorted account of public input do not carry identical consequences. The relevant boundary depends on what the task is for, what source can verify it, and who must answer for the final decision. O*NET and BLS descriptions are useful maps of occupational duties, not time-use shares or a census of any one planner’s week. They group workers under a common classification and cannot show whether a particular role is mostly GIS analysis, public engagement, development review, or project coordination. They also cannot establish local employer rules, data restrictions, staffing plans, or adoption. For a planner deciding what an exposure estimate means personally, the next step is therefore to list recurring duties in their own role, note which are consequential or difficult to verify, and keep the title-level description in its proper place: a starting point for questions, not a verdict about the worker.

Sources: Urban and Regional Planners; Urban and Regional Planners

What do observed-use studies add—and what can’t they tell planners?

Observed-use studies add a record of what people actually bring to particular AI services, which is a different question from what a model might be capable of doing in theory. They can show that users seek help with varied work tasks and whether recorded interactions look more collaborative or more delegated under a study’s coding scheme. They cannot, by themselves, tell a planner how many colleagues use AI, whether an agency has adopted it in an approved workflow, or whether a task or job has disappeared. The difference matters: a conversation is evidence of interaction, not a staffing decision or a verified work product. Google’s July 2026 report, “Understanding the AI economy,” describes ATLAS v1.0, an analysis of about 15 million aggregated interactions across the Gemini app, AI Mode, and Gemini API. Its authors classify workplace interactions and report that, within a typical job, AI appears in roughly 21% of tasks; fewer than 10% of workplace interactions are classified as fully automating tasks. The report describes much use as collaborative assistance, including ideation, information retrieval, and learning. The supporting “Methodology” page specifies that the underlying sample contains 14,653,926 de-identified interactions from April 6–19, 2026, mapped to U.S. occupational and task taxonomies. “Typical job” is an aggregate description, not a measurement of urban planners, a planning department, or each worker’s share of time. The less-than-10% figure concerns interaction classifications, not the percentage of jobs automated, hours removed, or people displaced. Anthropic’s “Economic Index: New building blocks for understanding AI use,” published in January 2026, analyzes a different service and period. It samples November 2025 conversations, predominantly using Claude Sonnet 4.5, and reports augmentation in 52% and automation in 45% of Claude.ai conversations in that sample. The report also draws on one million sampled Claude.ai conversations and one million first-party API transcripts, while the 52/45 comparison is specifically for Claude.ai. These labels are not interchangeable with ATLAS’s “fully automating” category: the studies use different platforms, collection windows, taxonomies, and units. Nor should the figures be combined into a single estimate. Anthropic says Claude.ai captures mostly consumer use and its API mostly business use, but both sets reflect interactions that reached one provider. Its report also cautions that users select tasks they are more confident will work and can split larger tasks into smaller steps with feedback. That selection shapes what usage logs reveal; it does not mean each observed attempt succeeded or that the workflow was adopted by an employer. Read together, the studies provide a useful constraint on sweeping claims. In the observed windows, the recorded work interactions do not amount to evidence that whole occupations are already being automated across the labor market. They do show multiple ways people engage with AI, with assistance and delegation both present. Yet neither dataset supplies a representative survey of planners, a count of planning agencies using AI, or an evaluation of planning output after human review. The records cannot establish net time saved once checking is included, the quality of a public-facing decision, future demand, or job displacement. They also exclude workers whose tasks never enter those particular services, including people bound by policy or data restrictions. For an urban or regional planner, treat these reports as evidence about available patterns of use, not local adoption rates or labor forecasts. A planner considering help with research, drafting, or organizing information still needs evidence from their own permitted workflow: whether the output is accurate enough, what it omits, how much review it needs, and who remains accountable. Platform-level use can make a task worth inspecting; it cannot decide whether that task changes in a particular office. The distinction is practical. A signal to examine work is not a reason, on its own, to infer that a planning role is being eliminated or to make a costly career move.

Sources: Understanding the AI economy; ATLAS Methodology; Economic Index: New building blocks for understanding AI use

What does the public-comment example reveal about task change?

