The Mid-Market's AI Lead Is an Optical Illusion

Ask which companies are buying the artificial-intelligence boom and the intuitive answer is the giants — the firms with the budgets, the data, and the platform teams. A network of large-employer job postings lets us ask the question a different way: not who says they are investing in AI, but who is actually writing AI skills into the jobs they post. The answer, taken at face value, is a surprise. It is the middle of the market — mid-sized employers, not the largest — that writes AI into the highest share of its openings. Taken apart, that surprise dissolves into something more useful, and more defensible.

What this corpus can see

The scope governs every reading that follows. This is a record of direct-employer postings from large U.S. employers — 93–99% of it flows from DirectEmployers, an applicant-tracking distribution network — not a survey of the labor market. The unit counted is the vacancy-location: one opening at one place, deduplicated against the same opening re-posted or re-scraped (methodology §1). Only April through June 2026 carries a claim; February is 87% one ingest day and March is a burn-in backfill, so both are greyed. And "AI" here is a concrete definition: a posting is flagged AI-skill when a specific AI tool or competency appears in the ad — machine learning, LLMs, PyTorch, a named model — rather than a bare, boilerplate mention of "AI." The flag is stricter than any-mention, but it is honest to say it fires on the concrete term wherever it appears in the ad, including a firm's standing company-description boilerplate; because that boilerplate is templated across all of an employer's postings, some of the size-band and within-occupation differences below could reflect which employers append AI language rather than which jobs require it — a confound this report flags but does not fully isolate. That tier runs about 8% of the corpus; the narrower AI-core tier (the job itself is an AI role) runs under 1%.

Company size is the enrichment that makes this report possible, and its softest joint. The corpus carries no size field. We classified the top 5,000 employers by posting volume (4,998 distinct after de-duplication) — which together account for 98.6% of size-classifiable, SOC-covered postings — into large (>10,000 employees), medium (500–10,000), and small (<500), from public knowledge, with a confidence tag on each. This is a public-knowledge estimate, not registry data. The Limitations section states what that costs.

The face-value finding

On the primary window, staffing and agency firms excluded, the AI-skill share of new postings rises with size to a peak in the middle:

— AI-skill, AI-core and VL columns. The named-tool column is not a cube field (it requires the per-posting genai_tools array, which lives only in the enrichment layer); it resolves to the supplementary exhibit r5_genai_tool_by_band.csv, recomputed on the same SOC-covered, staffing-excluded base — before the single-employer cell suppression the cube now applies to the AI-skill/AI-core/VL columns, so its denominator is a few percent larger than theirs.

Mid-sized employers lead: 10.0% of their openings ask for an AI skill, against 8.1% at the largest firms and 3.9% at the smallest. The ordering — medium, then large, then small — holds when staffing firms are included (9.1% / 7.6% / 4.5%), and holds again on the narrower named-tool measure, where mid-sized firms name a specific model (Copilot, Claude, ChatGPT, Gemini) in 1.9% of postings against the large band's 1.4%. It is a clean, reproducible ordering. It is also, on the medium-over-large half, largely an illusion.

The mid-market lead is an occupation-mix artifact

A share of AI-skill postings is two things multiplied: which occupations an employer hires for, and how often it asks for AI within each. Mid-sized employers in this corpus tilt toward management, computing, and engineering roles — occupations that ask for AI skills often. The largest employers carry enormous volumes of retail, pharmacy, healthcare, and sales postings — Walmart, CVS, and Walgreens sit at the top of the large band — that almost never do. That difference in what is being hired for, not how AI-hungry each employer is, produces most of the gap.

The decomposition is stark. Of the 1.86-point lead mid-sized firms hold over large ones, 1.79 points is occupation mix and just 0.07 points is the within-occupation rate. Standardize the two bands to a common occupation mix and the lead all but vanishes: mid-sized firms land at 8.18%, the largest at 8.11% — a statistical tie. Look inside the most AI-intensive fields and the giants, if anything, lead: in computer-and-mathematical occupations the large band asks for AI skills in 33.5% of postings against mid-sized firms' 31.5%; in business and financial roles, 11.1% against 7.7%. The largest employers are at least as AI-hungry as the middle of the market once occupation is held constant.

Two limits on that "held constant" are worth stating plainly, because they bound how far the tie can be pushed. First, the occupation control is coarse: it standardizes to the 23 SOC major groups, not to detailed six-digit roles or seniority, so a residual composition or seniority-fill difference inside a group — large and mid-sized firms need not hold the same mix of roles within "Computer & Math" — cannot be excluded. Second, every occupation figure rests on a title→SOC crosswalk whose measured precision is 68.4% at the six-digit grain and about 72% at the minor-group grain (methodology §12); the sub-two-point computing gap (33.5% vs 31.5%) sits inside that measurement error and should be read as "the giants at least tie," not as a firm lead. The robust claim is the tie itself, not its sign.

