{"id":326063,"date":"2026-09-22T14:40:00","date_gmt":"2026-09-22T11:40:00","guid":{"rendered":"https:\/\/ceotudent.com\/middle-management-after-ai-which-manager-tasks-automate-what-remains"},"modified":"2026-09-22T14:40:00","modified_gmt":"2026-09-22T11:40:00","slug":"middle-management-after-ai-which-manager-tasks-automate-what-remains","status":"publish","type":"post","link":"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains","title":{"rendered":"Middle Management After AI: Which Manager Tasks Automate Away and What Remains"},"content":{"rendered":"<p><strong>TL;DR.<\/strong> Alongside its 2025 to 2035 projections, the US Bureau of Labor Statistics released a new dataset assigning every projected occupation a relative AI exposure category, built by combining five external sources including two that use observed AI usage rather than theory. We downloaded it. Four findings. First, of 37 detailed management occupations, <strong>zero<\/strong> are rated low exposure, compared with 25.6% of all 831 occupations; 91.9% of management occupations are high or very high, against 49.6% overall. Managers are the most exposed occupational group in the economy. Second, BLS nonetheless projects management employment to grow 6.17% over the decade against 3.48% for all occupations, while first-line supervisors, the actual coalface of middle management, grow just 1.93%. Third, exposure barely predicts projected employment at all: across all 831 occupations the rank correlation between exposure category and projected growth is +0.103, and the median growth by category is not even monotonic. Fourth, at the task level the picture finally resolves. Across 843 core management tasks, only 11.3% are unexposed by agreement of both human raters and GPT-4, against 38.4% of all core tasks in the economy. But sorted by the task&rsquo;s leading verb, the gradient is clean: analyse and evaluate tasks are 63.7% exposed and 1.0% unexposed, while hire, train and discipline tasks are 12.8% exposed and 38.3% unexposed. The information half of the manager job is being automated. The accountability half is not.<\/p>\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_84 counter-hierarchy ez-toc-counter ez-toc-grey ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">Table of Contents<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#A-new-dataset-that-changes-what-can-be-said\" >A new dataset that changes what can be said<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#Managers-are-the-most-exposed-occupational-group-in-the-economy\" >Managers are the most exposed occupational group in the economy<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#The-paradox-most-exposed-fastest-projected-growth\" >The paradox: most exposed, fastest projected growth<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#So-go-to-the-task-level\" >So go to the task level<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#Which-tasks-exactly\" >Which tasks, exactly<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#The-managers-two-jobs\" >The manager&rsquo;s two jobs<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#Where-the-squeeze-actually-lands\" >Where the squeeze actually lands<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#What-a-manager-should-actually-do-with-this\" >What a manager should actually do with this<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-9\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#The-CEO-and-the-student-in-a-management-role\" >The CEO and the student in a management role<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-10\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#Limits-of-this-analysis\" >Limits of this analysis<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-11\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#FAQ\" >FAQ<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-12\" href=\"https:\/\/ceotudent.com\/en\/middle-management-after-ai-which-manager-tasks-automate-what-remains\/#Sources\" >Sources<\/a><\/li><\/ul><\/nav><\/div>\n<h2 id=\"a-new-dataset-that-changes-what-can-be-said\"><span class=\"ez-toc-section\" id=\"A-new-dataset-that-changes-what-can-be-said\"><\/span>A new dataset that changes what can be said<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Most writing about AI and management runs on anecdote, vendor surveys, or a single 2023 exposure paper applied loosely. That changed with the 2025 to 2035 projections round, when BLS introduced relative AI exposure categories for every detailed occupation it projects.