A MathEinstein scenario explorer, version 1.0, September 2026
What will one-on-one tutoring look like in 2030?
We don't know yet how AI will reshape tutoring. Will every child get the kind of personal teaching that only some families can afford today? Will human tutors become unnecessary? Neither, or something else? How can we tell?
We teach maths one-on-one online for a living, and for the past two years we've taught with AI tools inside the lesson itself. So we built a small model of how AI might change what a tutor does, how many hours families buy from humans, and what a tutor's hour is worth. Below you can read about three possible futures, then enter your own expectations about AI tutors and see the tutoring economy they imply.
One tutor's thirty-two tasks, and what AI does to them in each future.
Why we're doing this
Tutoring is one of the most studied interventions in education. Across dozens of randomised trials, regular one-on-one tutoring moves a student's attainment by around a third of a standard deviation, which is large by the standards of anything schools do; Benjamin Bloom's famous 1984 experiment claimed two full standard deviations. It's also the most exposed teaching format there is. A tutor works alone, online, with one child, in a subject whose answers can be checked. If AI is going to change any kind of teaching first, it's this one, and maths first of all.
We want the change to be good for children and for the people who teach them. Seeing the possible futures more clearly is the first step to choosing among them.
Here's the short version. In scenarios ranging from business as usual to one where the AI does half the explaining, the hours families buy from human tutors stay inside the range tutors have seen before, and rates hold or rise depending on what the tutor sells. But in the scenario where AI explains better than most humans and most families default to it, hours and prices for plain explanation collapse. In that world far more children get one-on-one teaching than ever before; the hard questions are what the human is for and how the money is shared.
As you scroll, you'll see a lesson broken into its tasks and how AI touches each one under each scenario. Then you can plug in your expectations for how capable AI tutors will be and how widely families use them. The model shows what tutoring might look like in 2030 if you're right, and which of our three scenarios your answers sit closest to.
A lesson is made out of tasks
This model treats a tutor's job as a bundle of tasks. AI can help a person do a task better or faster. It can take the task over. It might not touch it at all. And it can create tasks that didn't exist before.
A tutor's tasks
Think of an hour in the life of an online maths tutor
It's 4.55 pm. She opens the whiteboard five minutes early and rereads last week's note: the student can find a fraction of a quantity but freezes when the fraction is bigger than one. She looks at the homework the parent photographed last night, decides to start on a number line, and drafts three questions that climb from two-thirds of 12 to five-halves of 12. At 5.00 a thirteen-year-old appears from three time zones away, quiet. She spends two minutes on nothing in particular, a football result, a test at school, and only then begins.
For the next fifty minutes she explains, draws, watches the student attempt, interrupts at the right moment, gives a hint that isn't the answer, notices the shoulders drop after a wrong answer and resets the mood, and chooses each next question from the last one. At 5.55 she sets homework. Afterwards she writes the parent a short note, logs what was covered and what comes next, chases a reschedule, and answers a new enquiry from someone in a fourth time zone.
Each of those is a task. Some happen before the lesson, some inside it, some after it, and some are about running a business rather than teaching. We list thirty-two of them here and, as you keep scrolling, colour each one by what AI does to it.
Jobs change over time
Tasks leave the bundle: almost nobody posts worksheets any more, and few tutors still draw every diagram by hand. New tasks arrive: fifteen years ago no one taught over a video call, and two years ago no one screenshotted a whiteboard into an AI to build a practice app before the lesson ended. The bundle isn't fixed, and the job changes as the tasks do.
Only humans can do some tasks
Or, more carefully, only humans can do them in a way families will pay for. The AI can't be the person a thirteen-year-old doesn't want to let down at 5 pm on a Wednesday. It can't tell a parent, with a face they know, that their child is going to be fine. Whether it can keep a child motivated over months rather than minutes is an open question: the early trials find AI tutors engaging, but nobody yet has a year of data on a real child.
Some tasks get augmented
AI helps the tutor do them better or faster: building the diagram, drafting the three climbing questions, spotting from the homework photo which step the student skipped, writing the first draft of the parent note, choosing the next question.
Some tasks may get fully automated
Marking. Scheduling. The practice set. The progress log. The parent note. In some scenarios, the explanation itself, and the diagnosis before it.