The month-by-month view says the same thing from another angle. The mid-market's lead is front-loaded to April (10.3% versus the large band's 6.5%) and erodes as the window runs. Large-employer AI-skill demand rises across the three clean months — April 6.5%, May 8.0%, June 9.5% — and that rise is robust: it survives the mandatory drop-one-ingest-day jackknife (methodology §13), whose April, May and June bands do not overlap (the monthly series here is the un-suppressed panel series the jackknife runs on, not the cell-suppressed cube aggregate — see the provenance note). The mid-market's monthly demand, by contrast, is flat and noisy. By June the two bands converge — large 9.5%, the mid-market 9.1% — but we stop short of calling that a crossing: the June ordering does not survive the jackknife. Dropping a single high-volume ingest day moves the large band across a [9.0%, 9.8%] range and the mid-market across [9.1%, 9.5%], and on some drops the mid-market ends up ahead. Convergence is the honest reading; a June "large passes medium" would not be publishable as a directional fact. Whatever "mid-sized firms lead" means, it is not a stable or a widening one.

One caveat cuts the other way, and it is worth stating because it complicates the giants' case. The large band's AI-skill demand is concentrated: remove its five largest posters and its share falls from 8.1% to 7.2%, with two firms — Oracle and Deloitte — carrying much of the AI weight. The mid-market's demand is broad-based; removing its top five raises its share, to 11.1%. So the honest composite is this: at the aggregate the mid-market's edge is an occupation-mix mirage, but the large band's AI intensity leans on a handful of technology and consulting giants rather than being spread across the Fortune 500.

The gap that survives everything

Strip the illusions away and one division remains, robust to every control: the smallest employers demand AI skills far less often than everyone else, and not because of what they hire for. Standardized to the large band's occupation mix, small firms land at 4.12% against the large band's 8.11% — a 4.0-point gap that is not composition. They trail within every major occupation measured: in computing, 14.6% against the giants' 33.5%; in management, 8.0% against 13.1%. On the named-tool measure they name a specific model in 0.5% of postings against the large band's 1.4% — a gap of roughly two-and-a-half-fold, wide but far short of the AI-skill gap.

Two scoping caveats keep this from being over-read. "Small" here does not mean small firms in general: it means sub-500-employee employers that post through DirectEmployers, a large-employer, compliance-driven distribution network — a non-representative slice of the small-firm universe. And the small band is the thinnest and least-measured part of the corpus: its postings carry a mappable occupation only 65% of the time, against about 75% for the larger bands, so it conditions on a narrower, differently-selected base, and its size classifications are unaudited. The honest sentence is scoped: if the AI boom is reaching sub-500-employee firms inside this network through their hiring, this record does not yet see it. That — not the mid-market headline — is the finding that would survive a hostile read.

What this does and does not establish

It establishes that, among direct-employer postings from large U.S. employers, AI-skill demand is a big-company and mid-market phenomenon in roughly equal measure once occupation is controlled, and is markedly thinner at the smallest employers. Every size comparison here is defined inside this one ATS network — "large," "mid-market" and "small" are bands of employers that post through DirectEmployers, not a census of firms — and rests on an unaudited agent size estimate and a coarse, major-group occupation control. So the finding must not be restated as a claim about which real-world companies are buying AI. It does not establish anything about AI spending, headcount, or investment in dollars — only about the skills employers write into public job ads. It does not speak to firms outside this network, to the two-thirds of small businesses that rarely post through an ATS at all, or to any month before April 2026. Posting behavior is a leading signal of intent, not a ledger of outlays.

