The Signal: Human Advancement — Edition #8 — October 8, 2026

The Excelsior Group · The de-siloed AI × (education · work · talent acquisition · L&D · HR) letter · Covering October 1 – 7, 2026

The Read

A university ran a randomized trial on its own AI study tool and published a result nobody selling one wanted: in sections where the tool was offered, course grades fell — and only about 15% of the students offered it ever opened it. The same week, the best administrative data yet on workplace chatbots came out of Denmark: where employers encourage and equip people, 93% use the tools, and through the end of 2024 the effect on earnings and hours was indistinguishable from zero. Put the two together and the week's finding is plain. Availability is not adoption, and adoption is not outcome; the tool does little until the work around it is redesigned. Capital and incumbents acted as if they agree. SAP bought TechWolf, a company whose product is a map of what people in an organization can actually do, and Anthropic committed $100 million to training 10,000 engineers inside its customers rather than to another model. The labor market stayed still while this happened: 29,000 jobs added in September, unemployment at 4.2%, and the weakest September hiring plans Challenger has recorded since 2011. Old problems, new physics: intelligence is now the cheap input, and the scarce ones are a description of the work and people able to rebuild it.

Tide status

T1 — the entry rung is breaking: held; no new reading from the series it rests on, and one new study that complicates the mechanism. Stanford's Digital Economy Lab published nothing in the window; its headline gap for 22-to-25-year-olds in AI-exposed occupations remains 19% (data through June), and its dashboard is updated monthly. Two housekeeping points on the ADP Research monthly we cited last week. First, the September 23 release carries August data; Edition #6 described that refresh as July data, which was wrong. Second, last week's tide note printed the 22-to-25 comparison without its sign: employment for that age group fell 4.4% year over year in high-exposure occupations and also fell 2.0% in the least-exposed ones. The new evidence is from Denmark. Humlum and Vestergaard, writing at Brookings on surveys of about 25,000 workers in 11 exposed occupations linked to administrative records through December 2024, report that Denmark shows the same fall in early-career jobs in exposed occupations seen in the United States — but that the early-career share of employment moved no differently at workplaces that encourage chatbot use, with estimates tight enough to rule out a gap larger than a third of a percentage point. That sits uneasily beside the Danmarks Nationalbank paper we reported in Edition #6, which found adopting firms hiring fewer young workers; the two use different measures and periods, and reconciling them is this letter's open question, not a settled one. Revelio Labs added a non-graduate version of the same pattern: global call-center headcount is 5.3% below its December 2023 peak, and across 12 middle-income countries the entry rate into call-center work has fallen from 25.1 to 17.2 per 100 workers since November 2022 while the exit rate also fell. Fewer people getting in, not more being pushed out — the mechanism this tide has rested on since Edition #6. T2 (instruction cost → ~zero): no movement; the week's evidence (Waves, W3) is about what free, default-on tools do not do by themselves. T3 (the credential's monopoly is ending): two data points, no movement — Gallup's share of Americans calling college “very important” fell to 31%, and Workforce Pell stands at 21 federally approved programs across six states. Cross-reference: the daily Signal marked its AI-as-worker tide “confirmed and strengthened” on October 2 and ran Anthropic's Frontier Academy that day; this letter grades the program below (W5). The lenses agree.

Waves

W3 — The evidence bar in learning: a university tests its own AI tool, finds lower grades, and shows what “access” is worth [research]

Researchers at the University of Maryland randomized 30 instructors, teaching 2,379 undergraduates, to offer or not offer the university's Virtual Study Assistant, a chatbot inside the learning management system that answers from course materials (EdWorkingPaper 26-1598, Annenberg Institute at Brown). Among sections of the same course, being offered the tool reduced final grades by 0.37 standard deviations — about four points on a 100-point scale — and recorded LMS participation by 0.90 standard deviations; the losses were larger among first-generation students. Across the full sample the grade effect was smaller, about three points, and statistically weaker, while the participation decline was similarly large. The detail that determines how to read it: only about 15% of students offered the tool ever used it, typically for a handful of sessions, and 73.8% of requests asked for information, explanations or solutions against 0.7% seeking feedback on the student's own work. The tool defaulted to a direct-answer mode and only two instructors switched it to the step-by-step tutoring mode. Phil Hill's On EdTech argues the paper therefore tested an answer-giving assistant switched on with light instructor support, not tutoring, and that a grade effect across whole sections cannot run through the 15% who used it. Provost Jennifer King Rice, a co-author, told Inside Higher Ed “we don't have clear findings from the study,” and the university is holding off on wider availability. In corporate learning, Philippa Hardman made the parallel point about AI video: a vendor survey (Synthesia) finds 52% of L&D teams now use AI to make video and 88% report value from time saved, while the research she cites (Noetel and colleagues, 105 randomized trials) finds video added to active teaching has a far larger effect than video swapped in for it. And the 2027 EDUCAUSE Top 10 put “Determining where AI adds real value” first, Campus Technology reports. (edtech) (L&D)