<\/p>\n<p>What makes the dataset unusual is its construction. BLS did not build its own measure. It combined five external sources, three theoretical and two based on observed usage, converted each to a percentile rank so they sat on a common 0 to 1 scale, took the median of the theoretical ranks and the median of the observed ranks, and then clustered occupations into four categories: low, moderate, high and very high. The observed sources are real AI traffic mapped to occupational tasks, one from Anthropic&rsquo;s Claude usage and one from Microsoft Copilot data, as BLS describes them.<\/p>\n<p>BLS is also unusually blunt about what the categories are not. Its limitations section states that an exposure category &ldquo;is not a forecast of employment growth or decline&rdquo;, &ldquo;is not a worker replacement estimate&rdquo;, and that the categories &ldquo;do not distinguish between AI impacts from automation versus augmentation&rdquo;. It also states that the theoretical sources &ldquo;conceptualize AI capabilities as those available no later than mid-2023&rdquo;. Every number below inherits that ceiling, and we come back to it in the limits section.<\/p>\n<p>With those caveats attached, the file lets us ask a question that previously could only be answered with opinion: where does management actually sit?<\/p>\n<h2 id=\"managers-are-the-most-exposed-occupational-group-in-the-economy\"><span class=\"ez-toc-section\" id=\"Managers-are-the-most-exposed-occupational-group-in-the-economy\"><\/span>Managers are the most exposed occupational group in the economy<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>We filtered the file to the 37 detailed management occupations, meaning Standard Occupational Classification codes beginning 11, covering 13.67 million workers in 2025.<\/p>\n<table>\n<thead>\n<tr>\n<th>Group<\/th>\n<th>Occupations<\/th>\n<th>Low<\/th>\n<th>Moderate<\/th>\n<th>High<\/th>\n<th>Very high<\/th>\n<th>High or very high<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>All detailed occupations<\/td>\n<td>831<\/td>\n<td>25.6%<\/td>\n<td>24.8%<\/td>\n<td>24.8%<\/td>\n<td>24.8%<\/td>\n<td>49.6%<\/td>\n<\/tr>\n<tr>\n<td>Management (SOC 11)<\/td>\n<td>37<\/td>\n<td><strong>0.0%<\/strong><\/td>\n<td>8.1%<\/td>\n<td>62.2%<\/td>\n<td>29.7%<\/td>\n<td><strong>91.9%<\/strong><\/td>\n<\/tr>\n<tr>\n<td>First-line supervisors<\/td>\n<td>19<\/td>\n<td>10.5%<\/td>\n<td>21.1%<\/td>\n<td>57.9%<\/td>\n<td>10.5%<\/td>\n<td>68.4%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>Source: BLS AI exposure categories and 2025 to 2035 employment projections data file. Counts are of detailed occupations, not workers.<\/em><\/p>\n<p>Not one management occupation is rated low. Across the whole economy roughly a quarter are. On an employment-weighted basis, 89.7% of management employment sits in high or very high exposure occupations, against 56.0% of all employment.<\/p>\n<p>The three management occupations that escape the high bands are worth naming, because they make the mechanism visible: administrative services managers, food service managers, and farmers, ranchers and other agricultural managers. All three are moderate, and all three are the management jobs where the work is most physically present and least document-shaped.<\/p>\n<p>If you want to run this logic on your own role rather than on an occupational average, we built a procedure for that in <a href=\"https:\/\/ceotudent.com\/en\/how-to-audit-your-job-for-ai-replaceability\">how to audit your job for AI replaceability<\/a>.<\/p>\n<h2 id=\"the-paradox-most-exposed-fastest-projected-growth\"><span class=\"ez-toc-section\" id=\"The-paradox-most-exposed-fastest-projected-growth\"><\/span>The paradox: most exposed, fastest projected growth<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Here is where the story that usually gets written falls apart. If exposure meant displacement, management should be shrinking. BLS projects the opposite.