Forty-year-old proof
One-to-one tutoring works. It just never scaled.
Benjamin Bloom proved it in 1984. One-to-one tutoring combined with mastery learning produces two standard deviations of improvement — the average tutored student outperforms 98% of their classmates. The research has been sitting there for forty years.
The reason it never scaled was simple: personal tutors are expensive. Most families cannot afford them. Most schools cannot provide them. So we built classrooms around the average student, handed everyone the same textbook, and accepted that some children would fall behind.
That constraint no longer exists.
Built for one child
A tutor that starts with what the child already cares about
My son is obsessed with The War of the Worlds. We read the original H.G. Wells novel together last year, and it caught his imagination completely. He draws the Martians. He daydreams about them in school.
So when I decided to build him an AI tutor, I did not start with a curriculum spreadsheet. I started with what he already cared about. The tutor is a custom GPT named after H.G. Wells himself. When my son asks for a maths challenge, it knows his age, knows what the British curriculum expects of him right now, and frames every problem inside the world of the book. A multiplication question becomes a calculation from the alien invasion. An English exercise becomes decoding scrambled orders.
The subject matter does not change. The engagement does.
Mastery, not pace
Why personalisation changes the outcome
Most educational tools still work like a printed textbook. Content is produced for an average student at a given year group, and every child receives the same material regardless of where they actually are.
A personal tutor works differently. It meets the child where they are, generates problems at the right level of difficulty, tracks what they have already mastered, and adjusts when they make mistakes. When my son gives a wrong answer, the tutor confirms what he got right, identifies the specific error without embarrassment, and suggests a related challenge rather than making him repeat the same question. Progress is saved between sessions, so it picks up where it left off.
A student should demonstrate understanding before moving forward, not simply move on because time ran out.
That is what Bloom described as mastery learning. For decades, delivering it at scale was the problem Bloom himself called the Two Sigma Problem. The tutoring worked; replicating it for more than one child at a time did not. AI solves the scaling problem.
No code required
How this was actually built
The tutor runs on a custom GPT, available to anyone with a £16/month subscription to ChatGPT Plus. Building it required no software development, no technical background, and no external tools. The setup asks a series of questions — who is the tutor, how should it speak, what are its boundaries — and from there you add the instructions that shape how it behaves.
For this tutor, the instructions were drawn directly from Bloom’s research recommendations:
- Mastery learning: students must demonstrate understanding before advancing.
- Immediate, specific feedback: confirm what is correct, address what is not.
- Adaptive questioning: vary the challenge based on current performance.
- Growth mindset framing: encourage persistence, not just accuracy.
- Collaborative elements: position the tutor as a thinking partner, not an examiner.
Continuity between sessions is handled by downloading the exercise log at the end of each session as a text file and uploading it at the start of the next. The tutor reviews what was covered and builds on it rather than starting fresh — a tutor that knows your child, remembers their progress, and generates new material on the spot rather than cycling through a fixed bank of questions.
The same pattern at work
Knowledge, Workflow, Agents — mapped to your firm
I am a busy professional. I do not have hours each evening to devise targeted maths problems and English exercises tied to my son’s interests and current curriculum position. Almost no parent does. This tutor does not replace me. It means that when I do sit down with him, we are working with well-structured exercises designed for him, not generic worksheets printed for a class of thirty.
The same logic applies inside your firm. The KWA framework — Knowledge, Workflow, Agents — maps directly onto what is happening here. The knowledge layer is the curriculum content and the child’s history. The workflow layer is the mastery learning process. The agent layer is the tutor acting in real time, generating exercises, evaluating responses, adjusting difficulty, and keeping the child moving forward. When those three layers work together, the agent produces something that would have been prohibitively expensive to create manually.
The wider point
Specific beats generic, every time
The villain in most AI conversations is hype. Vendors promise transformation, deliver a demo that works on their example data, and leave you with a tool that does not fit how anyone in your firm actually works.
This is different. A custom GPT built around a specific child, a specific interest, and a specific curriculum is the opposite of generic. It is built for one person, it improves with use, and it costs less per month than a single hour of private tutoring.
For professional services leaders applying the 5 Steps for AI Leadership, this example lands at the Activate and Amplify stages: a specific tool, built for a specific need, producing measurable results, and designed so others can replicate it. The same approach — identify a specific knowledge base, map it to a workflow, deploy an agent that acts on it — is what drives the 20-40% capacity improvement we consistently see in firms that move past the pilot stage. You do not need a large budget to start. You need a clear brief.