And new tasks appear
New technology has always created work as well as removing it. For a tutor, the new work looks like reviewing the AI's diagnosis before the lesson, checking the AI's questions and answers before a child sees them, supervising the student's AI tutor between lessons, and teaching a child how to learn with AI without letting it do the thinking.
The result: the tutor's job changes
As the tutor takes AI into her work, she oversees more and grinds less. The live hour goes to the parts that need a human: the diagnosis, the stuck moment, the confidence, the parent. The machine carries the rest. More learning happens per hour of her effort. Productivity rises.
Switch the scenario at the top of the figure to watch the bundle change.
Every lesson happens millions of times a day
Somewhere between 100 and 180 billion dollars a year is spent on private tutoring worldwide, depending on who is counting. In the last large global census of students, 43% said they had a private tutor, and maths was the most tutored subject by a distance. The big online marketplaces list tens of thousands of maths tutors between them. Every one of the tasks above is being done, right now, somewhere.
From lessons to the tutoring economy
Add up every lesson, every hour of preparation, every parent note, and you get the tutoring economy. How AI reshapes it depends on how it affects each task, how many new tasks it creates, how fast families and tutors adopt it, and how much more a tutor can do with it. Those answers move four numbers we care about: how much one-on-one teaching happens, how much of it humans deliver, what a tutor's hour is worth, and how much of a family's money reaches the tutor.
There are many possible futures, but we're highlighting three
Which future arrives depends on how capable AI tutors become and how families and tutors take them up. Our three scenarios each capture a distinct kind of change. The numbers come from the same model you can play with below; they are illustrations of a direction, not forecasts.
Modest scenario
AI is a better whiteboard
AI's effect on tutoring is about what the internet's was. That was real: video calls and marketplaces turned a neighbourhood service into a global one. Tutors prepare faster and follow up better. Families use AI homework helpers the way they use YouTube, to look something up. The human lesson remains the product. Hours are flat, the better tutors charge a little more, and the platforms stay platforms.
The tutor's thirty-two tasks, plus new ones, in this scenario
Substantial scenario
The AI does half the explaining
By 2030, AI can do about half of a tutor's tasks at the level of a good tutor, mostly without a human in the loop, and most tutoring families use one every week. But not for everything: the live human hour continues, less often. Weekly becomes fortnightly, the tutor supports twice as many students, and the lesson becomes diagnosis, the hard part, and the relationship. Plain explanation earns no more than it does today; coaching and specialist expertise earn more; the mid-market generalist gets squeezed; the platforms bundle AI and take a larger cut.
The tutor's thirty-two tasks, plus new ones, in this scenario
Extreme scenario
The human is optional
AI explains, diagnoses, adapts, draws and, to most families' satisfaction, motivates better than most human tutors, at almost no marginal cost, and families default to it. Human tutoring shrinks to three products: accountability and relationship (the coach), premium human attention (the luxury), and vouching (someone families and schools trust to say what a child can actually do). Human hours fall by about three-quarters. Half of all tutors are still between offers in 2030, with the old product gone and the new one not yet sold. More children get one-on-one teaching than at any time in history. Roughly 40 cents of each family dollar reaches a human.
The tutor's thirty-two tasks, plus new ones, in this scenario
How might powerful AI change tutoring? Make your predictions
How do you think AI tutors will develop over the next four years, and what would that mean for tutoring? Answer five questions. The model updates as you move each slider. It starts on the substantial scenario's answers, and nothing you enter is saved or sent anywhere. The small letters under each slider mark where the modest (M), substantial (S) and extreme (E) scenarios sit.
Your 2030
Five questions in, the numbers out. All numbers are indexed to 2026 = 100 unless marked otherwise.
By 2030, what share of a maths tutor's tasks can an AI do at least as well as a good human tutor?
50%
What share of families who pay for tutoring will use an AI tutor every week?
60%
Of the teaching the AI does, how much happens with no human tutor in the loop?
Most of it
How much more productive does AI make a tutor who uses it well, in students supported per hour of her own effort?
2×
When a tutor's hours dry up, how long does it take to reposition: a new offer, a new niche, or a new job?
About 9 months
What your answers imply for 2030
Teaching delivered: human and AI, with 2026 marked
How the numbers are computed
This is a cartoon model, kept simple on purpose so you can argue with it. Let c = capabilities, a = adoption, u = autonomy (0.05 to 0.95), p = productivity multiplier, t = adjustment time in months.