<\/p>\n<table>\n<thead>\n<tr>\n<th>Group<\/th>\n<th>Employment 2025<\/th>\n<th>Projected change 2025 to 2035<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>All detailed occupations<\/td>\n<td>170.3 million<\/td>\n<td>plus 3.48%<\/td>\n<\/tr>\n<tr>\n<td>Management (SOC 11)<\/td>\n<td>13.67 million<\/td>\n<td>plus 6.17%<\/td>\n<\/tr>\n<tr>\n<td>First-line supervisors<\/td>\n<td>8.57 million<\/td>\n<td>plus 1.93%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>Source: same BLS data file; percentages computed from the 2025 and 2035 employment columns summed across each group.<\/em><\/p>\n<p>The most exposed occupational group in the economy is projected to grow at nearly double the all-occupation rate.<\/p>\n<p>And this is not a quirk of management. Exposure barely predicts projected employment anywhere. Across all 831 occupations, the Spearman rank correlation between exposure category and projected percentage growth is <strong>+0.103<\/strong>, which is close to nothing, and the direction is positive rather than negative. The medians are not even monotonic.<\/p>\n<table>\n<thead>\n<tr>\n<th>Exposure category<\/th>\n<th>Occupations<\/th>\n<th>Median projected growth, 2025 to 2035<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Low<\/td>\n<td>213<\/td>\n<td>plus 1.90%<\/td>\n<\/tr>\n<tr>\n<td>Moderate<\/td>\n<td>206<\/td>\n<td>plus 3.35%<\/td>\n<\/tr>\n<tr>\n<td>High<\/td>\n<td>206<\/td>\n<td>plus 3.35%<\/td>\n<\/tr>\n<tr>\n<td>Very high<\/td>\n<td>206<\/td>\n<td>plus 2.70%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>Source: computed from the BLS data file. Low-exposure occupations have the lowest median projected growth of the four categories.<\/em><\/p>\n<p>Read that table twice. The occupations BLS rates least exposed to AI are the ones it projects to grow slowest. Exposure is largely a measure of how document-shaped and language-shaped a job is, and document-shaped jobs also tend to be the growing, higher-wage ones. It is not a death sentence, and BLS says so explicitly.<\/p>\n<p>Within management alone the rank correlation is +0.316 and within first-line supervisors it is minus 0.237, but with 37 and 19 occupations respectively those are too small to interpret, and we report them only so the reader can see we looked.<\/p>\n<h2 id=\"so-go-to-the-task-level\"><span class=\"ez-toc-section\" id=\"So-go-to-the-task-level\"><\/span>So go to the task level<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Occupation-level exposure cannot answer &ldquo;which manager tasks automate away&rdquo; because an occupation is a bundle. To answer the actual question we went underneath it.<\/p>\n<p>The GPTs are GPTs study by Eloundou, Manning, Mishkin and Rock released a task-level labelset: every O<em>NET task statement, labelled for exposure both by trained human raters and by GPT-4, on a three-point scale where E0 means no exposure and E2 means exposed with software built around a language model. That file contains 19,265 task rows. We restricted to core tasks, meaning the tasks O<\/em>NET marks as central rather than supplemental, and compared management against the economy.<\/p>\n<p>We counted a task as <strong>agreed unexposed<\/strong> only when both the human raters and GPT-4 said E0, and <strong>agreed exposed<\/strong> only when both said E2. Requiring agreement is deliberately conservative: it throws away every task where the two rater types disagreed.<\/p>\n<table>\n<thead>\n<tr>\n<th>Task set<\/th>\n<th>Core tasks<\/th>\n<th>Agreed unexposed<\/th>\n<th>Agreed exposed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>All occupations<\/td>\n<td>13,225<\/td>\n<td>38.4%<\/td>\n<td>21.8%<\/td>\n<\/tr>\n<tr>\n<td>Management (SOC 11)<\/td>\n<td>843<\/td>\n<td><strong>11.3%<\/strong><\/td>\n<td><strong>42.0%<\/strong><\/td>\n<\/tr>\n<tr>\n<td>First-line supervisors<\/td>\n<td>328<\/td>\n<td>33.5%<\/td>\n<td>20.1%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>Source: computed from the GPTs are GPTs full labelset joined to O<\/em>NET task type. Every figure recomputed by an independent second script reading the raw files.*<\/p>\n<p>For the average job in the economy, unexposed tasks outnumber exposed ones by nearly two to one. For management, the ratio inverts almost exactly: exposed tasks outnumber unexposed by close to four to one.<\/p>\n<p>First-line supervisors, meanwhile, look like the economy average rather than like management. That is the first clue about where the line actually runs, and it is not the line between senior and junior.