Roadmap implication: Read the Maryland result for what it measured: offering a course chatbot, without redesigning the course around it, did not help and may have hurt — and it is to the university's credit that it ran the trial on its own product and published. The procurement lesson is the one last week's EdReports scan pointed to: ask vendors and internal teams for usage among the students offered the tool, the mode it runs in by default, and an outcome measured against a comparison group. A tool most students never open cannot be the mechanism of a benefit either.

Links: The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial · An AI Tutor Study That Barely Tested Tutoring · A University's Research on Its Own AI Tool Raised More Questions Than Answers · L&D Is Making More AI Video Than Ever. But Is Any of It Helping People Learn? · AI Leads 2027 Educause Top 10

W2 — Private capital reprices incumbents: SAP buys the skills-inference layer [deal]

SAP agreed to acquire TechWolf, the Ghent company whose models infer the skills and tasks inside an organization from its systems of work rather than from self-reported profiles. Terms were not disclosed; SAP expects to close in the fourth quarter of 2026, subject to regulatory approval. SAP says TechWolf will remain an independent entity under CEO Andreas De Neve, stay available to non-SAP customers, and become a core of the SuccessFactors portfolio; named customers include HSBC, GSK, Ericsson, AMD, MetLife, PayPal and Booking.com. TechWolf says it is the largest acquisition of a venture-backed software company in Belgian history, per Tech.eu. Thomas Otter of Acadian Ventures, an early TechWolf investor and therefore an interested party, gives the strategic reading: agents cannot be trusted in production without a description of who knows what and how work is organized, and that description is what SAP bought. The next day SAP said AI assistants for payroll, HR service, learning, recruiting, onboarding and career development are available this month (vendor-stated; no customer numbers). Two smaller deals fit the pattern from the other side: Juicebox agreed to acquire the intellectual property and other assets of Fetcher, a 2016-vintage sourcing-automation company that is winding down after raising more than $30 million, and UK edtech company Kortext acquired StudyStash, an adaptive AI study platform. (HR) (L&D) (TA) (edtech)

Roadmap implication: Last week this letter said the education publishers were assembling the measurement layer faster than the HCM suites. SAP answered within seven days. Count it with Pearson–Workera and Phoenix–Fuel50: the strategic money of the past month has bought descriptions and measurements of human capability, not content and not workflow. For anyone building agents for HR or for work generally, the scarce asset is now explicit — a trustworthy, current map of skills and tasks — and the independent ones are being taken off the table. Workday's analyst day is October 13; watch whether it answers.

Links: SAP to Acquire TechWolf, Giving Enterprises Evidence-Based View of Work in the Age of AI · SAP buys TechWolf in record Belgian VC-backed deal · SAP agrees to acquire TechWolf · SAP Accelerates Autonomous HCM · Juicebox to Acquire Fetcher Assets · AI study platform StudyStash acquired by Kortext after rapid global expansion

W5 — The engineered apprenticeship: a frontier lab pays to train the people, not the model [pilot]

Anthropic announced the Claude Frontier Academy, a $100 million commitment to train 10,000 “Frontier Deployed Engineers” by the end of 2027. The design is a residency: a multi-day in-person program with a graded practical, then twelve weeks leading a real Claude project inside the engineer's own employer, with a final assessment. Organizations nominate candidates; first cohorts are in San Francisco, New York and London; named participants include Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk. The first full credentials are expected in early 2027, so the program has started and has certified no one yet. The demand side shows why a vendor would pay for this. The Conference Board's third-quarter survey finds 82% of chief HR officers say AI has changed workforce planning, 19% have built its effects into enterprise-wide workforce and financial planning, and 4% feel very prepared to lead through it. Workday's Global Workforce Report (vendor research on its own customer data and a survey of about 6,000) finds 40% of leaders expect AI to get more from current employees against 28% who expect it to reduce headcount; posting demand for basic AI skills fell 25% from a January 2026 peak while demand for hands-on skills such as building AI tools and automating workflows rose 51% between September 2025 and July 2026; mentions of training skills fell 13%; and internal moves fell at 57% of employers year over year. A new NBER paper shows the formation problem at the top of the education system: AI-generated writing is present in 29% of US STEM doctoral dissertations filed in 2026, and within the same program and cohort those graduates are less likely to enter academic research careers. Ethan Mollick's essay of the week concedes that agents now organize themselves better than he expected, which leaves choosing the problem as the human job. (L&D) (work) (HR) (edtech)

