Directing Intelligence Digest | November 18, 2026
If you run a team or a college, someone has probably asked how many of your people have finished AI training. A Bank of Korea study released yesterday suggests that number tells you less than you’d hope. Among workers it followed for two years, the ones who took AI training saw no clear extra gain in time saved. The ones who built working habits did: testing their prompts, checking what came back, and fixing what went wrong. Students say the same from campus, where 86% of Dartmouth undergraduates call themselves mostly self-taught in AI. People are learning AI inside the work, so a leader’s job is to shape those habits and say plainly what the AI is for.
One question to carry through today’s brief: in your organization, who is teaching people to check AI’s work, and who decides what the saved time is for?
The short version
Habits tracked the gains: among 2,024 Korean workers followed for two years, those who regularly tested and checked AI saw their time savings grow 3.6 points more than others. Training showed no clear link. (Bank of Korea)
Saved time isn’t more output yet: Korean AI users save about 3.7 hours of a 40-hour week, but time saved barely tracks work done. (Bank of Korea)
Students are teaching themselves: 86% of Dartmouth undergraduates say they’re mostly self-taught in AI, and three in four say a clear instructor rule guides how they use it. (NBER w36021)
AI at work: in Korea, the workers who built habits gained the most time
What was studied. The Bank of Korea’s employment research team (Oh Sam-il, Seo Dong-hyeon and Yoon Jong-won) surveyed 16,342 employed people aged 18 to 64 in August and September 2026. They re-surveyed 2,024 people from last year’s survey, so they could follow the same workers over time. BOK Issue Note 2026-29.
What it found. Two in three Korean workers (66.5%) now use generative AI for work, up from 51.8%. Those who used it last week saved 9.2% of their hours, about 3.7 hours of a 40-hour week, up from 1.5 hours a year earlier. The researchers then asked why some people’s savings grew more than others’. Training, which the BOK counts as formal courses, self-study or coaching from a colleague or boss, showed no statistically clear link. Active habits did: testing prompts, checking results, working out what went wrong, saving prompts that work, and building AI into the team’s procedures. Workers with those habits saw their savings grow 3.6 percentage points more. The BOK thinks much of the year’s gain came from the AI itself getting better.
Most saved time went back into work. Yet the link between time saved and more work done was a correlation of 0.08, where 1 would mean they always rise together. The BOK concludes that turning saved minutes into results takes changes to workflow, structure and staffing.
How it fits with other work. Gartner’s survey of 589 Australian employees, out yesterday, finds that of workers expected to use AI, 90% have company tools but only 54.6% get formal learning. Its recommended fix is practice built into the work: analyst Jonathan Tabah calls it “productive friction,” such as doing some AI-assisted tasks by hand now and then.
What one company did. Australian logistics firm Toll Group reported yesterday on its October AI Skills Challenge (the company’s own account, not a study). Its core was peers showing real work: one team set up the task and inputs first and cut a several-day process to about half a day. It reports adoption, not results.
Where it falls short. Time saved is self-reported. The BOK says its results are correlations, not proof of cause, and people who build habits may differ in ways a survey can’t see.
What to watch: whether next year’s survey shows those habits leading to better-quality work.
Read it: Bank of Korea | Toll Group | Gartner press release | On X: Korea Times on the report
Who’s Directing? If I ran an operations team, I’d stop reporting training completions as the AI number, because this study found they don’t track the gains. I’d ask each team to share one example a week where someone checked an AI answer and caught something, so the habit gets seen and copied. And I’d write down what the freed hours are for, since the BOK found they don’t turn into results on their own.
AI in learning: students are teaching themselves, and they want clear rules
What was studied. On Monday, Dartmouth economists Bruce Sacerdote and Mason Parris-Bacon released a working paper through the National Bureau of Economic Research (NBER), a nonprofit network of economists. It reports the first round of a survey they plan to repeat with the same students. From June to September 2026 they surveyed 1,201 Dartmouth undergraduates, 332 incoming first-years, and 1,003 students at four-year colleges nationwide through an online student survey panel (College Pulse). NBER w36021.
What it found. 86% of Dartmouth students say they’re mostly self-taught in AI (79% nationally), and only 38% say Dartmouth is preparing them to use it in a career. About four in ten feel they have to use AI on take-home work because so many classmates do. And 75% say that when an instructor clearly permits or bans a particular AI use, that rule guides what they do.
How it fits with other work. An MIT faculty committee reported in August that only 25% of responding MIT students felt the institute was preparing them to use AI. Yesterday’s Digest covered a 19-university trial where instructor-built AI raised scores on exams taken without AI. Students are learning AI on their own, and the clearest lever schools hold is the instructor’s design and rules.
Where it falls short. It’s self-reported, from one selective college plus an online panel, in a first round, and students may overestimate what peers do. It’s a working paper, not yet peer reviewed.
What to watch: Dartmouth is now surveying the employers who hire its students, so we’ll learn whether they see the same gap.
Read it: NBER w36021 | MIT committee report
Who’s Directing? If I were a community college dean, I’d ask every instructor to add one line to each assignment saying which AI uses are allowed and which aren’t, because three in four students say a clear rule guides them. Then I’d move the graded test into class, as many Dartmouth faculty already have, so the grade shows what the student knows.
Quick hits
The judgment you teach your AI may stay when you leave. A University of Vienna and Edith Cowan University paper (arXiv, November 16) argues that the prompts and memory holding a worker’s judgment sit in employer systems, so the company keeps them when the worker leaves. A framework, no data; useful for anyone writing AI policy. arXiv 2611.21456
Triangulation: signal and noise
Where the evidence agrees. A central bank panel, a Dartmouth survey, Gartner and MIT point the same way: people are learning AI inside their work and study, mostly on their own, and the gains follow habits of testing and checking. That’s Directing Intelligence: deciding what the AI does and checking what it gives back.
Those habits are also how people keep Cognitive Coverage, which means having the depth and breadth of expertise the work demands, so you can defend your work and grow from it.
What it can’t tell us yet. All of it is self-reported, and none of it proves cause.
What I’d do this week. Count how often people check AI’s work, because that habit tracked the gains, and write down what the freed hours are for.
Until tomorrow, that’s the intelligence. How it’s directed is up to you.
Todd





