Firm Size, Task Modularization, and the Uneven Labor-Market Effects of Artificial Intelligence
Keywords: artificial intelligence · automation · augmentation · firm size · task modularization · entry-level employment · hiring suppression
Empirical appendix: Available as a separate online supplement. All errors are my own. Comments welcome.
This paper argues that AI's labor-market impact is mediated not only by occupation-level exposure, but by the organizational architecture in which the occupation is embedded. The same "financial analyst" job presents a different automation surface in a modularized large firm — a narrow codifiable slice — than in a bundled small firm where one person handles finance, collections, and sales support at once. AI does not encounter "the analyst." It encounters a task bundle, and the bundle is set by firm structure. Large firms have already decomposed white-collar work into standardized, codifiable units, so AI operates on the automation margin, compressing routine and junior layers. Small firms bundle tasks into generalist roles and operate under scarcity, so AI more often augments incumbents while suppressing the marginal entry-level hire. The predicted outcome is not mass unemployment, but a thinning of the entry ladder: stable payroll aggregates alongside fewer junior roles, lower routine shares, and delayed cohort formation — breaking the career ladder before it breaks the payroll aggregate. I treat firm size as an observable proxy for modularization, not its cause. Six pre-specified hypotheses follow; the most distinctive predicts hiring suppression — the non-created job that never appears in separation data. The public evidence assembled here is consistent with this mechanism but does not confirm it: aggregate postings show no uniform white-collar entry-port collapse, and the within-role weakening that does appear concentrates in the roles whose entry tasks are most codified and tool-automatable (software, IT, routine legal) — precisely where modularization and task separability are highest — while AI adoption is itself steeply concentrated in large firms. The pooled entry-port collapse is not in the data; the automatable-entry-port collapse is. I specify a firm–occupation–time triple-difference design, document what the currently public evidence can and cannot establish, and identify the firm-level test that remains to be run.
The dominant public framing of AI and work poses a binary question: does AI destroy jobs, or does it raise productivity? This framing maps onto a long-running scholarly tension between displacement and complementarity. Yet as an empirical matter the binary has proven curiously unsatisfying. Aggregate U.S. employment has continued to grow through the diffusion of generative AI, which seems to falsify the destruction story; at the same time, a growing body of micro-evidence finds sharp, localized contractions in specific worker groups, which seems to falsify the benign-productivity story. Both sides can point to data. Neither side's data settles the question.
This paper argues that the binary is mis-specified because it omits the organizational unit through which AI is actually adopted and deployed: the firm. The same nominal occupation is embedded in radically different organizational structures across the firm-size distribution. A "financial analyst" in a 50,000-person firm performs a narrow, standardized slice of a fragmented workflow; a "financial analyst" in a 20-person firm is often the only person doing finance, marketing, and collections at once. AI does not encounter "the analyst." It encounters a task bundle, and the composition of that bundle is a function of firm size.
AI displaces labor where work has already been made machine-readable. The relevant moderator is task modularization — the degree to which an organization has decomposed human work into standardized, codifiable units. Large firms have modularized white-collar work more, so the automation surface is larger there; small firms bundle work into generalist, tacit-heavy roles, so AI augments rather than replaces. Firm size predicts modularization and therefore predicts displacement, but does not cause it.
The claim is not the brittle "big firms cut, small firms create." It is that modularization is the causal object and size is its most observable proxy. The distinction matters empirically: a highly modularized small firm should behave like a large one, and a large firm that has resisted modularization should not show the effect — predictions a pure size story cannot make.
The framework makes three contributions. First, it supplies a mechanism linking the task-based theory of automation to the cross-sectional distribution of firm size, via two organizational channels — task modularization (the automation surface grows as work is decomposed into narrow standardized tasks) and organizational slack (large firms carry redundant overhead that AI can compress). Second, it reconciles an apparent contradiction in the evidence: firm-level studies find little aggregate employment loss while worker-level studies find sharp contractions for specific groups. Third, it relocates the policy object of concern from aggregate employment to the architecture of entry: the framework predicts erosion of the junior rungs of the career ladder even when total employment is stable.
