Most writing on this topic is framed as a contest. Read the two best studies of 2025 side by side and the contest dissolves, because they are not disagreeing about AI. They are disagreeing about supervision.
That points to an arrangement usually called hybrid, or blended, tutoring: AI and a human teacher working with the same student, with the work split between them on purpose. This article is about how that split works in practice. If you are still deciding between an AI tutor and a human one, that head-to-head comparison has its own article.
The evidence behind the hybrid model
In Scientific Reports, a purpose-built AI tutor outperformed an active-learning physics classroom for 194 Harvard undergraduates, on both learning and time taken. The system had been deliberately designed around teaching principles by people who understood instruction.
In PNAS, around a thousand high school students given unrestricted GPT-4 during practice ended up roughly 17% worse on the real exam than students who never had it. A guardrailed, tutor-style version landed level with the control instead.
The common factor is not the model. In both cases, the outcome tracked whether a person had decided what the system should withhold. Where someone had made those decisions, results were good. Where nobody had, results were negative.
That is the argument for the hybrid model, and it comes out of the data rather than out of a marketing department.
The division of labour that works
| Responsibility | Who |
|---|---|
| Explaining a concept again, and again | AI |
| Generating practice at the right difficulty | AI |
| Availability at 11pm and on Sundays | AI |
| Marking routine work | AI |
| Deciding what this student needs this month | Human |
| Finding a misconception nobody has spotted | Human |
| Deciding what the AI should refuse to give | Human |
| Noticing disengagement and acting on it | Human |
| Being accountable for progress | Human |
The pattern: AI carries volume, humans carry judgement. Almost every failure case in the research is a volume tool being asked to make a judgement call.
What the human does that the AI cannot see
Three things, specifically, and they are the ones that justify the cost.
Setting the constraint. Someone has to decide that this student, this month, does not get worked solutions in algebra. That is not a setting a child will choose and not one a helpful system will impose on itself.
Reading the gap between confidence and ability. A student who answers quickly and wrongly is doing something different from a student who answers slowly and wrongly, and the difference is usually invisible in the data but obvious in a conversation.
Deciding to stop. Sometimes the right instructional move is to abandon the plan because the child is exhausted or demoralised. No current system makes that call, and a system optimised for engagement is structurally disinclined to suggest stopping.
A week in a hybrid programme
Concretely, rather than abstractly:
Monday, with a person (45 min). Review what the week's data showed. Attack one thing: the misconception, not the symptom. Set what the AI will and will not do this week.
Tuesday to Thursday, independent (20-30 min/day). Targeted practice generated to the right difficulty. Answers withheld; hints available. Everything logged.
Friday, closed-book check (10 min). Redo two problems from Monday with nothing open. This is the honest measure, and it is the mechanism the PNAS students never had.
Ongoing. The human sees the week's data before the next session, so Monday starts from evidence rather than from "how did it go?"
The closed-book check is the part most programmes omit and the part the research most directly supports. Without it, everyone is optimising for the feeling of progress.
Where the guardrails come from
A guardrail is not a technical feature you can buy. It is a decision about what a student should have to do unaided, and it is subject-specific, student-specific and time-specific. A child who has just grasped a method needs more support than the same child three weeks later.
That is why the hybrid model is not simply "an AI tutor plus occasional check-ins". The human is not supervising the child. The human is configuring the system, continuously, based on what they can see and it cannot.
What this costs, honestly
Hybrid is not the cheapest option. It is cheaper than the same number of human hours, because the drill moves to software, and more expensive than a subscription alone.
If budget is genuinely tight, the ranking we would defend is: a free assistant used with good method beats an expensive product used carelessly. Method is worth more than spend, and it costs nothing.
The other lever on cost is frequency. How many live sessions a hybrid programme needs is a separate question, and the honest answer is usually fewer than providers suggest: how much tutoring a child actually needs goes through what the research on dosage shows.
How eGunas is building this, and what is not ready
We should be specific, since this is the page where we describe ourselves.
What exists today: one-to-one teaching with human tutors, aligned to CBSE, ICSE and state boards. That is the whole current product.
What does not exist yet: our AI-assisted learning. It is in development. There is no AI tutor to sign up for, and any page implying otherwise would be wrong.
What we intend: the division of labour above, with the human deciding the constraints. Whether we get it right is a fair thing to judge us on later, against the twelve questions we published for judging everyone else.
If your child needs help this term, the thing we can actually offer is a human tutor and a free first class. If you want to know when the AI side launches, we can tell you when it does rather than when we hope it will.
The conclusion the evidence supports
Not AI instead of teachers. Not teachers refusing AI. A system that supplies volume and patience, and a person deciding what it should withhold, what to attack next, and when the plan needs to change.
That is the arrangement both trials point at. It is less exciting than either headline, and it is what the data actually says.
How eGunas helps your child
eGunas teaches one child at a time. A session follows the student's own board syllabus but is built around what that particular student needs next -- their pace, the topics they are actually weak at -- rather than one lesson delivered to thirty students at once. That is what personalised learning means in practice.
eGunas AI, our AI-powered learning platform, is planned for 2026 and not available yet. It is being designed to extend that same individual attention beyond the lesson hour: practice pitched at the right difficulty, available whenever your child sits down to work, so progress can continue between sessions instead of waiting for the next one.
- Browse our courses -- CBSE, ICSE and state boards, Classes 1 to 10
- How eGunas online tutoring works -- one-to-one lessons online, wherever in India you live
- Book a free demo class -- one-to-one with a tutor, at no cost
- Ask us about your child -- tell us the subject and year, and we will say honestly whether we can help
New articles are published on the eGunas blog.
