The Excelsior Group · The de-siloed AI × (education · work · talent acquisition · L&D · HR) letter · Covering September 8 – 14, 2026

The Read

Two numbers this week refuse to sit in the same story. Indeed's economists report the US labor force shrank by roughly 700,000 in 2026 — only the fifth calendar-year decline since 1948 and the second outside a recession — while Handshake's data, surfaced by Ryan Craig, shows internship postings down at least 10% for three straight seasons and 109 applicants per internship where there were 62. A shortage at the top of the ladder and a wall at the bottom, at the same time. That is the entry-rung tide restated as a labor-supply paradox, and the market answered it in three ways this week: Coursera unveiled Project Helix, the first incumbent attempt to collapse learning, credentials and skills intelligence into one AI-native record — shipping, note, in 2027; Jack & Jill raised $40M to put an agent on both sides of the hiring table, with 385,000 candidates already talking to theirs; and a 515-firm field experiment found that companies shown how to redesign a workflow around AI, rather than handed the tool, generated 1.9x the revenue. Old problems, new physics: the tool is free, the rung is gone, and the return belongs to whoever rebuilds the work.

Tide status

No tide movement this week. T1 (the entry rung is breaking) gets a paradox, not a revision: labor supply is contracting (Indeed: −700,000 in 2026; unemployment 4.1%; payrolls averaging 80,000 a month) while the entry rung stays shut (Handshake: 30% of tech and 42% of professional-services internships erased 2023–25; Revelio: tech postings asking fewer skills but more years of experience). Josh Bersin's 'AI-fueled labor shortage, not a jobs apocalypse' and Gad Levanon's 'white-collar hiring recession alongside an AI-jobs boom' are both right — about different rungs. Stanford DEL publishes September 23; the 19% gap stands. T2 (instruction cost → ~zero) logged a second big-district moratorium (Los Angeles joined New York) and its mirror image (Miami-Dade put Gemini in front of ~100,000 high schoolers) — diffusion is now decided district by district, and the cost curve is irrelevant to that decision. T3 (the credential's monopoly is ending) collected its strongest evidence yet without moving: a Swiss study across 200+ apprenticeable occupations finds practical-skill marks are by far the strongest predictor of earnings and only weakly correlated with classroom grades, and Coursera's Helix is built around a 'portable skills record' rather than a certificate. Watching; not calling.

Waves

W1 — Agentic recruiting absorbs the funnel: agents on both sides, and the funnel becomes a barbell [deployed]

Jack & Jill raised a $40M Series A led by Air Street Capital (Madrona joining; Creandum and Entrepreneurs First re-upping) less than a year after a $20M seed. Its two agents negotiate for opposite sides — 'Jack' for the candidate, 'Jill' for the employer — and introductions happen only when both confirm fit: 385,000+ people have talked to Jack, 25,000 interviews arranged and rising 5,000 a month, 5,350+ employers including Ramp, Attio and Multiverse. hackajob's Archer agent, per Chad & Cheese, did 3x the company's prior annual revenue in six months with Barclays, Comcast, Oracle and Sainsbury's named — vendor-reported, so weight accordingly. BoomerangHR came out of stealth with a $12.7M seed and Warby Parker as a customer, matching alumni back to open roles. And HR Executive's 2026 Top HR Products, the HR Tech Conference's flagship list, put two AI screeners (Eightfold's AI Interviewer, Harver's AI PREVAIL) and Sana from Workday on the podium. Johnny Campbell's frame in Brainfood #518 fits all of it: hiring after AI is not a funnel, it is a barbell — automation at both ends, humans in the middle (TA) (HR).

Roadmap implication: When both sides field an agent, the product is the negotiation protocol, not the screen. Jack & Jill's fit-confirmation handshake is the first real attempt at that layer; expect the ATS incumbents to either buy one or expose one. For buyers: ask any 'AI recruiter' what happens when the candidate also has one.