Task-based automation. The framework builds on the task-based approach (Autor, Levy & Murnane 2003; Acemoglu & Autor 2011; Acemoglu & Restrepo 2018, 2019). This literature predicts heterogeneity across occupations as a function of routine-task intensity, but is silent on where within the firm-size distribution a given occupation's tasks are automated. The present paper extends it by making the automatable task share endogenous to organizational structure.
AI as augmentation and the skill-leveling effect. A generative-AI assistant raised customer-support productivity by roughly 14% on average, with gains near 34% for novice workers (Brynjolfsson, Li & Raymond 2023). This result supports the small-firm augmentation channel but simultaneously underwrites the hiring-suppression threat: if AI makes a novice perform like an expert, the small firm's reason to hire the novice at all weakens.
Worker-level evidence of contraction. Workers aged 22–25 in the most AI-exposed occupations experienced roughly a 13% relative employment decline since late 2022, while older workers in the same occupations were stable or growing (Brynjolfsson, Chandar & Chen 2025). The present paper proposes that firm size is a complementary axis: large specialized firms are precisely where junior rungs are most fragmented and therefore most automatable, so seniority and firm size should interact.
Firm-level adoption and reallocation. The most direct antecedent: larger firms and high-AI-investment firms are more likely to reduce their share of routine workers, while smaller firms are more likely to expand technical employment (Baslandze et al. 2026). This is close to a direct descriptive statement of the present thesis; the contribution here is to give it a mechanism and specify a design that identifies it causally.
Internal labor markets and ports of entry. The "entry ladder" the framework foregrounds is the internal-labor-market tradition of Doeringer and Piore (1971), in which firms fill higher rungs by promotion from a limited set of entry-level "ports" rather than from the external market. If AI thins those ports, the damage is not a current-period separation but a foregone future promotion pipeline — exactly the stock-versus-flow distinction this paper draws. The labor-economics literature on graduating in a downturn (Kahn 2010; Oreopoulos, von Wachter & Heisz 2012) supplies the long-lag "scarring" logic behind the dynamic implication: cohorts that never form leave a persistent hole in the experienced-worker supply.
Modularity and the division of labor. The claim that standardization precedes substitution has two intellectual parents. Braverman’s (1974) account of the decomposition of work into separable, codifiable steps describes precisely the process this paper labels modularization — here treated as a measurable moderator rather than a normative thesis. Baldwin and Clark (2000) formalize modularity as design rules with returns to scale, which is why \(m(f)\) is increasing in firm size without being caused by it. Williamson’s (1985) markets-and-hierarchies framing supplies the organizational-slack channel: the redundant coordination layers that scale carries, and that AI can compress. These are recombined, not invented; the contribution is to make the automatable task share a function of organizational architecture and to specify a design that recovers it.
Architecture of the contribution. Autor and Acemoglu explain tasks; Brynjolfsson and co-authors explain seniority; the Atlanta Fed study explains firms. No existing work connects the three. This paper supplies the connecting mechanism — modularization — and a single estimable design.
Consider an occupation \(o\) with routine-task share \(r_o\) in the abstract. AI lowers the cost of executing routine-codifiable tasks. The novel step: the realized automatable task share is firm-conditional:
where \(m(f)\) is the modularization intensity of firm \(f\) — how far it has decomposed occupation \(o\)'s work into narrow, standardized, separately-staffed tasks — and \(s_{o,f}\) is the task separability within firm \(f\): the degree to which tasks can be detached from the human context in which they are embedded. Both must be high for AI substitution to be likely. The two are distinct, not redundant: \(m(f)\) is an organizational property — whether the firm has already split the work into separately-staffed roles — while \(s_{o,f}\) is a task property — whether a given step, once isolated, can be executed without tacit, relational, or physically co-present context. A firm can modularize a role on paper (\(m\) high) while its tasks remain context-bound (\(s\) low), in which case AI still cannot substitute; conversely, separable tasks buried in an un-modularized generalist bundle (\(s\) high, \(m\) low) are not exposed because no one has carved them out. Substitution requires the product, not either factor alone.
Firm size enters \(g(\cdot)\) because decomposing work into specialized roles is a return to scale — but size is one argument among several, which is precisely why the design can separate the proxy (size) from the cause (modularization).
Same occupation, different task bundle
reconciliation · variance reporting · dashboard updating · SOP compliance
High codifiability · large automation surface
FP&A · collections · supplier negotiation · sales support · founder support
Generalist bundle · tacit-heavy · small automation surface
AI does not meet "the analyst" — it meets the bundle, and the bundle is set by modularization.