The American Planning Association’s “Making Sense of Emerging Practice: AI in Planning Use Cases” describes one concrete example: Bowling Green, Kentucky, used a month-long online polling and commenting forum for its BG 2050 Comprehensive Plan. The forum used machine learning both to pose questions to participants and to understand and synthesize large amounts of community feedback into material plan makers could use in the long-range plan. This is evidence that AI-supported synthesis has been tried within a planning engagement workflow. It does not show that a system made the plan, replaced public participation, or changed staffing.

The example is easiest to assess by separating the workflow into steps. Residents submit input; a system helps organize or synthesize it; planners and plan makers interpret the resulting material; and the public process still has to connect that interpretation to decisions and explanations. Each handoff raises a different question. Can a summary be traced to the comments behind it? Does it preserve disagreement as well as recurring themes? Can staff identify what the synthesis left uncertain, and explain how public input informed the plan? The APA account establishes the forum’s purpose and use of machine learning for synthesis. It does not report a measured comparison with a manual process, an accuracy audit, or a result showing that every concern was represented.

That limit matters because condensation changes what is easiest to notice. Repeated themes may become simpler to scan, while an uncommon concern, a conflict between groups, sarcasm, or a place-specific detail could be less visible after comments are grouped. These are verification risks inferred from the nature of summarization, not failures reported about Bowling Green’s system. A defensible review would keep the original comments available, trace summary points back to source material, and inspect comments that appear to disagree with or fall outside the main themes. Planners could record where human interpretation altered, qualified, or rejected an output. Such checks would help assess coverage; they would not by themselves prove that the process was fair or that all residents had an equal chance to participate.

The wider “AI in Planning Use Cases Database” gives context but not a prevalence estimate. APA says it contains more than 70 examples, is updated quarterly, and includes cases from planning and local government in the United States and abroad. The organization also cautions that publicly available examples are only part of the work and experimentation underway, with pilots and internal tools sometimes undocumented. The collection can show that different applications exist and suggest questions for local governance. It has no representative denominator from which to calculate how many planners or agencies use these methods, and it does not establish employment effects.

For a planner considering a similar workflow, the practical question is therefore narrow: does assistance make this specific engagement easier to review without losing the source, context, or disagreement that public decisions require? Compare the time and quality of the assisted process with the existing method, including the time needed to check it. Any test should follow employer policy and protect sensitive information, public records, and local procedures. The APA case supports examining a task and its safeguards; it cannot establish measured time saved, validated sentiment accuracy, or a case for replacing deliberation.

Sources: Making Sense of Emerging Practice: AI in Planning Use Cases; AI in Planning Use Cases Database

Why do planner-specific exposure claims need extra scrutiny?

A planning-specific number can still be a modeled estimate rather than a measured planning outcome. The roughly one-third figure in “Attention economics, artificial intelligence, and the future of the planning profession” is explicitly tentative: it is not a survey finding about planners, a count of adopted systems, or a measure of jobs lost. Reading it carefully means following how the number was produced, then stopping before the paper’s occupational task estimate is turned into a workforce forecast.

The paper first describes “GPTs are GPTs: Labor market impact potential of LLMs,” which uses occupational tasks in the U.S. O*NET database to assess how language-model capabilities might affect work. The Science article proposes a framework for estimating potential task impacts, using human and model judgments. Its abstract reports estimates of jobs whose tasks could be affected under different capability assumptions, including a larger share when complementary software is considered. These are modeled exposure estimates. They do not observe planning offices adopting tools, calculate net hours saved after review, or track whether employers reduce staff. The study presents a way to assess potential impacts, not a measure of labor-market outcomes.

The planning essay then makes a separate step. It says its author applied the Eloundou et al. method to the O*NET profile for Urban and Regional Planners, using a language model, and describes the resulting conclusion as tentative. The paper does not present this as a new representative study of planners or a direct measurement of their workdays. The distinction matters: the original study provides a general task-assessment framework; the planning paper applies that framework to one occupation’s standardized task descriptions; neither stage records actual employer decisions or displacement. The estimate should therefore be described as a tentative task-exposure result from an indirect application, not as the percentage of planners who will lose work.

Why can’t an exposed-task share be translated directly into a share of workers? An occupational profile lists activities, but not how much time every planner spends on each, how tasks combine in a particular job, or whether a tool performs them reliably in local conditions. A task may be assisted while its worker remains responsible for checking sources, applying regulations, handling exceptions, and explaining choices. A department may use time saved to process more work or redirect attention; it may also change staffing. Those outcomes depend on budgets, demand, implementation, and management decisions that a capability rubric does not measure. Further evidence is needed before making a workforce forecast.