<\/p>\n<h2 id=\"which-tasks-exactly\"><span class=\"ez-toc-section\" id=\"Which-tasks-exactly\"><\/span>Which tasks, exactly<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Now the question the headline asks. We classified all 843 core management tasks by the leading verb of the task statement and computed exposure within each class. The classification is a simple, transparent heuristic on the first verb; it covers 570 of 843 tasks, or 67.6%, and the unclassified remainder are mostly compound statements that begin with something other than a clean action verb.<\/p>\n<table>\n<thead>\n<tr>\n<th>Manager task class<\/th>\n<th>Tasks<\/th>\n<th>Agreed unexposed<\/th>\n<th>Agreed exposed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Analyse or evaluate<\/td>\n<td>102<\/td>\n<td>1.0%<\/td>\n<td><strong>63.7%<\/strong><\/td>\n<\/tr>\n<tr>\n<td>Prepare, write or report<\/td>\n<td>71<\/td>\n<td><strong>0.0%<\/strong><\/td>\n<td>53.5%<\/td>\n<\/tr>\n<tr>\n<td>Plan or formulate strategy<\/td>\n<td>151<\/td>\n<td>2.6%<\/td>\n<td>50.3%<\/td>\n<\/tr>\n<tr>\n<td>Direct or coordinate<\/td>\n<td>109<\/td>\n<td>17.4%<\/td>\n<td>37.6%<\/td>\n<\/tr>\n<tr>\n<td>Confer, meet or represent<\/td>\n<td>68<\/td>\n<td>22.1%<\/td>\n<td>35.3%<\/td>\n<\/tr>\n<tr>\n<td>Hire, train or discipline<\/td>\n<td>47<\/td>\n<td><strong>38.3%<\/strong><\/td>\n<td>12.8%<\/td>\n<\/tr>\n<tr>\n<td>Inspect or physically check<\/td>\n<td>22<\/td>\n<td><strong>45.5%<\/strong><\/td>\n<td>13.6%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>CEOtudent classification of O<\/em>NET core task statements for SOC 11 occupations by leading verb, scored against the GPTs are GPTs human and GPT-4 labels. Covers 570 of 843 core tasks. The two smallest classes have fewer than 50 tasks and should be read as indicative.*<\/p>\n<p>The gradient is monotonic from top to bottom and it is not subtle. Of 71 core management tasks that begin with preparing, writing or reporting, <strong>not one<\/strong> is unexposed by agreement. Of tasks that begin with inspecting or physically checking, nearly half are.<\/p>\n<p>Here is what that looks like in the actual task language. On the exposed end, real O*NET core tasks for chief executives and general managers include analysing operations to evaluate performance and identify areas of potential cost reduction, preparing budgets for approval, reviewing financial statements and activity reports to measure productivity, and preparing reports concerning activities and expenses. These are the tasks that fill a manager&rsquo;s week and that a competent model now does quickly.<\/p>\n<p>On the unexposed end, the agreed-E0 management tasks read differently: appoint department heads and delegate responsibilities to them; preside over or serve on boards of directors; hire, train, evaluate or discharge staff or resolve personnel grievances; stop production if serious product defects are present; serve as a confidential point of contact for employees to report irregularities; represent the organisation at personnel-related hearings.<\/p>\n<p>Every one of those is an act of authority with a consequence for a named human being, or a physical presence that can be held responsible. None of them is an information problem.<\/p>\n<h2 id=\"the-managers-two-jobs\"><span class=\"ez-toc-section\" id=\"The-managers-two-jobs\"><\/span>The manager&rsquo;s two jobs<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>That is the finding underneath the numbers. The manager role bundles two different jobs that we have historically treated as one because the same person did both.<\/p>\n<p>The <strong>information job<\/strong> is gathering, analysing, synthesising and transmitting what is happening. Historically this was the bulk of the work and the source of the authority, because the person who held the information was necessarily the person who decided. This is the half sitting at 50 to 64% exposure.<\/p>\n<p>The <strong>accountability job<\/strong> is committing the organisation to a course of action, deciding about specific people, and being the person who carries the consequence. This is the half sitting at 12 to 14% exposure with a third to a half of tasks unexposed outright.<\/p>\n<p>AI does not automate the manager. It unbundles the manager, and it takes the half that used to justify the headcount.