Human Tutor Intelligence: Analysis of One-to-One Expert Lessons
An exploratory study of one-to-one teaching, and what it asks of infrastructure.
The question
Among the interventions a skilled tutor makes, which could a current AI system have made with the information available at that moment, and which could it not?
The second half of the question matters more. If some interventions depend on information no system records, the binding constraint on better teaching is capture, not capability.
Four categories
- Diagnostic reasoning: naming the specific cause of an error, not the error. Most frequent category, three to four moves per session.
- Metacognitive coaching: handing over a transferable procedure rather than an answer. Two to three per session.
- Contextual scaffolding: attaching a concept to this student's own interests and week. Around two per session.
- Affective calibration: reading motivational state and changing the plan. Lowest by count, highest by consequence.
What AI could already do
- Canonical content and framework delivery: thinker tables, specification content, structured study packs.
- Generic structural feedback at scale: prompting a student to state the causal link, or close each paragraph with a judgment. This is the most common error in the corpus and can be caught on every paragraph, tirelessly.
- Spec-bounded practice and marking: predicted questions, model answers, mark-scheme marking, retrieval drills.
What it could not
- Scaffolding from the student's own life. Teaching an abstract concept through the feed they were scrolling that morning requires knowing them and what is in front of them this week.
- Affective calibration to a disclosed pattern. A twenty-minute exam freeze was disclosed because of trust, and met with a changed plan and a real accountability structure outside the session.
- Cross-context diagnosis. Naming that a student's analytical strength is causing an application error, and linking it to the same pattern in another subject, needs a longitudinal model of the individual.
The asymmetry is the finding
Management information systems record attendance, summative grades, behaviour incidents and flags. They record the outcome and discard the reason.
The diagnostic insight behind the grade, the interests used as hooks, the motivational pattern, the frameworks a student has actually internalised: none of it is captured anywhere. It lives in the teacher's head and evaporates at the end of the hour.
The implication is not that schools should run tutoring companies. It is that the expertise their teachers already exercise is being lost for want of a layer that can hear it.
Limitations
One tutor, one exam context, A-level humanities, a single analyst coding structured notes rather than verbatim transcripts. There is no controlled outcome data, so nothing here shows that the coded moves cause learning gains.
Moving from what tutors do to what produces outcomes needs verbatim transcripts, a larger and more diverse sample, inter-rater reliability checks, and a design that links coded interventions to assessment trajectories.
Method and anonymisation
Approximately 25 one-to-one A-level sessions with five students across three subjects, March to June 2026, coded by a single analyst from structured session notes. Exploratory, not a controlled trial. Students are pseudonymised and every intervention is reported in paraphrase, never verbatim.
educave.ai · Free to copy, forward and use inside your institution, with attribution.