Roadmap implication: The apprenticeship this wave tracks is now being built by a model vendor, for its customers' staff, on its own credential. That is a rational response to the Maryland and Denmark results — the constraint is people who can redesign work — and it is also a dependency: the person who can rebuild your process is being trained and badged by your supplier and seated first at your consultancies. Employers should decide deliberately who owns that capability. And note Workday's quiet number: internal mobility fell at most employers in the same year everyone agreed reskilling was the answer.

Links: Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap · Survey: CHRO Confidence Remains in Positive Territory, But Continues to Inch Down · Workday Global Workforce Report: AI Is Rewriting Jobs More Than It's Cutting Them · How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations · The Dot and the Swarm

W1 — Agentic recruiting absorbs the funnel: the interviewer goes on sale, and the sourcing layer consolidates [deployed]

HackerRank moved Chakra, its AI interviewer for technical hiring, from roughly six months of beta toward general availability on October 5, per TechCrunch. The company says Chakra conducted more than 500,000 interviews in testing, that suspicious-activity flags ran 70–80% lower than in comparable traditional HackerRank assessments, and that it collapses what had been three rounds into one; Snowflake, Snorkel and Capgemini are named as testers. All of that is vendor-reported, and the chief executive's claim that AI is “way less biased than humans, if you tune it properly” came with no audit. Workday's customer data describes the volume these tools are sold against: the median number of applicants per filled job rose to 69 from 58 a year earlier, time to fill held at about 60 days, and 84% of job seekers in its pulse survey said they used AI in their search. Juicebox's purchase of Fetcher's assets (W2) is consolidation in AI sourcing. LinkedIn's Hiring Assistant 2, which this letter covered last week, begins rolling out in November according to LinkedIn — it is announced, not in use, and last week's Watchlist should not have tagged it deployed. (TA) (HR)

Roadmap implication: Applicants per hire up by a fifth with time-to-fill flat is the arithmetic of an arms race: both sides automated, and the employer's cost of reading rose with no gain in speed. Interview-stage automation is the buyers' response, and it is arriving faster than evidence about who it screens out. Neither of this week's launches came with adverse-impact numbers.

Links: HackerRank's AI interviewer offers a glimpse into what job interviews could become · Workday Global Workforce Report: AI Is Rewriting Jobs More Than It's Cutting Them · Juicebox to Acquire Fetcher Assets · LinkedIn Announces Hiring Assistant 2, its Next Generation Hiring Agent, to Help Recruiters Deliver Stronger Hiring Outcomes

Ripples

Denmark's register data: near-universal use where employers enable it, and no measurable effect on pay through 2024 [research]

Anders Humlum (University of Chicago) and Emilie Vestergaard (University of Copenhagen) summarize at Brookings two rounds of surveys, in late 2023 and late 2024, of workers in 11 AI-exposed occupations — the 2024 round drew about 25,000 responses across 7,000 workplaces — linked to monthly administrative records through December 2024. 43% of these workers have employers that explicitly encourage chatbot use and 6% are prohibited. Where employers combine encouragement with enterprise tools and training, 93% have used chatbots at work, 28% use them daily and 19% report saving more than an hour a day; where employers take no active steps, about 41% have used them. The earnings result is a precise zero: confidence intervals rule out average effects on earnings or hours larger than 2%, for workers and for workplaces, the null holds in each of the 11 occupations, and 98% of adopters say the tools have not affected their pay. The movement is in jobs rather than wages: adopters are more likely to change occupation, and those who do see earnings growth run 12 percentage points ahead of other Danish workers. (work) (HR)

So what: This is strong evidence that employer policy, not worker enthusiasm, drives adoption — and that adoption inside an unchanged job does not show up in a paycheck. The data stop in December 2024, before agentic tools, so it is a reading of the chatbot era, not of 2026. For an HR leader the usable finding is the gap between 41% and 93%: encouragement, enterprise access and training roughly double use. What to do with the time saved is the part no one has measured.