<\/p>\n<table>\n<thead>\n<tr>\n<th>Layer<\/th>\n<th>What it is<\/th>\n<th>Exposure evidence<\/th>\n<th>What happens to it<\/th>\n<th>What to do about it<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Reporting layer<\/td>\n<td>Producing the status update, the deck, the variance analysis, the weekly summary<\/td>\n<td>0.0% of prepare and report tasks unexposed<\/td>\n<td>Collapses first and fastest; this is the layer that shrinks<\/td>\n<td>Stop competing on it; a manager whose value is the report has no value<\/td>\n<\/tr>\n<tr>\n<td>Analysis layer<\/td>\n<td>Reading the numbers, spotting the pattern, forming the recommendation<\/td>\n<td>63.7% of analyse tasks exposed<\/td>\n<td>Becomes cheap and abundant; volume rises, differentiation falls<\/td>\n<td>Move from producing analysis to interrogating it; the scarce skill is knowing when the answer is wrong<\/td>\n<\/tr>\n<tr>\n<td>Coordination layer<\/td>\n<td>Sequencing work, resolving dependencies, scheduling<\/td>\n<td>37.6% exposed, 17.4% unexposed<\/td>\n<td>Partially absorbed by systems and agents<\/td>\n<td>Automate deliberately rather than defending it; it is the layer most worth handing over<\/td>\n<\/tr>\n<tr>\n<td>Judgement layer<\/td>\n<td>Committing under uncertainty, owning the call<\/td>\n<td>Low exposure by construction<\/td>\n<td>Grows in relative weight as the layers above thin<\/td>\n<td>This is where the remaining job is; practise deciding with incomplete information<\/td>\n<\/tr>\n<tr>\n<td>Accountability layer<\/td>\n<td>Hiring, developing, disciplining, being answerable for a person&rsquo;s livelihood<\/td>\n<td>38.3% of staffing tasks unexposed, 12.8% exposed<\/td>\n<td>Does not transfer; it is not a capability question but a legitimacy one<\/td>\n<td>Invest here; it is the least substitutable thing a manager does<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>CEOtudent editorial framework. The exposure column is measured from the sources above; the layer model and the recommendations are our argument, not a research finding.<\/em><\/p>\n<p>The practical reading: a manager whose week is mostly the top two layers is in a genuinely exposed position regardless of title or seniority. A manager whose week is mostly the bottom two is doing work that neither dataset can see a substitute for.<\/p>\n<h2 id=\"where-the-squeeze-actually-lands\"><span class=\"ez-toc-section\" id=\"Where-the-squeeze-actually-lands\"><\/span>Where the squeeze actually lands<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>The first-line supervisor numbers are where this stops being abstract, because that is where most middle management actually lives.<\/p>\n<table>\n<thead>\n<tr>\n<th>Occupation<\/th>\n<th>Employment 2025<\/th>\n<th>AI exposure<\/th>\n<th>Projected change to 2035<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>First-line supervisors of office and administrative support<\/td>\n<td>1.51 million<\/td>\n<td><strong>Very high<\/strong><\/td>\n<td><strong>plus 0.1%<\/strong><\/td>\n<\/tr>\n<tr>\n<td>First-line supervisors of retail sales workers<\/td>\n<td>1.42 million<\/td>\n<td>High<\/td>\n<td><strong>minus 3.7%<\/strong><\/td>\n<\/tr>\n<tr>\n<td>First-line supervisors of food preparation and serving<\/td>\n<td>1.24 million<\/td>\n<td>High<\/td>\n<td>plus 5.4%<\/td>\n<\/tr>\n<tr>\n<td>First-line supervisors of construction trades<\/td>\n<td>0.91 million<\/td>\n<td>Moderate<\/td>\n<td>plus 5.0%<\/td>\n<\/tr>\n<tr>\n<td>Medical and health services managers<\/td>\n<td>0.64 million<\/td>\n<td>High<\/td>\n<td>plus 24.2%<\/td>\n<\/tr>\n<tr>\n<td>Computer and information systems managers<\/td>\n<td>0.69 million<\/td>\n<td>Very high<\/td>\n<td>plus 15.8%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><em>Source: BLS AI exposure categories and 2025 to 2035 employment projections data file.<\/em><\/p>\n<p>The largest supervisory occupation in the country, office and administrative support supervisors, is rated very high exposure and projected to be flat at plus 0.1% while the economy adds 3.48%. Retail sales supervisors decline outright. These are the supervisors of exactly the work that is most document-shaped and most rule-shaped.<\/p>\n<p>But look at the last two rows. Medical and health services managers are rated high exposure and projected to grow 24.2%. Computer and information systems managers are very high and grow 15.8%. Same exposure bands, opposite trajectories.