Six complementary approaches form the composite \(\hat{m}(f)\), combined via principal component analysis on the firm–year panel:
No single measure fully identifies modularization — each is a noisy signal. The claim is that the common factor across these measures captures the latent organizational architecture that is the true causal object. The composite \(\hat{m}(f)\) is a latent factor estimate, not a clean observation.
Because \(r_{o,f}\) is increasing in \(m(f)\) and \(m(f)\) is increasing in firm size, the automatable share of any exposed occupation is larger in more-modularized firms. Large firms have, in effect, pre-processed their white-collar work into an automatable form. Standardization is the precondition for substitution — the same logic by which assembly-line fragmentation made manufacturing tasks automatable a century earlier.
Large organizations carry redundancy — duplicated coordination, layered management, internal reporting, support and compliance functions — that is an equilibrium outcome of monitoring and coordination costs at scale. AI need not match the median worker's full productivity to justify headcount reduction; it need only compress coordination. A 5% reduction in a 50,000-person firm is 2,500 positions; in a 20-person firm, the same percentage is one load-bearing person. Small firms exhibit scarcity, not slack.
The size effect should be amplified where routine-task intensity is high, where the entry-level share is high, and where the firm over-hired in the 2021–22 expansion. In the long run the more consequential effect may be the non-formation of junior cohorts — a stock effect that flow data will reveal only with a lag.
Core
Conditional on AI exposure, more-modularized (typically larger) firms reduce employment in routine/clerical/support/junior-analyst/back-office roles more than less-modularized (typically smaller) firms.
The adoption-gradient objection — stated and answered. The most obvious alternative: "Large firms adopt AI earlier, so effects appear there first." The paper does not predict that large firms are affected because they adopt more. It predicts that, conditional on adoption intensity, the employment response differs because the same technology meets a different task architecture. The triple-difference in §6.1 holds adoption intensity constant; the H5 mediation run tests whether the residual size gap is absorbed by \(\hat{m}(f)\). If the gap collapses under adoption controls alone, the mechanism is adoption timing, not modularization — and the paper says so.
The dominant effect in less-modularized firms is not separations but a reduced rate of entry-level vacancy creation. The distinctive empirical object is the non-created job — invisible in layoff data.
Formal test — first-vacancy hazard. Let \(T_{f,k}\) be the time until firm \(f\) posts its first junior-specialist vacancy in function \(k\). The estimating equation:
Prediction: \(\beta < 0\). Placebo: non-exposed functions. Falsification: WARN separation data should be flat where postings fall.
AI lowers the routine/junior share and raises the technical/data/commercial share, leaving total employment little changed: the level is stable while the entry rung contracts.
Mechanism
The H1 effect is mediated by direct modularization measures (§3.1a). If \(\beta_1-\beta_2\) attenuates substantially once \(\hat{m}(f)\) enters, firm size was a proxy and the causal object is structure. This is the paper's central identifying claim, not a robustness check.
The displacement effect is larger where pre-period slack was larger (prior over-hiring, compressed margins, decelerating growth); here AI partly legitimates restructuring that slack would eventually have forced.
Output side (supportive, not load-bearing)
Conditional on adopting and using AI at comparable intensity, less-modularized firms show higher output per worker with smaller net employment reductions. H2 is supportive, not load-bearing: the paper's core rests on H1 and H3. Small-firm augmentation is directionally predicted but likely statistically invisible in current data (small firms adopt less; AI spend is opex not capex).
Each hypothesis has a stated null; for H1 the null is \(\beta_1 - \beta_2 = 0\).
What would prove the framework wrong. The framework should be rejected if: (1) exposed junior-role contraction is not stronger in more-modularized/larger firms; (2) the size effect disappears after controlling for adoption intensity and seniority; (3) less-modularized firms show separations rather than suppressed postings; (4) direct modularization measures do not mediate the size effect. Each is a clean kill.
The natural unit is the firm–occupation–time panel, supplemented by firm-level financials.
Outcomes. Total firm headcount; headcount by occupation and seniority; job postings by firm × occupation × seniority; layoffs; the entry-level share of postings; and — as the structural profit-side signal — the labor share (payroll ÷ value added) and value added per worker.