- Share of human tutoring hours replaced by AI:
r = c × a × u - Human hours (2026 = 100):
100 × (1 − r) - AI hours, quality-weighted:
400 × a × c. The 400 assumes an AI tutor is used about four times as many hours as a weekly human tutor. - Teaching delivered: human hours + AI hours
- Tutors who must reposition:
r. Still between offers in 2030:r × min(1, t ÷ 36), because displacement spreads across four years and the average displaced tutor has about three years to adjust. - Price of a plain explanation hour:
100 × (1 − 0.85 r) - Income per effort-hour for an AI-leveraged coach:
100 × (1 + 0.6 log₂ p) × (1 − 0.4 r) - Students per tutor:
25 × p - Family spending: human tutoring
100 × (1 − r), of which 75% reaches the tutor; AI tutoring25 × a × c, of which none does. The tutor's share is the first divided by the sum.
The model ignores families who never bought tutoring before and start buying human coaching because AI made them ambitious. That is the main reason it may be too pessimistic about human hours.
The line charts in the findings join 2026 to 2030 with a straight line. The model only computes the 2030 end point; the path between is not modelled.
What people expect
Anthropic surveyed more than ten thousand people for their version of this exercise, and found the typical answer sat close to their middle scenario. We haven't run a survey yet. If you're a tutor, a parent or a student, your answers above are the first row of our dataset: copy your prediction and send it to us, or to your tutor group, and the next version of this page will show what people actually expect.
Where your answers sit
Each line runs from the lowest to the highest possible answer. Grey dots are our three scenarios, light to dark: modest, substantial, extreme.
You predicted one possible future for tutoring. In each finding below, it appears as Your 2030, next to our three.
Finding 1: teaching deliveredOne-on-one teaching grows in every scenario
AI increases the amount of personal teaching that happens in every scenario, and the differences between scenarios are large. Two things drive it. An AI tutor is used daily where a human tutor is used weekly, so every family that adopts one gets far more hours of attention than they bought before. And the AI reaches the majority of students who have never had a tutor at all, because the marginal cost is close to zero.
One-on-one teaching in 2030, by scenario
Hours of personal teaching, 2026 = 100, marked by the black line. AI hours are weighted by how capable the AI is in each scenario.
But the amount of teaching isn't the only thing we care about. What would each future mean for how much of that teaching is done by a person, for how many tutors have to find something new to sell, and for how much of the money reaches them? This model isn't a full map of reality, but it shows a pattern: the amount of learning grows in every scenario, while the human share of it, and the human share of the money, can shrink even as the total grows.
Finding 2: tutor reallocationIn more transformative scenarios, more tutors have to change what they sell. That can mean a long, unpaid transition.
There is always churn in tutoring. Students finish exams, families move, a tutor's evenings fill up and empty again. In normal times the churn is painful for the individual and invisible in the aggregate; most tutors refill their roster within a few months.
In our substantial and extreme scenarios, plain explanation, the thing most tutors sell, is largely automated. At the individual level, a tutor who sells "I'll explain quadratics to your daughter" has to become a tutor who sells "I'll make sure your daughter learns quadratics this term, I'll be the one checking what the AI told her, and I'll be the person she doesn't want to disappoint on Wednesday." That is a different product, sold to a different buyer, at a different price.
Where tutors are in 2030
Every tutor working in 2026, followed to 2030. Pick a scenario.
Switching is hard for three reasons. Tutors may not want to change; the explaining was the part they loved. They need skills they weren't hired for: diagnosis, system design, parent communication. And even with the skills, finding the new customers takes time. The longer that takes, the more tutors are in limbo in 2030.
Finding 3: ratesThe price of a plain explanation falls. The income of a tutor who leverages AI rises. The gap between them widens.
This happens for the same reason wages diverge in any automation wave. When AI can explain a topic well for nearly nothing, families pay less for a human to do the same, and the price of an explanation hour falls. At the same time, a tutor who runs AI between lessons for fifty or a hundred students earns more per hour of her own effort, even if she earns less per student. The same 2030 is a bad year for one tutor and the best year yet for another, and which one you are depends on what you sell, not on how well you explain.
What a tutor's hour is worth in 2030
2026 = 100. Two tutors, same year, opposite experiences. Faint lines are the other scenarios.
