The Excelsior Group · The de-siloed AI × (education · work · talent acquisition · L&D · HR) letter

The Read

For three editions this letter has tracked the entry rung breaking. This week the most rigorous evidence yet arrived, and it complicates the rescue now underway. David Autor and colleagues ran a three-month randomized trial across 133 patent lawyers at 11 firms: AI raised work quality by roughly a third of a standard deviation, and junior lawyers gained the most. Then the researchers took the AI away. The advantage that survived belonged almost entirely to the seniors (+0.45 SD), while juniors split into more good performers and more poor ones — the largest gains, as the authors put it, accrued to the lawyers who retained the least. Set that against Philippa Hardman's count of live L&D postings, where 1% of 379 roles were junior, and against a Bocconi trial where ChatGPT lifted rubric scores 0.86 points while critical-thinking training lowered them, and the week's shape is clear: we are automating the practice that built expertise, our assessments still reward exactly what the machine now does for free, and roughly $1B of corporate money has started rebuilding the ladder in the trades rather than in the professions. Old problems, new physics — and this week, for the first time, a control group.

Tide status

No tide movement this week — the default, and a feature. But T1 (the entry rung is breaking) gets its sharpest refinement since seeding. The Autor patent-drafting RCT is the first controlled evidence that the entry-rung problem is not only about jobs disappearing but about whether the humans still in those jobs are accumulating anything. It cuts against both comfortable stories: AI did not deskill everyone (treated lawyers beat controls unaided at 90 days), and it did not level the field (the retained gain concentrated in seniors). Stanford DEL published nothing this window — the next indicators release is September 23, and the 19% gap stands unrevised. T2 (instruction cost to ~zero) logged its first serious counter-current: New York City barred student-facing generative AI through eighth grade for ~600,000 students, the largest such moratorium in the country. Free and default-on is running into a governance wall, which changes the diffusion curve without changing the cost curve. T3 (the credential's monopoly) — no movement; the OpenAI Jobs Platform remains unlaunched roughly twelve months past its announced target.

Waves

W5 — The engineered apprenticeship gets its first transfer evidence — and it is uncomfortable [research]

Last edition logged the corporate build-out of simulated practice (KPMG's TaxSIM, Deloitte's rebuilt audit scheme, the six-firm DepoSim pilot) and named the open wedge: every number was a satisfaction number, and no vendor had published transfer evidence. Academia published it first. NBER w35720 (Autor, Rodchenko, Martin, Iscenko, Strand, Pearl, Ferere) randomized AI access across 133 patent lawyers at 11 US IP firms for three months: quality rose +0.34 SD at 10 days and +0.38 SD at 90 days, with juniors gaining most. The decisive test was the unaided post-test — treated lawyers still beat controls by +0.32 SD, so no net skill erosion on average. But the retained advantage sat with senior lawyers (+0.45 SD), while juniors bifurcated into more good and more poor outcomes. The authors' reading: foundational expertise may be necessary to convert AI assistance into durable capability. Meanwhile Hardman's field count found the rung that builds that foundation has largely gone — 1% of 379 live L&D postings were junior, a 28:1 senior-to-junior title ratio (edtech) (L&D).

Roadmap implication: Every simulation vendor now has a benchmark to clear and a harder question to answer: if durable gains require prior expertise, a simulator sold as a substitute for the missing entry rung may not work on the people who need it most. Buyers should ask for unaided post-tests at 90 days, segmented by seniority — the Autor design is the template, and 'satisfaction' should stop clearing procurement.

Links: NBER w35720 — Does AI Assistance Enhance or Erode Expertise? · Hardman — The Disappearing Bottom Rung in L&D (Sep 3)

W3 — The evidence bar rises again: our assessments reward what AI now does for free [research]

A four-arm RCT at Bocconi (1,053 first-year students, a 45-minute business case, 20 trained raters, three independent ratings per submission) found ChatGPT access raised conventional rubric scores by 0.86 points on a 5-point scale against a control mean of 2.09 — while causal-reasoning training raised the diversity of ideas and slightly lowered rubric scores. Read together, that is an indictment of the rubric, not the training: the instrument rewards the polished standard answer that is now free. OpenAI launched a Learning Lab research network the same week and cited the Bocconi work as founding evidence, though it disclosed neither funding nor participant numbers. Underneath the research, capacity is missing: a Royal Society survey of 9,200+ English teachers found 14% confident across the full range of AI literacy, 18% having received any AI CPD, and 15% with whole-school guidance (edtech) (L&D).

Roadmap implication: The bottleneck is moving from model quality to measurement quality. A product that can show it improves unaided performance — or an assessment that AI cannot trivially ace — is now the scarce asset. Expect 'what does your rubric measure that a frontier model cannot do' to become a real procurement question, and note that 86% of teachers are not yet equipped to ask it.