Treatment / exposure. Firm-level AI adoption; AI-skill share of postings; AI mentions in earnings calls; occupational AI-exposure indices; sector AI use from federal surveys.
Candidate sources. Lightcast / Burning Glass (postings by firm, occupation, seniority, skill); Census BTOS (AI adoption by size and sector); Revelio Labs (function-level headcount); Compustat / Capital IQ (public-firm productivity); WARN notices (formal layoffs — placebo for H3); earnings-call transcripts (adoption timing); Orbis / EIBIS (European cross-section); Census RDC (full-distribution productivity/labor-share signal; see §6.5); BLS QCEW/JOLTS; ADP payroll micro-data (via partnership).
Minimum viable empirical stack. (1) Postings (Lightcast/Revelio) for H3/H4; (2) headcount and flows (Revelio/LEHD/ADP) for employment by seniority and firm size; (3) adoption/treatment (BTOS, ABS AI supplement, earnings-call timing); (4) output (Census RDC for the full distribution). Layers 1 and 3 alone permit a first pass at H1/H3/H4.
Measurement caveats. Postings ≠ hires; AI adoption is self-reported and lumpy; public-firm financials over-represent large firms; enterprise firm size ≠ establishment size; the AI-exposure index is contested.
The object of interest is \(\beta_1 - \beta_2\) — the differential employment response to AI exposure between large and small firms. H1 predicts \(\beta_1 - \beta_2 < 0\) for routine/junior roles.
Three specification points. First, size is not binary: estimation uses size bins (1–19, 20–49, 50–249, 250–499, 500+), a continuous log-employment interaction, and a spline. Second, modularization replaces size in the mediation runs (H5). Third, exposure is split by type:
The decisive test, stated plainly. The informative comparison is large vs. small holding constant occupational exposure, AI adoption intensity, sector, pre-period growth, and seniority mix. Only the residual size gap that survives all of these supports H1/H5.
Pre-trend coefficients \(\theta_{k<0}\) are the key falsification check. The expected pattern is divergence after \(k=0\): junior/routine headcount in large firms turning down, output per worker in small firms turning up, and entry-level postings in small firms flattening.
Candidate instruments: (i) shift-share (Bartik) exposure — pre-period occupational mix × national leave-one-out AI-exposure growth; (ii) technological supply shifters — model-release timing interacted with pre-period task mix; (iii) geographic compute/cloud-infrastructure rollout interacted with exposure. All treated as candidate instruments with contestable exclusion restrictions; IV complements rather than substitutes for the event study and triple-difference.
The FSRDC is the only U.S. infrastructure observing the full firm-size distribution. Linkage: LBD (employment, payroll, firm size, age) + ABS AI supplement and BTOS (treatment) + Economic Census/ABS receipts (denominators) + LEHD (worker × firm × age for entry-level/seniority outcomes).
\(Y\) taken as (i) labor share (payroll ÷ value added); (ii) value added per worker; (iii) log receipts per worker. Prediction: \(\beta_1 < 0\) and \(\beta_2 \approx 0\). Census measures receipts and payroll, not net income — hence labor share, not profit margin, is the deliverable. FSRDC access requires an approved project, Special Sworn Status, and disclosure review. This is a 12–18-month build, run in parallel with the faster postings-based tests of H1/H3.
Said once, firmly: this section is motivating, not confirmatory — none of it is a clean test of \(\beta_1-\beta_2\).