The paper is useful for a different reason. It discusses how tools might alter professional attention, including the possibility that routine-task assistance could leave more capacity for issues receiving too little consideration. This is a possible direction for practice, not evidence of a benefit across planning departments. The article also considers less favorable possibilities and emphasizes that adoption choices matter. That interpretation prevents reading the estimate only as a threat while preserving uncertainty about whether benefits will materialize.

For planners, when a role-specific exposure number appears, check whether it comes from observed workplaces or from applying a capability rubric to occupational descriptions. Ask what the measure counts and whether it reports adoption, productivity after verification, staffing, or employment outcomes. Evidence supports a tentative estimate about potential applicability to listed planning tasks and a discussion of workflow change. It does not establish how common exposure is in a planner’s job, how many employers will implement AI, or how many workers may be displaced. Keep uncertainty layers intact; do not convert exposed tasks into exposed people or lost jobs.

Sources: Attention economics, artificial intelligence, and the future of the planning profession; GPTs are GPTs: Labor market impact potential of LLMs

Does the labor outlook contradict task-exposure studies?

No. The U.S. Bureau of Labor Statistics (BLS) projects that employment of urban and regional planners will grow 4 percent between 2025 and 2035, about as fast as employment overall. It estimates an average of roughly 3,200 openings each year over that decade. Those figures describe expected occupation-wide employment and openings; they do not measure whether AI is changing particular planning tasks, how widely agencies will adopt it, or how many jobs a technology might eliminate. An exposure estimate and an employment projection answer different questions, so they can both be accurate without contradicting one another. The distinction matters because a job title is an aggregate category, while employment totals reflect the combined effect of many forces. BLS says demographic, transportation, and environmental changes are expected to drive demand for planners. It points to population growth and change, housing needs, transportation systems, revitalization, environmental degradation, resource scarcity, and the need to develop or improve infrastructure such as roads, sewer systems, parks, and schools. It also cautions that federal, state, and local government budgets may affect public-sector planner employment because development projects depend on available funds. The projection therefore reflects a mix of anticipated work and fiscal conditions, not a controlled estimate of AI’s contribution. This allows task change and occupational growth to occur at the same time. A planning office might alter how it searches documents, organizes data, or prepares a first draft while still needing staff to assess site conditions, interpret local rules, discuss trade-offs with stakeholders, and explain recommendations. This illustrates how the measures can coexist; BLS did not observe this workflow or predict adoption. If assistance changes the time needed for one task, the effect on staffing would still depend on what work follows, whether demand for planning services changes, what budgets permit, and how the organization redesigns responsibilities. The occupation-level projection does not resolve those steps. Nor should annual openings be read as the number of newly created positions. BLS expects many openings to arise when workers transfer to other occupations or leave the labor force, including through retirement. Openings combine replacement needs with employment change; they are not an AI-specific count, nor a promise of vacancies in a particular city or agency. A positive projection means BLS expects higher total U.S. employment in 2035 than in 2025 under its assumptions. It does not mean every specialization, employer, region, or career stage will experience the same conditions. For a planner deciding what to do, the projection is useful context about broad U.S. demand, not an answer about personal exposure or job security. Local budgets, current vacancies, procurement rules, and actual changes in assigned duties are more relevant to a decision about a particular workplace. Task-exposure research can identify duties worth examining, but cannot explain BLS’s projected growth or show AI has no effect. The practical reading is modest: do not treat the 4 percent figure as reassurance that work will remain unchanged, and do not treat an exposure result as evidence that the occupation is shrinking. Use labor projections for the outlook question and task-level evidence for the workflow question; neither establishes causation between AI and employment. BLS’s figures cover the United States and its 2025–2035 projection period, so they should not be treated as a local forecast for another country or as a guarantee for any individual planner.

Sources: Urban and Regional Planners

What is a proportionate response to the evidence?