Links: Bocconi RCT — ChatGPT scores vs critical-thinking training (Sep 2) · OpenAI Learning Lab (Sep 2) · Royal Society — only 14% of teachers confident (Sep 2)

W1 — Agentic recruiting absorbs the funnel: the vendor now arms the candidate [deployed]

Eightfold opened its enterprise AI Interviewer to any job seeker over 18 for free — 5-to-10-minute simulated screens with feedback, no requisition or resume required, data not shared with employers, not used for training, deleted in 30 days, with a NYC Local Law 144 bias audit cited. A vendor arming the candidate side of its own product is a new move in an arms race this letter has tracked from CV prompt injection to candidate copilots. The legal exposure is hardening alongside it: a Morgan, Brown & Joy analysis notes a Duke study finding prompt injections in 1% of roughly 200,000 resumes, and Stanford work finding leading AI detectors misclassified over 60% of non-native-English essays as AI-generated — the risk sits with teams deploying screening they cannot inspect. Candidate appetite remains the gating fact: Gartner data cited in the same coverage puts only 30% of candidates as open to AI-led interviews, with 31% told in advance (TA) (HR).

Roadmap implication: Free candidate-side practice is a distribution play dressed as goodwill — it trains applicants on the format Eightfold's enterprise buyers use, and it generates interaction data. Watch whether rivals match it; and note that a detector misclassifying 60% of non-native speakers is an EEOC exhibit waiting to happen, not a tuning problem.

Links: Eightfold opens AI Interviewer for AI Career Day (Sep 8) · HR Executive — AI delegation, exposure and exclusion (Sep 7) · Recruiting Brainfood #517 (Sep 6)

W4 — Frontline is the proving ground, and the ladder is being rebuilt in the trades first [deployed]

The Lowe's Foundation launched the Building Futures Skilled Trades Coalition — 75+ organizations, a $250M commitment, a target of 1M tradespeople by 2035 (250,000 directly), Gable Grants across 73 community colleges and nonprofits in 30 states, with Nvidia, AT&T, Bank of America, Carrier, GM, DeWalt and Duke Energy named. Paul Fain's reporting puts the combined corporate pledges near $1B against ~1.5M workers, driven substantially by data-center construction: Lightcast counts 20M skilled-trade jobs, 2.1M sought annually against 800,000 completing training — a 1.3M annual gap — with 315,000 trade jobs added in five years from data centers. Kelly Services puts data-center employment at 650,000 by 2026, up 30% on 2023, with 25% of its placed personnel poached by hyperscalers and commissioning PMs at $135K (work) (L&D).

Roadmap implication: The most serious money answering the entry-rung problem is not going into the professions that generated the evidence — it is going into trades where the shortage is priced and the training pathway is legible. For anyone building in professional L&D: the trades have a funded, coordinated, employer-led model running three years ahead of you. Study it or lose to it.

Links: Work Shift / The Job — Anchor Partners (Sep 3) · HR Dive — Lowe's coalition to train 1M workers (Sep 4) · HR Dive — data center skilled-worker shortage (Sep 3)

Ripples

New York City bans student-facing generative AI through eighth grade [policy]

NYC Public Schools barred student-facing generative AI from 2-K through grade 8 for 2026-27, plus companion chatbots at all grades — roughly 600,000 students, about two-thirds of enrollment — while permitting tightly capped high school pilots: up to 50,000 students (~5%), maximum five classes per school, five approved tools with weekly time caps (Quill 15 minutes, Edia 20, Brisk Boost 10-20), two 45-minute AI literacy modules a year, and screen-time caps of 30-45 minutes. Exemptions cover students with disabilities, multilingual learners, and CTE/CS students (edtech) (policy).

So what: The largest school system in the country just decided that default-on frontier AI is a risk to manage rather than a capability to distribute — one week after OpenAI expanded free teacher access to 300,000+ educators. K-12 AI is now a two-track market: teacher-facing tools that clear procurement, and student-facing tools that increasingly must clear politics.

Links: NYC one-year AI moratorium (Sep 3)

The labor market's AI-exposed middle keeps shedding while the headline recovers [research]

Indeed Hiring Lab's read on August: +162,000 nonfarm payrolls, unemployment steady at 4.1%, three-month average job growth up to ~71,000/month from 38,000 in July. Underneath, information and financial activities shed a combined 34,000, while local government education added 42,000 and leisure and hospitality 62,000. July JOLTS showed openings at 7.3M with the hires rate down 0.2pp to 3.2% and quits at 1.9% — what Hiring Lab calls low-hire and low-fire starting to look 'less like a phase and more like the new normal.' Challenger counted 52,881 August cuts, up 58% month-on-month but down 38% year-on-year, with AI dropping out of the top attribution slot for the first time since February (116,175 AI-attributed cuts year-to-date) (work).