| Hypothesis | Observable implication | Dataset | Outcome | Main threat | Falsification |
|---|---|---|---|---|---|
| H1 | Exposed junior roles fall more in more-modularized firms | Lightcast × Revelio/ADP; BTOS | Junior postings; headcount by seniority | Adoption gradient | Effect vanishes once adoption + seniority controlled |
| H2 | Output per worker rises in small adopters | Census RDC; Compustat | Labor share; receipts per worker | Small firms rarely adopt; opex not capitalized | No productivity response among adopters |
| H3 | First junior vacancy hazard lower post-ChatGPT in exposed small firms | Lightcast full text; Revelio | First-vacancy hazard \(\lambda_{f,k,t}\) | Demand decline in exposed functions | Separations rise; or effect in non-exposed functions |
| H4 | Routine share ↓, technical share ↑, total stable | JOLTS; BTOS; BLS CES; Liu & Webber 2026 | Routine share; aggregate payrolls | Macro confounders | Total employment collapses, or composition unchanged |
| H5 | \(\beta_1-\beta_2\) attenuates when \(\hat{m}(f)\) enters | Lightcast title diversity; process-language index | Triple-diff with \(\hat{m}(f)\) | \(\hat{m}(f)\) correlated with size | Size gap remains after modularization controls |
| H6 | Displacement larger in over-hired, margin-compressed firms | Compustat pre-period trends | Triple-diff × overhiring indicator | Selection into over-hiring | Slack-free firms show same effect |
Adoption rises steeply with size — and the gap is widening, not closing. Higher-frequency firm-size-resolved data (Ramp AI Index via Bloomberg, monthly, Jan 2023–May 2026) put magnitudes on the confound. Adoption among large businesses rose from 7.8% to 61.3%; among small businesses, from 3.7% to 44.4%. Fitted monthly trends are steepest at the top (\(\beta_{\text{large}}=1.27\) vs. \(\beta_{\text{small}}=1.14\) pp/month; Appendix Table A1b), and the large–small level gap did not narrow as diffusion matured — it widened, from 4.1 to 16.9 percentage points (Appendix Table A3). The firm-weighted Census BTOS national rate sits near 19.8% over the same window, reflecting a different surveyed universe; the two agree on the monotone size ordering and the trend, not on levels. The Ramp figures should be read as a diffusion index within a specific business sample (card-using firms), not as a population adoption rate — 61.3% is not the share of large firms in the economy using AI. This is the central confound, not just support: any AI labor signature appears in large firms first because they adopt first, so the design must isolate the gap that survives equal-adoption comparison.
Investment is highly concentrated. AI spending runs ~$2,000 per employee economy-wide, with professional/business services near ~$3,470 per employee (Atlanta Fed study, 2026). The frontier-firm concentration is where the modularization-and-slack channels should be most active.
Stated reshuffling in the predicted direction. Larger and high-investment firms are likelier to cut routine workers; smaller firms are more likely to expand technical employment (Atlanta Fed study, 2026). This is nearly a descriptive statement of H1/H4/H5 — but from stated executive intentions, not realized causal effect.
No aggregate separation shock. The JOLTS layoff-and-discharge rate held in a ~0.9–1.2% band throughout the AI era — below the 2019 expansion baseline. Pre-ChatGPT mean = 1.11%; post-ChatGPT mean = 1.04%. Supports H4. See Appendix Table A5.
The white-collar signature is an openings collapse, not a separations spike. JOLTS gross flows pulled by industry show the asymmetry H3 and H4 predict: in white-collar segments, vacancy creation has fallen sharply while separations stayed near their historical band. From the 2021H2–2022 average to the 2024H2–2026 average, professional & business-services openings fell ~42% while layoffs and discharges rose only ~22% off a far smaller base; the pattern repeats in information (−51% vs. +23%) and finance (−24% vs. +29%). The vacancy-to-separation ratio in professional services collapsed from ~5.9 to ~2.8. A level-only structural break confirms a significant post-ChatGPT downshift in openings (≈−18%, \(p<0.001\)); the break is not significant once a linear trend is included, because vacancies were already declining — so the claim rests on the level comparison and the ratio collapse, not the two-parameter break. This is the macro shadow of hiring suppression: a separations-only monitor (WARN, the JOLTS layoff rate) registers almost nothing, while the vacancy side has more than halved. This evidence does not identify AI; it identifies the margin on which a hiring-suppression mechanism would appear. See Appendix §A5.3, Tables A5c–A5d, and Figure A7.
The net-flow data agree, and locate the offset. ADP sectoral net employment change (monthly, by industry) cross-validates the vacancy picture and shows where the aggregate is held up. Over the trailing twelve months (May 2025–Apr 2026), net private hiring of roughly +47k/month is carried almost entirely by one less-exposed, relational sector: education & health added ~+52k/month (+624k cumulative, more than 100% of the +567k total private gain), while the three AI-exposed white-collar service sectors — professional services, information, and financial activities — together netted approximately zero per month (professional services −54k cumulative; information +13k; finance +26k). Net change equals gross hires minus separations; with separations flat (JOLTS, A5.1), a collapse toward zero net hiring in exposed white-collar services implies a fall in gross hires — the hiring-suppression signature in a flow measure. This is the §10 aggregation logic made visible: a stable aggregate resting on a relational sector while the exposed white-collar middle stops being a net employment engine. See Appendix §A5.4, Table A5f, and Figure A8.
