Where AI Belongs in a Learning Function
Memo · 4 min read
To: a leadership team deciding how much to automate Re: the rule that decides where AI goes
AI belongs in two places in a learning function, production and the workflow, and it has to be handled with care in a third, the learning itself. One rule governs all three: it may remove a difficulty after the capability that difficulty used to build has formed, never instead of it. Everything else in this memo is that rule applied.
In production: real leverage, with a ceiling
The first place AI earned its keep in my function was the least glamorous one. Structured course designs render into first-draft deliverables. Narration is generated instead of scheduled. Presenter segments and visual assets are produced without a shoot or a queue. Assessment items are drafted from objectives that were written at scoping. Engineers stopped producing first drafts and started editing them.
The review gate never moved. Every generated asset clears the same subject-matter and portfolio review as anything built by hand, because a faster wrong course is still a wrong course. That is also why the leverage has a ceiling. It reaches drafting and never judgment, and drafting is only part of the work: analysis, design decisions, SME time, review, and implementation are untouched. In the capacity model I run, production leverage applies to the draftable share of build work only, so doubling drafting speed yields roughly one and a half times the capacity, never twice. A model that promises the full doubling does not survive its first quarter, and a function that plans on it will miss its commitments by exactly the difference.
What the compression buys is time, and the time goes to the parts of the job that cannot be compressed: deciding what to build at all, designing the assessment before the content, and holding the quality bar. Used this way, before any headcount request, AI turned a staffing conversation into a smaller one.
In the workflow: the allocation rule
The second place is more consequential. Knowledge systems that put the right information inside the work at the moment it is needed. Agents that absorb whole categories of procedure so nobody has to carry them. This is where AI stops being a production tool and becomes a decision about where capability lives: in the person, in the work, or in the machine.
The rule is offloadability. If a performance is specifiable, a defined outcome in a bounded context, it can move out of the person into a checklist, a sensor, a workflow rule, or an agent, and nothing of value is lost. Most of these moves are pure gain. Nobody develops as a technician by memorizing torque values, and the error rate drops when the system carries them.
Some difficulties were load-bearing, though. They never appeared as objectives, yet they were where situational awareness and ownership got built, as byproducts. The torque value was trivia; diagnosing why it was wrong was the development. Offload indiscriminately and both go, and the dashboards will not show it for years, because they measure the performance and the performance, for a while, improves. So the allocation decision is made deliberately, per task, and recorded with its reason.
In the learning itself: the place to be careful
The third place is where most of the enthusiasm is pointed and where the rule bites hardest. AI tutors, copilots, and assistants promise to remove difficulty from the act of learning. Sometimes that is the right thing to remove. Often it is the thing the learning was made of.
The calculator is the usual counterexample, and it proves the rule rather than breaking it. Calculators removed the difficulty of arithmetic execution after number sense had formed; nobody hands one to a child before the child can count. The sequence is what made the offload safe. An AI that removes the difficulty of the formative act itself, at every stage, with no gatekeeper, removes the formation and leaves the output, and the output no longer certifies anything about the person who produced it.
The evidence is already in. In a randomized trial of roughly a thousand secondary students, an unrestricted AI tutor raised practice-problem performance by 48 percent, and a tutor built with safeguards (hints rather than answers, teacher-written solutions) raised it by 127 percent. On the exam, taken without access, students who had used the unrestricted tutor scored 17 percent worse than students who never had it. The safeguarded group showed no such loss. The difference between the two tools is the whole design brief: one removed the friction, the other amplified it.
Four operating rules
- AI drafts; people decide. Every generated asset clears the same review gate as hand-built work. The gate is the quality system, and it does not move for speed.
- Count the leverage honestly. Model AI production gains against the draftable share of build work only, and plan capacity on that number.
- In the workflow, offload what is specifiable and keep what is load-bearing. Record the decision and the reason for every task that moves into a system.
- In learning experiences, AI assists after the capability has formed or works as friction's amplifier. Hints over answers, feedback over completion. It never substitutes for the act that builds the capability.
The question in front of every learning leader is no longer whether to use AI. It is which difficulty a given use removes, and whether that difficulty was the training.
Source for the trial: Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman, "Generative AI Can Harm Learning," Proceedings of the National Academy of Sciences 122 (2025). Figures verified 2026-10-15.