Finding 4: who gets the moneyThe pie grows, but a larger share may go to platforms and AI providers
Today, on the big marketplaces, roughly three-quarters of what a family pays reaches the tutor; the rest goes to the platform. If the AI layer that does more and more of the teaching is owned by someone other than the tutor, a growing share of the family's money goes to whoever owns it. That can happen even while individual tutors earn more.
In the modest scenario the tutor's share barely moves. In the substantial one it slips to around two-thirds. In the extreme one, about 40 cents of each dollar reaches a human, and total human tutoring income is a fraction of today's while the amount of teaching delivered is several times larger. The challenge in that world isn't delivering learning. It's making sure the people who teach still get paid for it.
How each family dollar is shared
Marketplace tutoring. One dot per cent.
The future is not predetermined
What tutoring looks like in 2030 depends on things nobody has settled yet.
- Whether AI tutoring produces learning that lasts. The best-known trial so far, in a Harvard physics course, found students learned more than twice as much in less time with an AI tutor than in a well-run class. Another, in Turkish high schools, found students with unrestricted AI did far better on practice and worse on the exam once the AI was taken away, and that a version built with tutor-like guardrails removed most of the harm. A six-week World Bank pilot in Nigeria found gains of about 0.3 standard deviations. All of these are weeks long. Nobody has followed a child for a year.
- How schools and exams respond. If assessment moves toward supervised, in-person work, the human who can vouch for a child becomes more valuable, not less.
- Policy. In 2021 China banned for-profit tutoring in core school subjects almost overnight, and the listed tutoring companies lost most of their value within months. Governments can reshape this market faster than technology can.
- What families choose when the AI is good and the human is expensive. Nobody knows yet. The extreme scenario assumes they choose the AI; it may be that a good AI makes families more ambitious and they buy more human coaching, not less.
- What tutors choose. The one variable in this model that a tutor controls is the adjustment time: how fast they change what they sell.
What the model suggests a tutor can do now
Every scenario rewards the same moves. The extreme one merely punishes not making them.
- Move your hours toward the tasks that stay plain in every scenario: diagnosis, the stuck moment, confidence, and the parent.
- Sell an outcome and a relationship, not explanation by the hour. "Your child will pass this exam and I'm accountable for it" survives every scenario. "I explain topics for $25 an hour" survives one.
- Own the between-lesson layer instead of renting it. A tutor whose students learn on the platform's AI is a supplier to the platform. A tutor whose students learn on her own system is the product.
- Become the verifier. In every scenario with AI in the loop, someone has to check what it told the child. Parents will pay the person they trust to do that.
- Keep the parent relationship direct. Parents are the buyers, and a tutor they never hear from is the easiest thing in the bundle to replace.
Limits of this model
This is version 1.0, and it is a cartoon of a complicated thing. It isolates a few forces and leaves out many that may matter more.
- It leaves out policy, school-system responses and exam reform, any of which can move the market faster than the technology.
- It treats tasks as separable. In a live lesson they're braided: the diagnosis happens inside the explanation, the confidence work happens inside the hint. Automating "the explanation" may be harder, or easier, than the figure implies.
- It doesn't follow individual tutors, so it can only paint a coarse picture of what displacement costs a person.
- It assumes learning gains from AI tutors are durable, which is contested, and it assumes families judge tutoring by learning rather than by reassurance, which is doubtful.
- It assumes marketplaces persist in something like their current form, and it ignores tutors who go direct.
- It ignores in-person tutoring, physical presence, and the countries where tutoring is a system rather than a service. It is calibrated on online, one-on-one, maths work, which is the most exposed slice of the whole market.
- Several of its parameters are guesses. The formulas are shown above so you can disagree with the specific ones.
The extreme scenario is better read as a thought experiment than a prediction, and the modest one may understate what's already visible in our own lessons. Actual outcomes may differ materially. If you think we've coloured a task wrong, tell us; that's the most useful kind of feedback we can get.
Written by Varun at MathEinstein with an AI research assistant, September 2026. The task-bundle framework and the shape of this piece are adapted from Anthropic's Scenarios for our Economic Future (Korinek, Jones, Sacher, Cotter and McCrory, 2026), which models the whole US economy the way this page models one job. Tutoring effect sizes: Nickow, Oreopoulos and Quan (2020); Bloom (1984). Census figures: Cambridge International Global Education Census (2018). AI tutoring trials: Kestin et al. (2025); Bastani et al. (2024); World Bank (2025). Market-size range: published industry estimates, which disagree with each other by nearly a factor of two.