Start with one recurring task whose inputs and output can be checked. Compare the whole workflow, including review, with the method already in use. Consider training or a career move only after identifying a real capability gap or a goal that matters to you. This is a practical recommendation drawn from the evidence, not an intervention proven in a study. Exposure is a reason to inspect work, not an instruction to leave a profession. First, map the task before choosing a tool. Note how often it occurs, what information it uses, how sensitive that information is, what an error could affect, and whether a qualified person can verify the result against its source. A planner might list research, report preparation, data organization, or public-comment synthesis alongside consequential duties such as assessing a site plan or explaining a recommendation. The Bureau of Labor Statistics’ “Urban and Regional Planners” profile describes planners gathering and analyzing data, meeting stakeholders, assessing site plans, and recommending approval or denial. That profile helps locate a bounded administrative step; it does not show how much time a particular planner spends on it or whether AI handles it well. If a task uses confidential records or produces a decision that cannot be independently checked, the sensible result may be not to test it. Second, test only when workplace policy permits and the material can be handled appropriately. One possible trial is a first-pass summary of public comments already available for a planning process. Define what it must preserve: recurring themes, disagreements, location-specific concerns, and unusual but relevant comments. Keep the originals available. Check the draft against them for omissions and conflicts, then record the review time. The American Planning Association’s “Making Sense of Emerging Practice: AI in Planning Use Cases” documents a Bowling Green example in which machine learning supported synthesis of public feedback for the BG 2050 comprehensive-plan process. This shows that the workflow has been used; it does not report a controlled comparison of completeness, bias, time saved, or staffing effects. The checks here are proposed safeguards, not findings attributed to that case. Public-record obligations, privacy rules, procurement requirements, and local procedures still govern. Third, compare the complete process with the current approach. Ask whether the result is accurate enough for its intended use, whether it leaves out important concerns, whether a planner can explain the summary, and whether review consumes the time supposedly saved. Google’s “ATLAS Methodology” describes classifications of interactions on Google AI services; those observations do not measure the quality or net time cost of a planning department’s workflow. A permitted, documented local comparison speaks more directly to that question, though one small trial cannot establish broad workforce effects. Stop or redesign the test if checking cannot catch errors, the output obscures context, data handling creates unacceptable risk, or review adds more work than it saves. That result is useful too: this task may be a poor candidate under current conditions. If the workflow changes and a specific gap remains, choose the smallest learning step that addresses it: practice data checking, improve GIS work, learn to evaluate structured outputs, or study an approved tool in one named planning task. No route is automatically necessary because a study flags exposure. Durable skills include framing the planning question, checking data and sources, judging omissions, and explaining trade-offs; a particular interface may change. A larger career move is separate. Consider it when it fits your interests or circumstances, or when evidence from your role reveals a persistent mismatch that a bounded upgrade will not address. Compare options against your salary floor, geography, training time and cost, health, family responsibilities, and required credentials. The proportionate next step may be a reversible workflow test, a targeted learning project, or no test because policy or verification makes it unsuitable. An occupation-level exposure signal does not establish your personal likelihood of job loss.

Sources: Urban and Regional Planners; Making Sense of Emerging Practice: AI in Planning Use Cases; ATLAS Methodology

What should a planner do next?

Choose one recurring task, check whether your employer permits a test with the relevant data, and record whether quality and review time improve. That is a more defensible next step than treating occupational exposure as a forecast about your position. The evidence reviewed here does not support a planner-wide displacement conclusion. Start with a task that occurs often enough to evaluate, such as preparing a first-pass summary of public comments. Define what a useful result must preserve: major themes, conflicting views, less common concerns, and links to original comments. Ask who can verify these points. The American Planning Association’s AI in Planning Use Cases Database documents planning examples. It shows that experiments exist, not that they are accurate, broadly adopted, or proven to save labor. Before moving material, check privacy, public-record obligations, procurement terms, data governance, and agency policy. If any rule is unclear, pause and ask the appropriate internal owner. If a test is permitted, compare the output with the existing method and note omissions, corrections, and checking time. A faster first draft is not an improvement if verification erases the time saved or important context disappears. If the test exposes a concrete gap, choose a bounded learning step tied to it, such as improving data checking or evaluating AI-assisted summaries. A broader role change should reflect your goals and constraints too. The task checker organizes task-level change-pressure signals and first actions; it does not estimate job-loss probability. The career roadmap can compare staying and redesigning with adjacent paths against salary, location, and training constraints, without guaranteeing employment or income.

Sources: AI in Planning Use Cases Database

Questions readers ask