Links: Our investment in Jack & Jill: a job market that works for you · Jobs Utopia or Fiasco? · New AI tool looks to bring world-class offboarding and alumni management to every HR team · The products that defined the year: A closer look at the 2026 Top HR Products · Recruiting Brainfood - Issue 518

W5 — The engineered apprenticeship: the return is in the redesign, and the expertise has to be there first [research]

Last week the Autor patent-drafting RCT showed AI's durable gains accrue to people who already have expertise. This week the firm-level version arrived. Philippa Hardman's read of Kim, Kim & Koning's 515-firm field experiment: companies shown case studies of redesigned workflows — same tools, same training as controls — found 44% more uses for AI, completed 12% more tasks, were 18% more likely to win paying customers, and generated 1.9x the revenue. McKinsey's State of AI 2026 (1,719 respondents, 97 countries) says roughly 75% of top performers now redesign workflows, up from 55%, while 80% of companies report personal productivity gains and organization-level bottom-line impact unchanged year on year. On the skills side, the Swiss practical-skills study behind Gap Letter #201 finds occupational know-how, not classroom grades, predicts earnings across 200+ apprenticeable trades, and the Philadelphia Fed's 280,000-apprentice study finds 40% of apprentices who started with no college later earned a bachelor's — and that only 26% of the apprentices studied were in registered programs. The counter-signal: an Academy of Management Review paper warns of 'epistemic de-skilling' in managers who outsource judgment to AI, and a GoTo survey has 46% of Gen Z saying reliance makes them feel less intelligent (L&D) (work) (edtech).

Roadmap implication: Three separate literatures now converge on one design rule: build the judgment first, then add the tool, and measure the workflow rather than the seat. Any L&D program that ships 'AI access' without workflow redesign is, on this evidence, paying for the 80% that shows up in surveys and none of the 1.9x that shows up in revenue.

Links: The AI Workflow Redesign Method: How the Top 6% of Users Get Real Returns from AI · College Classrooms Delivering Deep-Fried Degrees · Ties to Industry · Generative AI overuse may hurt leaders' judgment, researchers warn

W3 — The evidence bar in learning is rising: procurement, not the model, is now the bottleneck [policy]

EdTech Insiders' procurement research (EdSignals/Decision Lab) puts numbers on why evidence does not travel in K-12: only 40% of purchasers call adoption frictionless, 94% rely on peer recommendation, formal evidence intermediaries account for 8% of awareness, 62% of teachers say AI guidance is nonexistent or vague, and 15 of 30 states rely on voluntary guidance with 5 in statute. Into that vacuum, Microsoft, the AFT and the UFT announced a 'National AI Safety & Privacy Standard' written to be contract-enforceable — no training on student data, human oversight of high-risk decisions — which Insiders reads, fairly, as a procurement bar incumbents can absorb and startups cannot. The districts are sorting themselves: New York and Los Angeles imposed student-facing moratoria; Miami-Dade rolled Gemini to ~100,000 high schoolers and trained all 18,000 teachers. OECD's PISA 2025 reports declining reading, math and science in most countries — the denominator for all of it (edtech) (policy).

Roadmap implication: The scarce asset in K-12 AI is no longer a better tutor, it is a contract a district can sign. A vendor-authored standard that big systems adopt becomes the de facto evidence tier — which means the evidence bar is rising, but the people setting it sell the products being measured. Watch who else signs Microsoft's standard, and whether any startup can.

Links: How AI is Reshaping K-12 EdTech Procurement · Quick Deep Dive: Microsoft Is Vying to Be the "Safe" AI Company for Schools · A Tale of Three Cities: NYC, LA, Miami-Dade · New York and Los Angeles Ban Student-Facing AI (For Now). Should Other Schools Follow?

W2 — Private capital reprices incumbents' AI risk — two exits, six cents on the dollar and a $140M health-sciences roll-up [deal]

upGrad closed its acquisition of Unacademy at $206M — about 94% below Unacademy's ~$3.4B peak, on $42M of revenue, with the CEO retained. Perdoceo agreed to buy South University for $130–140M cash plus up to $56M earn-out (South: $291M FY2025 revenue, ~10,500 students, its fourth owner in under a decade). Blackboard's first post-bankruptcy deal was Cursive, a process-based academic-integrity tool used by 100+ institutions — buying the AI-era assessment problem rather than an AI feature. McGraw Hill took Teachally (AI lesson generation); Lemnis took Wayfinder, its third acquisition in eight months. On the HCM side, Josh Bersin shipped Galileo's 'Jupiter' release embedding his firm's HR-intelligence assistant inside Microsoft Copilot, Workday, ServiceNow, HiBob and Gloat — an analyst becoming a layer — and Workday's only in-window news was a CMO hire ahead of its October 13 Analyst Day (edtech) (HR).