Limitations

• Size is an estimate, not a registry. Company size was assigned by an AI agent from public knowledge for the top 5,000 employers by posting volume, banded large/medium/small, with a confidence tag. After the de-duplication below, 4,998 distinct employers remain: 1,353 are large, 1,819 medium, 1,791 small, and 35 unknown; by self-rated confidence, 1,246 are high, 3,040 medium, 712 low. A 300-employer human-scored audit is specified but not yet complete — until it is, band-level error is characterized only by the model's own confidence mix. Findings are made at band level, never for individual employers' sizes, precisely because any one classification may be wrong.
• Coverage, differential by band, and the unclassified tail. The classified employers cover 98.6% of size-classifiable, SOC-covered postings; the long tail (1.2% unclassified, 0.15% unknown) is reported, not silently dropped. Every share here is computed on the SOC-covered base — about 74% of in-window vacancy-locations (the widely-quoted 71% is the posting-level coverage; the vacancy-location grain this analysis runs on is a few points higher). That coverage is not uniform across bands: small-firm postings carry a mappable occupation only 65% of the time against roughly 75% for large and mid-sized firms, so the small band conditions on a narrower, differently-selected slice than the others. Postings without a mappable occupation are outside this analysis entirely.
• Two enrichment defects in the size file, both now corrected. (1) Two employers (Ecolab, Midwest Machinery Co.) appeared twice under one company ID each in the pre-refresh file; both duplicate rows agreed on band, so the effect was a fractional over-count (well under 0.01 point on any headline share). This has been de-duplicated in the current file (5,000 → 4,998 rows), and every number in this article is computed post-dedup. (2) A separate, systematic off-by-one misalignment in enrichment batch s04 — a block of rows that carried each other's size_band/confidence/note shifted by one, most visibly Jack Henry & Associates (a public fintech) labeled small with a note describing a different company — has now been realigned: Jack Henry and the other affected rows carry a coherent band, confidence and note in the current file. The affected employers were all low-volume and posted no AI-skill jobs, so the correction does not move the 9.97 / 8.11 / 3.87 headline shares; it is recorded here for provenance.
• Occupation coding is a crosswalk with measured error. Every within-occupation rate and the whole occupation-mix decomposition depend on a title→SOC crosswalk, whose measured precision is 68.4% strict at the six-digit grain and about 72% at the minor-group grain (methodology §12). The standardization controls to the 23 SOC major groups only — not to detailed occupation or seniority — so residual within-group composition is possible, and small within-occupation gaps (notably the ~2-point computing gap) carry unquantified crosswalk error.
• Three clean months, and the mandatory jackknife. April–June 2026 is three data points. Per methodology §13 every published monthly share ships with a drop-one-ingest-day jackknife band; the large band's monotonic April→June rise survives it (non-overlapping bands), but the June large-vs-mid-market ordering does not (its bands overlap and can flip), so the article reports June convergence rather than a crossing. Nothing here is extrapolated forward.
• Suppression, now enforced at both grains. Every cited cell clears the ≥30-vacancy-location floor by a wide margin (the thinnest cited in prose are the small band's computing and management cells, n≈7,700 and n≈7,100). The size cube now also enforces the single-employer half of STANDARDS §7, which an earlier draft could not: cube_size_band_monthly carries a single_emp_share column, and any (month, band, occupation) cell whose top employer exceeds 50% of the band's non-staffing volume is dropped at build time — 121 cells, 2.9% of pre-suppression size-band volume. The most material drops are the medium band's Transportation cell (one carrier, Roehl Transport, is 61–73% of it across April–June) and the large band's Protective-Service cell (Allied Universal, 51–53% in April and May). Because these single-employer-dominated cells are overwhelmingly low-AI, removing them raises the headline shares against the earlier un-suppressed cut — medium 9.19%→9.97%, large 7.97%→8.11%, small 3.81%→3.87% — and every number in this article is post-suppression. On top of that, the full decomposition table that ships with this article suppresses its two remaining sub-30 cells (Military: medium n=11, small n=10).
• AI is a regex. The AI-skill flag is a curated, reviewable definition with documented false-positive guards; because it fires on a concrete AI term anywhere in the ad, including templated company boilerplate, part of a band's rate may reflect employer-level ad templating rather than per-job requirements (see above). The AI-core and any-mention tiers are reported alongside so a reader can see the whole distribution rather than one number.
• Single-source scope. 93–99% of the corpus is one ATS network of large U.S. employers. Every figure is a claim about that record, not about "the economy," and the smallest-employer band in particular is the thinnest slice of an already large-employer-tilted corpus.

Every headline, decomposition, within-occupation and standardization number above resolves to cube_size_band_monthly at corpus snapshot 2026-07-12 (now with single-employer cell suppression baked in); the concentration checks resolve to the employer-level cube_employers. Two families of figure fall outside that cube by construction and resolve to supplementary exhibits instead: the named-tool shares (which need the per-posting genai_tools array) to r5_genai_tool_by_band.csv, and the monthly trend with its drop-one-ingest-day jackknife bands (which need the raw ingest-day field, and run on the un-suppressed panel series so the jackknife is meaningful — its monthly values therefore differ from the suppressed cube's by-month cells) to r5_jackknife_by_month_band.csv. Reproduction inputs and the full by-month, ex-top-five, jackknife, and decomposition tables ship with this article.