Caveat, stated plainly. Professional-services net hiring was already weak across the entire post-2022 ADP window, not only recently, so this is a persistent compositional feature consistent with the framework — not a clean AI-dated break. ADP revised its methodology in 2022 (comparisons are kept within the post-2022 regime), professional services bundles rate-sensitive temporary-help employment, and none of this isolates an AI channel. Motivating, not identifying.
Two independent series triangulate the same picture. First, the size–adoption gradient reappears in transcript text, not just spend: normalizing earnings-call mentions of "AI" near "adoption" by constituent count, large-cap S&P 500 firms discuss AI adoption roughly three times as often as small-cap Russell 2000 firms (≈0.27 vs. ≈0.09 mentions per firm in Q2 2026), an independent corroboration of the Ramp gradient and of the earnings-call component of \(\hat{m}(f)\) (§3.1a). Second, the H3 placebo is consistent with the data: national WARN layoff notices averaged ~379k workers/year in the AI era (2023–25) versus ~304k in 2017–19 — a modest rise within reporting-coverage drift and an order of magnitude below the 2020 mass-separation event (2.31M). Openings in exposed white-collar sectors more than halved and net hiring there fell to zero, yet formal layoffs show no spike — the door that does not open, not the worker pushed out. See Appendix §A5.5, Tables A5g–A5h, and Figures A9–A10.
A postings-side cut by seniority moves closer to the mechanism. Disaggregating postings by seniority shows the predicted divergence directly: indexed to August 2024, junior postings remained depressed (trough 88, recovering only to 93) while senior postings recovered above baseline (104), opening an 11-point senior–junior gap by August 2025. On the worker side, recent-graduate unemployment has inverted relative to its own history — from roughly 1 point below the all-worker rate through the 2010s to about 1.2 points above it in 2024–25 (5.6% vs. 4.2%, March 2026). This is not an AI-identified result — the sector cross-section cannot separate AI from the remote-work unwind (AI and remote intensity are collinear at \(r=0.90\)), and the seniority series is economy-wide, not resolved by AI exposure — but it is the closest public aggregate analogue to H3's non-created junior job. Identifying AI as the cause still requires the firm×function×seniority panel of §6.5. See Appendix §A5.6, Tables A5.6a–A5.6b, and Figures A11–A13.
A within-role test does not support a uniform entry-port effect. The closest cut available — Revelio postings by role×seniority×month, across thirteen roles (six AI-exposed, five low-AI controls) — tests whether entry postings fall more than experienced postings inside the same role. Pooled across the six AI-exposed roles, the entry-port effect is absent (coefficient +1.7%, \(p=0.86\)); entry and experienced postings fell together. A triple-difference against the low-AI controls produces a large negative coefficient, but that is a fixed-effects artifact — an equal-weighted comparison of role-level effects shows no AI-vs-control difference (\(p=0.95\)) — so we report it and decline to interpret it. The honest result is heterogeneity: three of six AI-exposed roles lean the predicted way — software engineering (\(-19\%\), \(p=0.04\)), routine legal drafting (\(-15\%\)), and IT operations (\(-13\%\)) — while accounting, administrative support, and systems engineering hold up or expand. This does not confirm H3, and it positively contradicts a uniform "white-collar entry collapse." If anything, it locates where the mechanism may bind first — the roles whose entry-level tasks are both highly modular and directly automatable today — which is consistent with the \(m(f)\times s_{o,f}\) structure but is a hypothesis from a nine-role sample, not a result. Critically, these data carry no \(firm\_id\), so the firm-size/modularization channel that H3 actually concerns cannot be tested here; that still requires the firm-level hazard of §6.5. Three successive role expansions (5, 9, 13) return the same null, indicating that adding roles has reached diminishing returns and that the firm dimension, not more roles, is the binding next step. See Appendix §A5.7, Table A5.7a, and Figures A14–A15.