Links: Still waters, rapid currents: Early labor market transformation under generative AI

September's labor data: hiring stalls, AI is cited for 9% of the month's announced cuts, and the call center shrinks from the entrance [research]

BLS reported payrolls up 29,000 in September with unemployment at 4.2%, both “changed little,” and revised July to a loss of 10,000 and August down to a gain of 133,000. Challenger, Gray & Christmas counted 43,281 announced job cuts, down 18% from August and 20% from a year earlier; employers cited artificial intelligence for 3,961 of them, about 9% and the fifth-most-common reason for the month, though at 120,136 it remains the most-cited reason for the year to date, about 21% of all announced cuts. Announced hiring plans of 90,787 were the lowest September total since 2011. These are reasons employers state in announcements, not measured causes. One example of why the label needs care: HubSpot is cutting about 660 jobs, roughly 7% of its workforce, as it reorients around AI, and its chief executive's letter says the decision “is not driven by AI-related efficiencies.” Revelio Labs supplied the week's structural reading: global call-center headcount is 5.3% below its December 2023 peak and has fallen year over year in each of the last eight quarters — the only eight such declines in 70 quarter-end readings since 2009. Of workers who leave, 49.7% move sideways into other customer-service, sales or office work and 10.8% reach higher-paying technical-support, customer-success or software roles. Revelio does not claim to observe which firms replaced people with chatbots. (work) (TA)

So what: The pattern is the one this letter has reported since August, now in a second occupation and outside the graduate market: few layoffs, weak hiring, and a narrowing entrance. Call-center work has been a first rung into the formal economy in a dozen middle-income countries; an entry rate down by a third matters more to those workers than any US headline. And when a company says its cuts are about strategy rather than AI efficiency, the honest count records what it said.

Links: The Employment Situation - September 2026 · Job Cuts Fall in September; Hiring Plans Up 3% Over 2025 On Weak Early Seasonal Hiring · After a Decade of Growth, Global Call Center Employment Is Shrinking · HubSpot to cut 660 jobs as the company pivots to AI

California, read closely: what the “No Robo Bosses” law requires — and a correction to last week [policy]

The employment bar spent the week reading the two laws Governor Newsom signed on September 30, and this letter owes a correction. Edition #7 said SB 947 requires employers to tell affected workers which tools were used and to provide summaries of the personal data considered. It does not. As Kilpatrick Townsend and Jackson Lewis describe the statute: an employer may not rely solely on an automated decision system to discipline or terminate; where it relies primarily on one, a human must independently review and corroborate the output; the worker receives a written notice that an automated system was primarily relied upon, that a human reviewed it, and whom to contact; and a description of the worker's own data is provided on the worker's request. Kilpatrick puts the civil penalty at $500 per violation, enforced by the Labor Commissioner and public prosecutors, with no express private right of action; the law takes effect July 1, 2027 and, as both firms describe it, applies to discipline and termination decisions. Jackson Lewis adds the point HR-technology buyers should hear: productivity, attendance and insider-risk tools can qualify as automated decision systems even when IT or security bought them. On SB 951, Littler and Fisher Phillips report that from January 1, 2027 a covered layoff notice caused “in whole or in substantial part” by AI or other automated technology must carry the heading “This notice is for a technology displacement” and state the job functions being automated and the type of technology; the two firms note that “in substantial part” is undefined. (HR) (work) (policy)

So what: The compliance work is narrower and more procedural than “AI cannot fire you”: an inventory of which systems feed discipline decisions, a documented human review that uses evidence beyond the system's output, a notice, and a data description on request. Vendors should build the audit trail and the export; employers should start with the inventory, because the tools in scope are not all in HR's budget.

Links: No Robo Bosses: California Requires Human Review Before AI-Driven Workplace Discipline · When the Productivity Monitor Becomes the Boss: California's No Robo Bosses Act and Employee Monitoring Platforms · California Amends its WARN Act (Again) · California Employers Will Soon Need to Say When AI Causes a Mass Layoff: 5 Steps to Prepare

An AI-screening bias suit fails on the evidence, not on the merits of the tools [policy]

A federal court granted Sirius XM summary judgment on September 30 in Harper v. Sirius XM Radio, HR Dive reports. The plaintiff, a Black job seeker, applied to about 150 roles between November 2023 and November 2024 and alleged that the company's iCIMS software screened him out in a way that discriminated by race. The court found no direct or circumstantial evidence of discriminatory intent, and on disparate impact held he had not shown how the employer's alleged use of AI could have discriminated against him on the basis of race. iCIMS said its customers “decide whether and how to use AI features, and their hiring teams make the hiring decisions.” HR Dive does not name the court, and this letter has not read the order. (TA) (HR) (policy)

So what: The ruling says a rejected applicant cannot get to trial on volume of rejections alone; it does not say the screener is fair. The evidence that would decide such a case — selection rates by group — sits with the employer and the vendor, and with federal disparate-impact enforcement curtailed, almost no one is obliged to produce it. Mobley v. Workday, where the fight over that kind of data is still being litigated, remains the case that matters.