<\/p>\n<p>The variable that separates them is not AI. It is whether demand for the underlying service is growing. Exposure tells you which of your tasks will change. Demand tells you whether your job still exists. Conflating the two is the single most common error in this entire genre of writing, and the BLS data makes the separation unusually visible.<\/p>\n<p>This is the same structural pattern we found at the other end of the career ladder in <a href=\"https:\/\/ceotudent.com\/en\/the-entry-level-squeeze-disappearing-junior-roles\">the entry-level squeeze<\/a>: the roles built mainly on producing routine documented output are the ones compressing, at both the junior and the supervisory end.<\/p>\n<h2 id=\"what-a-manager-should-actually-do-with-this\"><span class=\"ez-toc-section\" id=\"What-a-manager-should-actually-do-with-this\"><\/span>What a manager should actually do with this<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Three moves follow from the evidence rather than from opinion.<\/p>\n<p><strong>Move your week down the stack.<\/strong> If the reporting and analysis layers are 50 to 64% exposed and the accountability layer is not, then time spent on the former is depreciating and time spent on the latter is appreciating. That is not a motivational claim, it is what the exposure gradient says. The practical test is simple: what fraction of your last two weeks produced a document, and what fraction produced a decision about a person or a commitment?<\/p>\n<p><strong>Hand over coordination deliberately.<\/strong> The coordination layer is the one most worth automating on purpose, because it is partially exposed and it consumes the most time for the least differentiation. Doing that well is a skill in itself, and we set out the concepts it requires in <a href=\"https:\/\/ceotudent.com\/en\/agent-literacy-7-concepts-before-delegating-work-to-ai-agents\">agent literacy<\/a> and the test for which meetings survive in <a href=\"https:\/\/ceotudent.com\/en\/meeting-that-should-have-been-an-agent-decision-protocol-sync-vs-automated-work\">the meeting that should have been an agent<\/a>.<\/p>\n<p><strong>Get better at reading output you did not produce.<\/strong> When analysis is cheap, the scarce capability is judging whether a given analysis is sound. That is a different skill from producing it and most managers have never had to practise it separately, because they produced the analysis themselves and therefore knew where the weak assumptions were.<\/p>\n<h2 id=\"the-ceo-and-the-student-in-a-management-role\"><span class=\"ez-toc-section\" id=\"The-CEO-and-the-student-in-a-management-role\"><\/span>The CEO and the student in a management role<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>The CEO half of this is straightforward and slightly uncomfortable. A chief executive&rsquo;s core tasks are among the most exposed in the entire dataset, and the exposed ones are exactly the tasks people imagine to be the job: analysing operations, preparing budgets, reviewing performance, producing reports. What is unexposed is appointing people, presiding over the board, and conferring to resolve problems. The job was never really the analysis. The analysis was the visible artefact of a role whose actual content is committing the organisation and owning the result.<\/p>\n<p>The student half is the part that determines who survives this. The layers that are thinning are the ones a person can become excellent at once and then coast on for a decade. The layers that are thickening, judgement under uncertainty and responsibility for other people&rsquo;s development, are the ones nobody ever finishes learning and where a twenty-year veteran can still be visibly better than a five-year one.<\/p>\n<p>Which is a genuinely good outcome for anyone who actually likes the job. The part of management that is being automated is the part most managers complain about. What remains is the part that was always supposed to be the point, and we looked at which capabilities compound that way in <a href=\"https:\/\/ceotudent.com\/en\/career-capital-ai-era-what-compounds-what-decays\">career capital in the AI era<\/a>.<\/p>\n<h2 id=\"limits-of-this-analysis\"><span class=\"ez-toc-section\" id=\"Limits-of-this-analysis\"><\/span>Limits of this analysis<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>Four, stated plainly, because they matter more than usual here.<\/p>\n<p><strong>The capability ceiling is mid-2023.