The limit of public evidence. Taken together, the descriptive layers in this section have reached the edge of what public and aggregate data can establish. They show: a steep, widening firm-size gradient in AI adoption (the confound); a broad openings slowdown with stable separations; net hiring narrowing to one relational sector; an economy-wide junior–senior posting divergence and recent-graduate inversion; and within-role entry-port weakening concentrated in the most codifiable roles. What they cannot show — because no public series carries firm identity linked to occupation and seniority — is the object of the title: whether large, AI-adopting firms suppress junior roles more than small firms in the same function. That single interaction is H3 proper, and it requires firm-level microdata (§6.5). Further aggregate series would not change this; the binding constraint is now data structure, not data quantity. Appendix §A5.9 reports a sequence of deliberately simple econometric tests — scatterplots, bivariate OLS, controlled OLS, and role×seniority fixed-effects panels with placebo comparisons — to discipline these descriptive patterns before the firm-level design: the adoption gradient and the junior–senior scissors are strong and significant, while the within-role, AI-resolved entry-port effect is not uniform and concentrates in the most automatable roles.
AI does not break every entry ladder; it breaks first where entry work has already been modularized into automatable tasks. The public evidence assembled here can be read at a glance:
| Test | Result | Interpretation |
|---|---|---|
| AI adoption by firm size | strong positive (+0.57 log/rank, R²=0.92) | adoption-gradient confound — must be neutralized, not the mechanism |
| Junior/senior scissors (economy-wide) | significant negative (−9.2%, p<0.001) | consistent with entry-ladder compression |
| Revelio pooled within-AI H3 | null (+1.7%, p=0.86) | no uniform white-collar entry collapse |
| Per-role H3 (Revelio) | software / IT / routine-legal negative | supports the automatable-entry claim, not a general one |
| True H3 (firm-size × modularization) | pending | requires firm-level data (firm × role × seniority × month) |
The sharpest within-occupation contraction is by seniority. ~13% relative employment decline for 22–25-year-olds in the most AI-exposed occupations, older workers stable (Brynjolfsson, Chandar & Chen 2025). The framework's bet is that seniority and modularization interact; a size split of these data is the natural next test.
Adoption-gradient confound (most serious). Large firms adopt more, so a size gradient in outcomes can arise mechanically from differential treatment intensity. Two independent measures make this concrete and quantify how large the confound is. Ramp spend data show a large–small adoption gap of 16.9 pp, widening over diffusion (§7; Appendix A3); Indeed AI-posting data show an even steeper gradient in AI hiring — top-1% firms post AI roles at ~50% versus ~1% for the smallest, a ~38× ratio, also widening (Appendix §A5.8, Figure A16). Either is large enough to generate a size gradient in employment outcomes with no role for task architecture. Mitigation: condition on adoption intensity; use the continuous-exposure specification; test whether the size effect survives within bands of equal adoption intensity. If it does not, the result is an adoption story, not a modularization story — which is exactly why the H5 mediation by \(\hat{m}(f)\) is load-bearing rather than a robustness check.
Modularization vs. size (H5). Size may merely proxy for modularization. If \(\beta_1-\beta_2\) attenuates when \(\hat{m}(f)\) enters, the policy object is organizational design, not size, and that attenuation is the finding, not a failure.
Reverse causality / restructuring narrative (H6). Firms already planning restructuring may adopt AI to legitimate it. Mitigation: event study with pre-trend tests; IV using supply-side technology timing; explicit interaction with pre-period over-hiring and margin pressure.
Macro and sectoral confounders. The post-2022 window coincides with rate hikes, the tech-sector hiring slowdown, post-pandemic normalization, and remote-work reversal. Mitigation: industry-by-time fixed effects; firm-level rate-sensitivity controls; excluding tech firms and remote-amenable occupations. For each major alternative, the question is whether it can generate a modularization-conditional, exposure-conditional, junior-concentrated pattern with flat pre-trends. Most cannot generate the full conjunction.
Measurement. Postings ≠ hires; self-reported adoption is noisy; public-firm financials select toward large firms; enterprise firm size ≠ establishment size.
If the framework is right, the policy-relevant story is a compositional story with a dynamic sting. The economy-wide payroll can keep growing while the architecture of entry erodes: large firms compress the junior and routine layers that have historically been the training ground for mid-career talent, and small firms quietly decline to make the first junior hire because a subscription substitutes for it. Neither move registers as a layoff. Both register as a missing rung.