So what: A frozen market hurts entrants most: when nobody quits, nobody backfills, and the entry rung stays closed regardless of what the headline print says. Watch the hires rate, not payrolls, as the leading indicator for early-career hiring.

Links: Indeed Hiring Lab — August jobs report (Sep 4) · Indeed Hiring Lab — July JOLTS (Sep 1) · HR Executive — August job cuts up 58% (Sep 4)

OpenAI's own data shows AI adoption running inverted against seniority [research]

OpenAI published two pieces of usage economics this week. Frontier firms — its top decile by AI usage — now generate 8.3x the output tokens per active user of typical firms, up from 2.6x in January, a fast-widening adoption gap. More pointed for this letter: six months after adoption, early-career employees were sending 13 more messages per week than executives. OpenAI also reports running 3.1 agent-workdays per human workday internally, against 1B+ weekly active users and 2.5M businesses (work) (L&D).

So what: The people with the least accumulated judgment are using the tool the most — which is precisely the population the Autor RCT found bifurcating rather than reliably improving. That is a training design problem with a measurable population, and it belongs to L&D this quarter, not next year.

Links: OpenAI — AI-native company workflows (Sep 1) · OpenAI — The Work Now Within Reach (Sep 8)

Anthropic opens real usage data to outside researchers [research]

Anthropic gave three independent academic teams privacy-protected access to roughly 250,000 real Claude conversations from April-May 2026 through 'Anthropic Insights,' aggregating without exposing raw text. Partners include Stanford's Social and Language Technologies Lab, Oxford's Human Information Processing Lab, METR, with Imperial College London running a privacy audit. Reported findings: over 50% of conversations involved consequential work affecting others, and about 75% had users setting direction rather than delegating wholesale (work) (research).

So what: The single biggest gap in this beat is that nobody outside the labs can see how AI is actually used at work. A frontier lab handing that to independent auditors sets a disclosure precedent every enterprise AI vendor will eventually be measured against — and gives L&D its first external baseline for what good delegation looks like.

Links: Anthropic opens Claude usage data to researchers (Sep 7)

Salesforce puts $166M into HiBob at $3.2B — the incumbents still hold the data [deal]

Salesforce led a $166M round in HiBob at a $3.2B valuation, bringing total raised to roughly $700M with an IPO pushed to 2027 or later. Chad & Cheese read it as evidence against the SaaSpocalypse thesis: agents need systems of record, and the incumbents own them. The read is consistent with Workday's August print, which this letter graded a pass — AI above 25% of new ACV and agent ARR now a disclosed line (HR) (work).

So what: A CRM leader buying into an HCM platform is a bet that the employee data layer is where agentic HR gets arbitrated. If agent ARR becomes a standard disclosure, the HCM category gets repriced on it within three quarters — and the private-market comps move first.

Links: Chad & Cheese — HiBob Beast Mode (Sep 4)

Nearly half the time spent with AI goes to fixing its output [research]

A BambooHR survey of 1,600+ US salaried workers found roughly 1.5 hours a day spent on AI — about 47 days a year — of which some 20 days go to troubleshooting and prompt iteration. Enthusiasm is high anyway (65% confident or enthusiastic, 58% motivated by time savings), and more than a third say knowledge transfer now happens mainly through AI. Separately, an Idealis study found 72% of AI users worried about job security versus 46% of non-users (HR) (work).

So what: Self-reported, so treat the precision loosely — but the direction matters: the rework tax is large enough that productivity claims quoting gross time saved are close to meaningless. And if a third of workers now learn primarily through AI, the knowledge-transfer channel organizations never designed has quietly become the main one.

Links: HR Dive — half of AI time spent fixing output (Sep 2) · HR Executive — 3 of 4 gen AI users fear for job security (Sep 8)

Watchlist — what moved

Scoreboard — 19%. The AI employment gap for workers 22-25 in AI-exposed occupations (Stanford DEL, ADP payroll data) — no update this window; next release ~Sep 23. The number we track until it bends.

The Tape — pre-seed to Series B, week of Sep 1

Domestic (US)

International

4 rounds, $15.5M disclosed (one undisclosed) — 1 domestic / 3 international. Third consecutive quiet week, and the honest read is quieter still: Ollie and Catch are agentic-workspace plays that qualify only under the broad 'work tech' leg. Score the Tape strictly to the human-capital stack and the week is 2 rounds, $3M, zero domestic. Zero US edtech, TA, L&D or HR-tech venture rounds were announced in the window. Eleven candidates were rejected as republished older rounds — including Humanly's $25M Series B (actually April 2026) and CandorIQ (July 2025) — with HRTechFeed the single largest source of stale datelines.

Read the archive: https://excelsiorgroup.ai/insights/signal/human/