<\/strong> BLS states that the theoretical sources conceptualise AI capabilities as those available no later than mid-2023, and the Eloundou labels we used at the task level are GPT-4-era. Model capability has moved since. The honest reading is that these figures are a lower bound on exposure, particularly for the analysis and writing classes, and that the unexposed classes are the more durable finding because a capability increase is less likely to move a task that is unexposed for reasons of authority rather than difficulty.<\/p>\n<p><strong>Exposure is not displacement.<\/strong> This is BLS&rsquo;s own warning and we are repeating it rather than softening it. Nothing here shows any manager losing a job. The employment column is a projection, and the previous article in this series on <a href=\"https:\/\/ceotudent.com\/en\/vision-horizon-how-far-ahead-should-you-plan-ai-era\">the vision horizon<\/a> documents how badly ten-year BLS projections have missed before.<\/p>\n<p><strong>The verb classification is ours, not a standard.<\/strong> It is a transparent heuristic on the leading verb of each task statement, covering 67.6% of core management tasks. A different classifier would produce somewhat different cell values. What we would expect to survive any reasonable reclassification is the ordering, because the gap between the top and bottom classes is close to fifty-one percentage points on the exposed measure and more than forty-five on the unexposed one.<\/p>\n<p><strong>Tasks are treated as independent.<\/strong> BLS flags this about all five of its sources, and it applies to our task-level work too: the data does not model bottlenecks or complementarities. A job where 42% of core tasks are exposed is not a job that is 42% done, because the remaining tasks may be the ones that gate everything else. In management specifically, we think that is exactly what is happening, but the data cannot prove it.<\/p>\n<h2 id=\"faq\"><span class=\"ez-toc-section\" id=\"FAQ\"><\/span>FAQ<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p><strong>Is middle management disappearing?<\/strong><br \/>\nNot according to the numbers. BLS projects management employment up 6.17% over 2025 to 2035 against 3.48% for all occupations. What is happening is narrower and more specific: first-line supervisors grow only 1.93%, and the most document-shaped supervisory occupations are flat or declining, with office and administrative support supervisors at plus 0.1% and retail sales supervisors at minus 3.7%.<\/p>\n<p><strong>If managers are the most AI-exposed group, why is their employment growing?<\/strong><br \/>\nBecause exposure measures how much of the work a model could assist with, not whether the work is still needed. Across all 831 occupations the correlation between exposure and projected growth is +0.103, essentially nothing, and low-exposure occupations actually have the lowest median projected growth. Demand for the underlying service drives employment; exposure drives what the job consists of.<\/p>\n<p><strong>Which manager tasks are actually safe?<\/strong><br \/>\nOn these labels, the ones that involve authority over a specific person or physical presence. Hiring, training, evaluating and discharging staff, resolving grievances, presiding over governing boards, appointing and delegating to department heads, serving as a confidential reporting channel, and inspecting conditions on site. Staffing tasks are 38.3% unexposed against 12.8% exposed, the reverse of the manager average.<\/p>\n<p><strong>I am a first-line supervisor of office work. What does this say about me?<\/strong><br \/>\nThat your occupation is rated very high exposure and projected at plus 0.1% growth while the economy adds 3.48%, which is the sharpest single signal in the dataset. It does not say your job disappears. It says growth in your specific occupation is not where the market is heading, and that the part of your role most worth deepening is the people-accountability half rather than the coordination and reporting half.<\/p>\n<p><strong>Does this mean I should stop producing analysis?<\/strong><br \/>\nNo. It means stop treating the production of analysis as the thing that makes you valuable. The scarce and appreciating skill is judging whether an analysis is sound, deciding what to do about it, and owning that decision. Producing the analysis is now the cheap step.