The AI Tutoring Business Model: Costs, Evidence and Moats

The AI tutoring business model works when a company can deliver tutoring that changes learning, at an inference cost below what a school, parent or learner will pay, through a channel the model providers cannot copy for free. Most AI tutoring products in 2026 meet one or two of those conditions, not all three. The most proven business so far is not the student-facing tutor but AI that saves teachers time or supports human tutors.

The economic case starts from a famous gap. Tutoring is the best-evidenced intervention in education: a meta-analysis of 96 randomized trials of PreK–12 tutoring found a pooled effect of about 0.37 standard deviations. But high-dosage human tutoring costs a reported $3,500–4,300 per student per year in the original Chicago model. Large language models promise to close that gap. Since March 2023, when Khan Academy launched Khanmigo and Duolingo launched Max on GPT-4, the question has been whether that promise survives contact with costs, evidence and competition.

What jobs does AI do in education?

The book separates four jobs that marketing often blurs, each with its own buyer and evidence standard.

The distinction matters for the business model. A teacher copilot that saves time can be valuable even if test scores do not move. A student-facing tutor competes for scarce instructional minutes and can easily replace the student's thinking instead of supporting it, so it has to show learning.

What does an AI tutor cost to run?

Classic intelligent tutoring systems were limited by authoring cost: commonly cited estimates ran 100–300 hours of development per hour of instruction. Large language models inverted the problem. Authoring is now cheap. Inference, the compute each conversation consumes, is a cost of goods that rises with every session.

The research behind the book works through the arithmetic with illustrative prices. A moderately active student, 20 sessions a month of 15 turns each, costs:

Set that against the revenue side. US K-12 supplemental software often sells for about $5–30 per student per year. A product that routes every conversation through a frontier model at district prices loses money on each active student. The book's conclusion is that model routing, caching and small models "are therefore not optimisations but survival." Per-token prices have fallen fast since 2023, but a student three times as active still costs three times as much.

Consumer tiers face the same squeeze. Duolingo cited AI costs as one reason its 2025 gross margin slipped to 72.2% from 72.8%. Khanmigo for individual learners was priced at about $4 a month or $44 a year in 2024.

Does AI tutoring work? What the evidence shows

The evidence from 2024–26 points both ways, and design decides the direction.

The pattern is consistent: AI that gives answers can replace the effort that produces learning, while AI designed as a tutor, or as support for a human tutor, can improve it. What is still missing, as of 2026, is large, pre-registered, multi-semester trials of the major products on standardized outcomes. Expect early results to shrink at scale, as they did for classic tutoring systems. The book suggests a credible effect of 0.1–0.2 standard deviations at low cost per student would be a major win. Our guide to evaluating edtech efficacy claims explains how to read these studies.

Usage is the other constraint. Khan Academy reported about 2.0 million users with access to Khanmigo in the 2024–25 school year, but its own staff acknowledged that many students did not use it unprompted. Access is not use, and use is not learning.

Who is making money from AI in education?

The 2023–26 market has sorted into layers.

Model providers entered education directly: OpenAI with ChatGPT Edu (2024) and study mode (July 2025), Anthropic with Claude for Education and a learning mode (April 2025), and Google with Guided Learning (2025). Their free or discounted plans are the largest threat to thin "wrapper" startups.

AI-native teacher tools grew fastest. MagicSchool AI, started in Denver in 2023 by a former teacher, reached about 1 million educators within about five months of launch (company claim) and raised about $62–63 million by February 2025. It described revenue as "eight figures" in early 2025 (CEO, reported). It follows the classic teacher-first ladder: free sign-up, then school and district licenses.

AI-native tutors in narrow domains raised at high valuations. Speak, a speaking tutor for language learners, raised $78 million at a $1 billion valuation in December 2024.

Incumbents embedded AI to defend renewals. Nerdy, the owner of Varsity Tutors, rebuilt its platform as "Live+AI" in 2025, combining human tutors with AI tools; its 2025 revenue was $179.0 million.

The disrupted were businesses that sold answers. Chegg lost about half its revenue in two years. Our case study on what happened to Chegg explains why.

What makes an AI education business defensible?

The book sets out six rules that follow from the costs and evidence above.

1. Ride the model curve; do not fight it. Each model generation makes generic tutoring cheaper and better. A product that is mainly a prompt over a general model will be matched by the provider's own free education mode.
2. Own something the next model release does not provide. Distribution into schools, curriculum alignment, proprietary learning data, teacher trust, workflow integration and evidence are durable. Model access is not.
3. Design for thinking, not answers. Withhold final answers, ask the learner to attempt the next step, and test them without the AI.
4. Start with teacher augmentation. The clearest early evidence and adoption are in teacher time savings and support for human tutors.
5. Price against inference; sell against outcomes. Know the cost per active student per month by model tier.
6. Report active learners, not accounts. "Millions of users" usually counts accounts with access.

The anatomy of a credible AI tutor

A credible tutor in 2026 combines several models routed by task and cost; a pedagogical policy that withholds answers and checks understanding; grounding in the curriculum the class actually uses; verification through math engines, code execution or a second model; a learner model and memory; safety layers including crisis escalation; transcripts and flags for teachers; and evaluation that asks whether the student learned without the tutor afterward. Every one of those layers adds cost, which is why the product decision and the business model are the same decision.

What policy risks does AI tutoring face?

Most district contracts now forbid training models on student data. Consumer chatbots generally require users to be at least 13, and the FTC's amended COPPA Rule, finalized in January 2025, tightened consent and data-retention requirements for children's services. The EU AI Act classes AI used to evaluate learning or steer students as high-risk.

South Korea is the warning case. AI digital textbooks were rolled out as official textbooks in March 2025, fewer than 30% of elementary schools used them, and in August 2025 the National Assembly downgraded them to optional supplements after complaints about errors, privacy, screen time and teacher workload (reported). A national mandate can reverse faster than a vendor can recover its investment.

Which factor of the learning equation decides AI tutoring?

AI changes the cost side of the learning equation without settling the rest. It makes content and conversation cheap, which helps the first factor, learning design, only if the design forces students to think. It does not create a payer: districts still budget per student per year, and consumers can get a general chatbot free. It does not create a cheap route to the payer, which still runs through teachers, districts and app stores. And it makes the fourth factor, evidence, more urgent, because a product that can make learning worse needs proof that it does not. The companies most likely to win are those that combine a model with something it cannot supply: a channel, a curriculum and a study.

The takeaway

AI tutoring has a real business case because it can deliver some of what human tutoring does at a fraction of the cost. But inference costs scale with use, free model-provider tutors cap what customers will pay, and the evidence shows that unguided AI can harm learning while well-designed AI helps. The near-term money is in teacher tools and AI support for human tutors; the long-term prize goes to whoever proves learning at a price districts can renew. The wider economics sit in our edtech business model guide, and Chapter 17 of The Learning Equation covers AI in education in full.