Roadmap implication: Edtech's 2021 cohort is now clearing at distressed prices while the strategics buy assessment and integrity — the parts AI broke rather than the parts AI enables. That is the pattern to underwrite: the durable acquisitions this cycle are in verification, not generation.

Links: The EdSheet Vol. 41 · For-profit Perdoceo Education to acquire South University for $140M · Introducing The Jupiter Release of Galileo: An Inside Story

Ripples

Coursera unveils Project Helix — the first incumbent attempt to collapse learning, credentials and skills into one record [demo]

At the first combined Coursera–Udemy customer event since the merger closed in May, CPO Patrick Supanc previewed Helix: a learner states a goal in natural language, an adaptive path is assembled from 30,000+ instructors and content partners, assessment runs continuously through observation of work, and the output is a portable skills record rather than a certificate. Scale behind it: 300M+ learners across the combined platforms and 12,000 enterprise customers. Broad enterprise availability is guided to the first half of 2027 — so this is [demo], and it stays there until a customer is named (edtech) (L&D) (TA).

So what: This is the de-siloing thesis stated by the largest learning incumbent in its own product roadmap: learning, credentialing and skills intelligence as one system. The eighteen-month gap between announcement and availability is the window in which every AI-native competitor has to prove it can do the same thing with a customer, not a keynote.

Links: Coursera Announces Project Helix, a New AI-Native Platform Connecting Skills Discovery and Personalized Learning to Verified Capability and Business Outcomes · Announcing Project Helix

Indeed: the goalposts moved — the US labor force is shrinking for the first time outside a recession in a generation [research]

Indeed Hiring Lab's economists argue the labor market must now be read against constrained supply: the labor force fell ~700,000 in 2026, the fifth calendar-year decline since 1948 and the second outside a recession; the civilian labor force is down 973,000 since August 2025; payroll growth has averaged 80,000 a month while unemployment fell from 4.3% to 4.1%. Women added 495,000 jobs over the year against 108,000 for men; foreign-born workers are 19% of the labor force and roughly a quarter of physicians, 17% of nurses and 40% of home health aides. AI appears once, as a hope for productivity — with the authors' caution that 'a labor market with no job growth is unlikely to result in a robust economy' (work) (HR).

So what: Every 'AI took the jobs' reading of soft payrolls now has to clear a simpler explanation: there are fewer workers. That strengthens the Bersin/labor-shortage case for the top of the ladder and does nothing for the bottom — a shrinking labor force with a closed entry rung is a compounding problem, not a self-correcting one.

Links: The Goalposts Moved: Interpreting the Labor Market in a New Era of Constrained Labor Supply · The AI-Fueled Labor Shortage and Why HR Is Now A Growth Function

The internship rung, quantified — and the employer survey that says the opposite [research]

Handshake's data, via Gap Letter #201: internship postings down at least 10% per season for three consecutive seasons; 30% of tech and 42% of professional-services internships erased between 2023 and 2025; 109 applicants per posting, up from 62. Revelio Labs finds tech postings since late 2024 demand fewer skills but more years of experience. HBS research on 150M resumes finds only 19% of workers with twenty years' experience ever reach the top seniority level. Against all of that, Accenture reports 52% of leaders plan to hire more entry-level workers to meet AI demand and 73% intend to create AI-focused entry-level roles within two years — stated intent, not postings (TA) (work) (edtech).

So what: Postings are behavior; surveys are aspiration. Until Accenture's 73% shows up in Handshake's counts, treat the rung as closed and the 'we'll hire juniors for AI' line as a leading indicator to verify next quarter, not a fact.