<\/p>\n<h2 id=\"sources\"><span class=\"ez-toc-section\" id=\"Sources\"><\/span>Sources<span class=\"ez-toc-section-end\"><\/span><\/h2>\n<ul>\n<li>US Bureau of Labor Statistics, Employment Projections program, Artificial intelligence exposure categories, published alongside the 2025 to 2035 projections. Methodology combining three theoretical and two observed-usage sources into percentile ranks and a four-category clustering; the limitations section stating that an exposure category is not a forecast of employment growth or decline, is not a worker replacement estimate, does not distinguish automation from augmentation, and that the theoretical sources conceptualise AI capabilities as those available no later than mid-2023; the statement that all five sources treat occupations as bundles of independent characteristics without modelling bottlenecks or complementarities.<\/li>\n<li>US Bureau of Labor Statistics, AI exposure categories and 2025 to 2035 employment projections, data file. Relative AI exposure category, 2025 and 2035 employment, projected percentage change, median annual wage and typical entry education for 831 detailed occupations; all group counts, employment sums, correlations and medians in this article computed from this file.<\/li>\n<li>US Bureau of Labor Statistics, Employment Projections 2025 to 2035 news release. Total employment projected to rise from 170.3 million to 176.2 million, growth of 3.5 percent, slower than the 10.9 percent recorded over 2015 to 2025; the announcement of the AI exposure categories as a new data product.<\/li>\n<li>Eloundou, Manning, Mishkin and Rock, GPTs are GPTs, task-level labelset. 19,265 O*NET task rows with human aggregate exposure labels and GPT-4 exposure labels on the E0 to E2 scale, plus task type and occupational code; the basis for every task-level figure and for the agreed-unexposed and agreed-exposed definitions used here.<\/li>\n<li>O*NET database version 31.0, Occupational Information Network, US Department of Labor. Task statements and core versus supplemental task type designations for management occupations; the task language quoted in this article.<\/li>\n<li>Felten, Raj and Seamans; Eisfeldt and colleagues; and the observed-usage measures derived from Anthropic Claude conversations and Microsoft Copilot activity. These are the remaining four of the five inputs to the BLS exposure categories and are described here only as BLS characterises them in its methodology, not from the underlying papers.<\/li>\n<\/ul>\n<hr>\n<p><em>This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>In September 2026 the US Bureau of Labor Statistics published AI exposure categories for every occupation it projects. We pulled the file and found something nobody has reported: of the 37 detailed management occupations, exactly zero are rated low exposure, against 25.6% of all occupations. Managers are the most AI-exposed occupational group in the US economy. And BLS simultaneously projects management employment to grow 6.17% over 2025 to 2035, nearly double the 3.48% all-occupation rate. To explain the paradox we went to the task level, classifying 843 core manager tasks from the O*NET database against the exposure labels from the GPTs are GPTs study, and found a clean gradient: manager tasks that begin with analyse, prepare or plan are 50 to 64% exposed with almost none unexposed, while tasks that begin with hire, train, discipline or inspect invert completely. The manager job is two jobs, and only one of them is being automated.<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,18],"tags":[],"class_list":["post-326063","post","type-post","status-publish","format-standard","hentry","category-is","category-strateji"],"_links":{"self":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/posts\/326063","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/comments?post=326063"}],"version-history":[{"count":0,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/posts\/326063\/revisions"}],"wp:attachment":[{"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/media?parent=326063"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/categories?post=326063"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ceotudent.com\/en\/wp-json\/wp\/v2\/tags?post=326063"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}