Links: Sirius XM beats allegations of AI-fueled racial bias

OpenAI publishes teen usage numbers the same day a watchdog rates its teen product an “unacceptable risk” [deployed]

OpenAI said that in one week nearly 1.2 million teens used ChatGPT's Learning Visualizations and more than 180,000 used Study Mode, that teens average under 15 minutes a day, and that a College Planner for US students in grades 10–12 is coming soon, alongside a partnership with College Advising Corps. These are usage figures from the vendor; none is a learning outcome. The same day, Common Sense Media's Youth AI Safety Institute rated ChatGPT for Teens an “Unacceptable Risk” after testing more than 4,000 prompts on accounts registered to 13-to-17-year-olds. On the learning features it found Study Mode often surfaced a “Show me the answer” option and completed full homework assignments when it was selected, and that caregiver-set study hours could be switched off by deleting a prefix from the prompt; on safety it reported zero alerts to linked parent accounts after extended conversations about self-harm on new accounts. Education Week reports OpenAI disputes the methodology. The Institute discloses that its funders include the OpenAI Foundation. (edtech)

So what: Two releases, one product, no overlap in what they measured: reach on one side, a structured stress test on the other. With New York City and Los Angeles already restricting student-facing generative AI this school year, districts will weigh the stress test more heavily. The gap neither fills is the one the Maryland trial addressed for a different tool — whether students who use it learn more.

Links: Helping teens learn, plan, and shape the future of AI · ChatGPT for Teens Poses Unacceptable Risk to Kids, Common Sense Media Finds · ChatGPT's 'Teen Mode' Is Unsafe for Students, Report Concludes

The degree's standing hits a new low while its statutory alternative moves at the speed of paperwork [research]

Gallup finds 31% of Americans now say a college education is “very important,” the lowest since it began asking in 1978 and down from 70% in 2013; 29% say “not too important,” up from 6% (1,200 adults, fielded in August, per Higher Ed Dive). The federal alternative is early: Paul Fain counts 21 Workforce Pell programs approved across six states so far, six of them at North Carolina's Forsyth Tech, with 48 more programs at 20 other North Carolina colleges awaiting a federal decision. Phil Hill's reading is that the announcements are running ahead of the rollout — approvals are arriving, with no signs yet of meaningful, at-scale adoption. (edtech) (work) (policy)

So what: Public confidence in the degree is falling faster than anything is being built to replace it. Twenty-one approved programs is a pilot, not a pathway system. The number to wait for is students enrolled and placed, which the program's completion and placement requirements will eventually force into the open.

Links: Share of Americans who say college is 'very important' hits a new low · Ambassador to Business · Workforce Pell: The press releases are ahead of the rollout

Watchlist — what moved

Scoreboard — 19%. The AI employment gap for workers 22–25 in AI-exposed occupations (Stanford Digital Economy Lab, ADP payroll data, through June) — the last published headline figure, unchanged. Latest monthly read (ADP Research, published September 23, August data): employment of 22–25-year-olds down 4.4% year over year in high-exposure occupations and down 2.0% in the least-exposed, the 35th consecutive month of contraction in the high-exposure group.

The Tape — pre-seed to Series B, October 1 – 7

Domestic (US)

No qualifying rounds verified this week.

International

Tally: 1 round, $9M — 0 domestic / 1 international; no new sector fund. This is the quietest Tape since the shutout of August 18–25, and the first week of the fourth quarter. The week's capital moved through M&A instead: SAP–TechWolf, Juicebox–Fetcher (assets), Kortext–StudyStash, and Personio's purchase of spend-management company Circula. Noted but not counted: Vitalize, a reported $31M Series A led by Oak HC/FT for hospital staffing and scheduling software — seen only on an aggregator, no primary announcement found, and at the healthcare-operations edge of this letter. Rejected: Scooter (about $180K pre-seed, Pune — a sales-hiring marketplace with no AI claim), Okulyo ($700K, Turkey — campus payments), Mallow (€11M, Paris — a children's voice-gaming device), and re-surfaced older rounds from Notey, Ema, Wild Zebra, Metaview and last week's five. Blind spots: Southeast and East Asia sources failed again, FinSMEs was only partly readable for October 2–6, and a small US round announced only by press release could have been missed.

The Signal: Human Advancement is published weekly by The Excelsior Group. Archive and subscribe: https://excelsiorgroup.ai/insights/signal/human/