Links: College Classrooms Delivering Deep-Fried Degrees · Business leaders plan to continue AI spend, even if bubble bursts · Most Careers Stall at Mid-Level, No Matter How Long People Work · Tech hiring 2026: a specialized job market

McKinsey State of AI 2026: 39% expect headcount cuts within a year; the bottom line has not moved [research]

McKinsey's annual survey (1,719 respondents, 97 countries): 39% of organizations expect AI to reduce headcount in the next twelve months; about one in five say AI operating and token costs are now limiting use; roughly 75% of top performers redesign workflows, up from 55%; and 80% of respondents report personal productivity gains while enterprise-level bottom-line impact is unchanged year on year. Exponential View adds the earnings-call view: about a third of corporate AI claims attribute direct financial impact, averaging 6.8 claims and 5.1 quantified per call (work) (HR) (L&D).

So what: The headcount expectation and the flat bottom line are the same finding: organizations are planning to take the return in labor cost because they cannot yet find it in revenue. That is a workforce-planning decision being made on survey sentiment — and, per W5, the firms that did find revenue are the ones that redesigned the work.

Links: The state of AI in 2026: On the road to ROI

Los Angeles joins New York — and Miami-Dade goes the other way [policy]

Within a fortnight of New York City's one-year moratorium for ~600,000 students through eighth grade, LAUSD adopted a broader moratorium on student-facing generative AI on all student-issued devices for 2026-27, imposed administratively. Miami-Dade chose the opposite: Gemini for Education for ~100,000 high school students with all 18,000 teachers trained. Chicago and Clark County allow district-approved tools. The American Psychological Association's September 3 report urges districts to pause and evaluate; IBM's teacher survey puts weekly AI use at 76% of middle-school and 73% of high-school teachers regardless (edtech) (policy).

So what: The two largest districts in the country are now natural experiments running in opposite directions, with teachers using the tools either way. Whoever instruments the comparison — outcomes in LA versus Miami-Dade over one school year — owns the most valuable dataset in K-12 AI.

Links: LAUSD adopts moratorium on student AI use, another big move to limit tech use · A Tale of Three Cities: NYC, LA, Miami-Dade

Consent and the notetaker: AI in the interview room is becoming a legal exposure [policy]

HR Dive's legal read: AI notetakers in recruiting interviews are 'more high risk than your typical Zoom meeting' under all-party-consent states (California, Florida, Illinois) and monitoring-notice laws; SHRM's CHRO Jim Link wants disclosure, consent and 'human beings make HR decisions.' The same week the Dutch regulator fined Uber €825M for automated driver terminations without human review, Phenom put synthetic-candidate fraud detection and adverse-impact monitoring on its AI Day agenda, and a California bill to ban emotion-reading AI at work advanced. Two of HR Executive's Top HR Products are AI screeners (TA) (HR) (policy).

So what: The legal perimeter around agentic hiring is being drawn by consent and human-review rules, not by AI-specific statutes — which means it applies today. Vendors that build disclosure and override into the product will sell; the rest will be someone's exhibit.

Links: HR must set boundaries as it embraces AI notetakers, experts caution · Phenom Unveils Agenda for AI Day 2026

Watchlist — what moved

Scoreboard — 19%. The AI employment gap for workers 22-25 in AI-exposed occupations (Stanford DEL, ADP payroll data) — next release September 23. The number we track until it bends.

The Tape — pre-seed to Series B, September 8 – 14

Domestic (US)

International

5 rounds, ~$60.7M disclosed (one undisclosed) — 1 domestic / 4 international. The first non-quiet week since Edition #1, carried almost entirely by Jack & Jill's $40M A. Quarter to date: 14 rounds, ~$144M of startup capital, one fund ($265M). Rejected as republished or pre-window: Humanly $25M B (still on HRTechFeed's front page; actually April), Wild Zebra $6M (Aug 6), PeopleX $34M A (Sep 2), Dawraty $2M (Sep 6), Emversity, Weskill, Alfa AI, and the Kalos $78.8M fund (May). Out of scope: Tanda/Workforce.com (Thoma Bravo growth stake, not venture), Aron, Synapse Analytics, two generalist funds. No funding was timed to HR Tech — the conference is October 20–22, not this week as assumed.

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