The distinctive empirical object is the non-created job: a vacancy that, absent AI, would have existed, and now does not. It is invisible in separations and WARN data and only partly visible in postings; H3 is designed precisely to surface it. The distinctive dynamic implication is a stock problem masquerading as a benign flow: if junior cohorts do not form, the future supply of experienced workers — the very workers whose tacit knowledge AI currently complements rather than replaces — is drawn down with a long lag.
Aggregate-employment monitoring will under-detect the phenomenon; the relevant indicators are entry-level posting shares, junior-to-senior hiring ratios, and modularization-conditional occupational composition. Interventions aimed at "saving jobs" miss the target; interventions aimed at rebuilding the entry rung — apprenticeship subsidies, training-linked hiring credits, portable credentialing — are closer to the mechanism.
The aggregate signature of AI is a size-weighted average of two opposing micro-responses. The weights are knowable and asymmetric: firms with 500+ employees are ~0.1% of U.S. businesses but employ ~54% of workers and pay ~52% of the wage bill (SUSB 2022; BEA WP2022-12). Three implications follow; the arithmetic is in Appendix Table A6.
Implication 1 — Productivity gains load on the large-firm within-firm component. The automation margin raises value added per worker in the large-firm segment, which carries the majority of value added, adoption intensity, and measured productivity gains per adopter. Small-firm augmentation is real but largely invisible in national accounts (opex-financed, capability-expanding rather than turnover-raising). The testable formalization is a size-weighted Olley–Pakes decomposition applied to the §6.5 RDC estimates; the illustrative magnitude is in the appendix and should be treated as accounting motivation, not a causal claim.
Implication 2 — Employment effects appear as compositional churn, not aggregate collapse. Neither segment pushes total employment sharply. The result is flat aggregate payrolls atop a sharp compositional shift — the Liu–Webber "no aggregate decline, but pockets" pattern, now with a structural mechanism for why the pockets cancel in the total.
Implication 3 — The right dashboard is not headline payrolls. The relevant leading indicators are: the large-firm labor share (§6.5); the economy-wide junior-to-senior posting ratio; and the first-vacancy hazard in exposed functions (H3). Aggregate-employment statistics show the last thing that moves.
Limits. This is an illustrative accounting bridge, not a general-equilibrium model. It omits the reinstatement effect, price adjustment, entry/exit dynamics, and long-run wage/capital responses.
The AI-and-work debate has been conducted at the wrong unit of analysis. By fixing on the occupation, it forces a binary — destruction or productivity — that the data refuse to settle, because the same occupation carries opposite labor signatures depending on the organization that houses it. This paper has argued that the missing moderator is task modularization — how far an organization has already turned human work into standardized, codifiable process — and that firm size matters only because it predicts modularization. Highly modularized firms (typically large) sit on the automation margin; bundled, generalist firms (typically small) sit on the augmentation margin.
The framework reconciles the field's central tension — stable aggregates alongside sharp localized contractions — as one compositional phenomenon. The empirical agenda is explicit and self-falsifying: a firm–occupation–time triple-difference whose estimand is \(\beta_1 - \beta_2\), estimated across size bins and a continuous size spline; mediation runs replacing size with direct modularization measures; an automate-versus-augment exposure split; event study; and placebo and falsification tests built around the distinction between separation and suppression.
If AI is breaking the labor market, it is not breaking it evenly. It is breaking the parts of the firm where work had already been made machine-readable — it breaks the entry ladder first where entry work has already been modularized — and it is breaking the career ladder before it breaks the payroll aggregate. A note of empirical discipline belongs here: the pooled entry-port collapse is not in the public data; the automatable-entry-port collapse is. A uniform "white-collar entry is collapsing" reading is not supported and is not the claim. What the public evidence supports is narrower and sharper — concentrated entry-port weakening where tasks are codifiable, against a backdrop of steeply size-concentrated AI adoption — and the step that would convert this from motivating pattern to identified effect is firm-level, not aggregate. The risk is not mass unemployment. It is a labor market that keeps its senior workers while quietly halting the production of the next generation of them.
Author to complete pagination and finalize working-paper numbers before circulation.
Data sources.