Directing Intelligence Digest | November 13, 2026
Some companies have started judging engineers by how much AI they use. One employee-built leaderboard ranked colleagues by their AI use, and at least one company has set minimum monthly spending targets. So researchers at JetBrains and Carnegie Mellon checked the number against the work. Across five weeks and 404 engineers, a week in which an engineer doubled their usual AI spend went with only about 10% more commits.
Today’s video moves from a wall of spinning meters into a 1960s driver-training classroom where the teacher sets the route. Watch it asking: in your organization, are you counting how much people use AI, or what they get done with it?
The short version
AI spend tracks use, not output: in five weeks of data on 404 JetBrains engineers, a week with twice an engineer’s usual AI spend went with about 10% more commits. (arXiv 2611.13396, Nov 10)
Classroom AI is ahead of the rules in Texas: 86% of school and district leaders said teachers were using AI, but only 29% had a district policy for student use. (Texas 2036, Nov 10)
Design decides whether practice sticks: in a high school math trial in Turkey, students who practiced with a plain chatbot scored 17% lower once it was taken away than students who never had it. A tutor version with safeguards largely avoided the drop. (PNAS, 2025)
A tool that starts with the teacher: Wharton’s Primer interviews the educator and builds nothing until the educator approves the design. It has not been tested in classrooms yet. (Wharton Generative AI Labs)
Here’s AI at work: AI spend measures how busy the assistants were, not what the work was worth
The research. Shyam Agarwal and Bogdan Vasilescu (Carnegie Mellon), Olga Lvova, Agnia Sergeyuk, Anastasiia Serova and Timofey Bryksin (JetBrains), and Thomas Zimmermann (UC Irvine) linked five weeks of coding-assistant use, from June 29 to August 2, 2026, to the commits of 404 JetBrains engineers. A commit is a saved change to a project’s code. AI spend here is the count of tokens, the units AI providers bill by, that each engineer’s assistants used. They also analyzed an anonymous company survey of 412 employees. arXiv 2611.13396.
What it finds. Spend depended mostly on who the engineer was, and the heaviest spenders stayed largely the same all five weeks. Within one engineer, a week with twice the usual spend went with about 10% more commits. Among weeks with similar commit counts, the engineer-week at the 90th percentile of spend used 39 to 240 times the tokens of the one at the 10th. The authors’ recommendation is plain: use spend at most for budgeting, and judge AI-assisted work by its tasks and outcomes.
How it fits. The survey shows where much of the rest goes. In nearly nine of ten reviews of the coding assistants, technical staff said they’d also used them that week for work beyond code, such as analysis, research and planning. The costs show up there too: 55% of the 299 people who answered said they’d handled colleagues’ material they suspected was AI-generated, at an average of 1.6 hours a week. In the authors’ words, token spend has been called “the lines-of-code metric of the AI era,” and their data support the comparison: both are easy to count and loosely tied to what engineers accomplish.
Limits. One company that builds AI coding tools, five weeks, and associations only, since engineers chose how much to use the assistants. Commits miss code quality, review and design.
What to watch: whether companies that ran AI-use leaderboards or spending targets replace them with measures of finished tasks and outcomes.
Read it: The AI Era’s Lines of Code? (arXiv) | Replication package | On X: Kuldeep Pisda on the paper
Who’s Directing? I run my own team of AI agents, so I read this one as a manager. A usage number tells you how much the agents ran. It can’t tell you whether the work was good. If I managed an operations team rolling out AI assistants, I’d keep AI usage numbers on the budget sheet and off every performance review. Then I’d ask each person to show me one finished piece of work the AI helped with and what it changed, because that’s the number the business actually cares about.
Now, here’s what’s happening in learning: the teacher’s design is the variable
The research. Texas 2036, a Texas policy organization, published “In Ctrl of AI in Schools” on November 10, by Renzo Soto, its deputy director of education and workforce policy. It reviews the research on AI and student learning, and pairs it with a fall 2025 survey of 322 Texas school and district leaders from 277 districts, run by the University of Texas at Austin and the Texas Association of School Administrators. Texas 2036 | Task force white paper with the survey (PDF).
What it finds. 86% of leaders said teachers were using AI and 61% reported student use, but only 29% had a district policy for student use and 20% had one for teachers and staff. By Texas 2036’s count, 42 states have issued statewide AI guidance for schools and Texas is one of the eight that have not. On learning, the report finds the evidence mixed, and says the difference “often comes down to how tools are designed, how teachers guide their use and whether students are equipped to think critically and independently.”
How it fits. The clearest case the report cites is a field experiment with nearly a thousand high school math students in Turkey (Bastani and colleagues, PNAS, 2025). With GPT-4 open during practice, grades rose 48% for a plain chatbot and 127% for a tutor version with safeguards. Once access was taken away, the plain-chatbot group scored 17% lower than students who never had it, while the safeguards largely prevented that loss. Wharton Generative AI Labs’ Primer is built on the same lesson. An intake agent interviews the teacher about the lesson, the students’ level and their usual misconceptions, a design agent writes a plan the teacher can read and revise, and nothing is built until the teacher approves it. It is open source, and the lab says it “has not been evaluated as a learning intervention under classroom conditions.”
Limits. The Texas report is a policy review, not new causal research, and its own authors say much of the evidence comes from outside the U.S. and from higher education. The survey is self-reported. Primer has no learning results yet.
What to watch: whether Texas lawmakers take up statewide guidance, and whether anyone runs Primer in real classrooms and reports what students learned.
Read it: In Ctrl of AI in Schools (PDF) | Bastani et al., PNAS | Primer | On X: Ethan Mollick on Primer
Who’s Directing? If I were a community college dean, I’d write the student AI rules for my division this term instead of waiting for a statewide answer. Then I’d ask faculty to set up each AI tool for their own course the way Primer starts: what the lesson is for, where students usually go wrong, and what they should be able to do alone. The research keeps saying the teacher’s setup decides whether the practice sticks.
Quick hits
Remote work, more than AI, tracks the slump in young-grad jobs. Stripe economist Ernie Tedeschi finds that for 22-to-26-year-olds with a bachelor’s degree, the employment rate was 88% in September 2026, a point below three years earlier. In his model, work from home accounts for 56% to 88% of the decline, while AI exposure shows a much weaker link that is statistically indistinguishable from zero. He calls the result suggestive, not causal. Stripe Economics, Nov 12
Triangulation: signal and noise
Signal. A company’s own records, a state policy review and a controlled classroom trial point the same way: the AI number that’s easiest to count says little until a person decides what good looks like. For engineers, that’s finished work. For students, it’s what they can do once the tool is gone.
Noise. The JetBrains study is one company over five weeks; the Texas numbers are self-reported; Primer is untested; the Stripe analysis is suggestive.
What I’d do. Take AI usage off the scorecard and put outcomes on it. Before a classroom tool goes live, have the teacher write down what students should be able to do without it.
Until tomorrow, that’s the intelligence. How it’s directed is up to you.
Todd





