Directing Intelligence Digest | November 11, 2026
For the smallest sellers on a leading Chinese online marketplace, paying market rates for product photos, videos and listing copy would have cost the median firm in the bottom size quarter 22.7% of a year’s sales. In the top quarter, 0.4%. Then the platform built generative AI into the software every seller already uses, and the smallest shops grew fastest. That’s from a new NBER working paper from the University of Wisconsin-Madison and two Zhejiang universities, passed around on X this week by Raffaella Sadun and Kevin Bryan.
Today’s video walks through a night market of tiny stalls, then into a classroom where the lights dim. Watch it asking: in your organization, is AI something each person has to figure out alone, or is it built into the tools your newest people already use?
The short version
Built in, the smallest gain most: small sellers who adopted the platform’s built-in AI saw monthly sales rise 51.6%; the bottom quarter by size grew 61.5%, the top quarter 10.3%. (NBER w36003)
Left to individuals, seniors pull ahead: in a PA Consulting survey of 2,000 U.K. employees, 70% of senior leaders say AI helps them do things they couldn’t before, versus 38% of entry-level and junior staff. (PA Consulting, Nov 9)
A teacher chatbot backfired: in a semester-long trial with 193 teachers in Türkiye, a GPT-4o teaching assistant lowered students’ motivation and confidence, and scores fell for students of lower-performing teachers. (Stanford SCALE review, Nov 10)
The guidance gap: 79% of surveyed students use AI for homework or assignments; 47% say they were shown how. (OpenAI, citing ThinkYoung, Nov 10)
Here’s AI at work: the smallest shops gained the most
The research. Panle Jia Barwick (University of Wisconsin-Madison), Lu Fang (Zhejiang University of Finance and Economics), and Jinhan Liu, Huilan Xu and Zhe Yuan (Zhejiang University) studied a leading Chinese e-commerce platform (anonymized) that in 2024 embedded generative AI in the software sellers use to run their stores: tools that generate product images, videos and titles, plus ready-made analysis of demand and competitors. With complete records of each firm’s AI use and monthly sales, they compared adopters with matched non-adopters among firms selling RMB 5 million a year or less. NBER Working Paper 36003.
What it finds. Adoption raised monthly sales 51.6%, mostly through 47% more orders, with prices barely moving. The gains tilt hard toward the smallest: 61.5% for the bottom quarter by size, 10.3% for the top. The smallest sellers’ negative reviews fell 42.5%, while human customer-service time per order fell 5.6% across adopters. Categories where the tools mattered most drew more new sellers, which the authors read as market expansion. Their line: “Scale economies in producing AI need not become scale advantages in using it.”
How it fits. Self-serve adoption looks different. PA Consulting’s survey of 2,000 U.K. employees, published November 9, found senior leaders using AI more than three times as often as entry-level and junior staff, and 70% of them saying it helps them do what they couldn’t before, versus 38% of juniors. A third of the workforce, 32%, hasn’t seen meaningful benefits. PA’s Kate Spencer: “The employees who already possess greater expertise and judgement are currently deriving the greatest benefits.”
Limits. A working paper, not peer reviewed. One platform in China, physical-goods sellers only. Sellers chose whether to adopt; the authors confirm the results with instrumental-variable estimates and other designs, but it isn’t a randomized trial. PA’s is a self-reported survey from a consulting firm.
What to watch: whether U.S. platforms that embed AI for small businesses publish seller-level results alongside adoption counts.
Read it: NBER | PA Consulting | On X: Raffaella Sadun, Kevin Bryan
Who’s Directing? I’ve built 35 of my own AI agents, and this paper names the thing I’d watch most closely: who carries the setup cost. If I ran a company or a campus, I’d stop counting licenses and list the three systems our newest people open every day, then ask what AI built into those would take off their plate.
Now, here’s what’s happening in learning: the teacher chatbot that lowered motivation
The research. Stanford’s SCALE Initiative (Seth Walker, Lily Fesler, JP Martinez Claeys, Chris Agnew and Susanna Loeb) updated its review of the strongest causal evidence on generative AI in education on November 10. Of 868 papers added to its repository, 201 were trials or quasi-experiments; 32 met its bar, bringing the evidence base to 52. Most new student studies are from colleges, not K-12. Stanford SCALE.
What it finds. The headline: AI helps more when students use it to understand material than to produce work, though neither reliably improves learning. The study that should stop school leaders is a randomized trial in Türkiye by Sungu, Lira and Duckworth (an SSRN preprint, as summarized in the review): 193 teachers and 2,816 students in grades 5 to 12. Teachers got a GPT-4o chatbot for lesson planning, assessment, feedback and admin for a semester. They mostly used it to generate materials and accepted them with minimal revision. Students’ motivation and confidence fell, even though they weren’t told their teachers were using AI. Average scores didn’t change; for students of below-median teachers, they fell. On the student side, a Brazilian trial found ChatGPT cut presentation prep from 5.8 to 3.2 hours, and students who prepared the traditional way did better on a surprise conceptual test 45 days later.
How it fits. OpenAI, citing new ThinkYoung research published November 10, reports that 79% of surveyed students use AI for homework or assignments, but only 47% were shown how. It adds that as AI improves, “the role of teachers becomes more important, not less.” The Stanford review says the same thing in evidence: effects depend on how AI is designed and used.
Limits. Many studies are preprints with small samples, and the review finds no evidence yet on whether gains last. The Türkiye result is one trial. ThinkYoung surveyed teens across Europe, the Middle East and Africa, and OpenAI is reporting it.
What to watch: district and campus AI contracts that require a test of what students can do without the tool.
Read it: Stanford SCALE | OpenAI | On X: David Ross
Who’s Directing? If I were a provost or superintendent, I’d add one line to every AI pilot: measure what students can do after the tool is gone. And I’d build teacher training around editing AI materials, because in the Türkiye trial teachers mostly accepted the drafts as written, and their students’ motivation fell.
Quick hits
Freelance demand is shifting. OpenAI and Upwork found that in 43 of 63 work categories, activities common in ChatGPT lost share of Upwork job posts. Company research on its own platforms. OpenAI, Nov 10
Saved time, not reinvested. Payscale’s AI Workforce Impact Report: 43% of employees say AI saves them three or more hours a week; 19% say the time goes into more strategic work. Vendor survey. Payscale, Nov 10
Triangulation: signal and noise
Signal. An economics paper on a marketplace, a U.K. workforce survey and a review of 52 education trials point the same way: gains go to people who get AI built into their work and keep their judgment in it.
Noise. The marketplace paper is one platform in China; PA, Payscale and OpenAI are reporting their own surveys; the Türkiye trial is a single preprint.
What I’d do. Build AI into the tools your least-resourced people already use, and keep a person editing what it produces. Then ask two questions: who on your team is figuring AI out alone, and what can your students do without it?
Until tomorrow, that’s the intelligence. How it’s directed is up to you.
Todd
Sources
PA Consulting, AI risks widening the workplace divide, new report finds (November 9, 2026)
Barcaui, ChatGPT as a cognitive crutch, Social Sciences & Humanities Open (2025)
OpenAI, Closing the AI guidance gap in education (November 10, 2026)
OpenAI, Work at the Frontier: AI’s effect on freelance work (November 10, 2026)