What EdTech Companies Are Actually Selling

In 2021, EdTech raised somewhere north of $20 billion in venture capital. Third-largest category globally that year, behind fintech and healthcare. BYJU’S was valued at $22 billion. Coursera went public at $5.1 billion. Duolingo priced at $5.6 billion.

The theory was that the pandemic had permanently broken education, students wouldn’t return to classrooms, online learning had proved itself, and the market was effectively everyone.

By 2024 BYJU’S was in bankruptcy proceedings after accounting irregularities, missed payrolls and a legal fight with its lenders. Coursera’s stock had lost the large majority of its IPO value. The boom was over, and it had not produced the transformation it sold.

Now AI EdTech is here with almost identical promises, and almost nobody is asking why the last round failed so completely.

To understand the pattern you have to understand who actually buys this stuff. In most districts the procurement decision sits with administrators, not teachers and certainly not students. And the administrator evaluating a product is not primarily asking whether it improves learning. They’re asking whether they can defend the purchase to a school board, a superintendent and a room full of parents.

Those are different questions with different answers.

A product that improves learning in ways that are hard to measure, deeper engagement or better long-term retention, is much harder to defend than one that emits usage dashboards and completion certificates. EdTech companies worked this out early and built accordingly.

The first generation, from the CD-ROM era through the early 2000s, was optimised for activity rather than learning. You clicked through lessons. The system logged that you had clicked through lessons. The report said students had engaged with the content. Whether anyone learned anything depended entirely on factors the product had no access to.

The MOOC era made the same mistake at a far larger scale. Coursera, edX and Udacity all launched with genuine idealism about opening up Stanford and MIT lectures to everyone. Then they looked at the data and found completion rates in the single digits to low teens.

Worse, the people completing courses were mostly already educated, already motivated, already partly familiar with the material. The students the democratisation framing was written about, first-generation applicants, workers retraining, anyone in an under-resourced school system, dropped out at rates that made the framing embarrassing.

Sebastian Thrun said the quiet part out loud in 2013, calling what Udacity had built a lousy product after reading their own completion data. He was right and the honesty was unusual. The rest of the sector kept raising on optimistic projections while holding the same numbers. The dropout rates were not the slide anyone led with.

MOOCs didn’t fail because the content was bad. MIT’s courses are excellent. They failed because learning isn’t primarily an access problem, and if it were, Wikipedia would have replaced school two decades ago. Learning needs feedback, struggle, social accountability, someone who knows more than you pushing back on your reasoning, and retrieval practice spread out over time. A recorded lecture supplies none of that. Lectures are the weakest part of education when delivered in a room; stripped of the room, they’re weaker still.

AI EdTech is running the same loop with better production values. The 2024 and 2025 decks show tutors that know every student’s learning style, adapt in real time, never lose patience and never sleep. The vision is genuinely compelling. The implementations aren’t matching it yet.

Khanmigo is the fairest test case, because Khan Academy is not a cynical operator and the product had serious money and serious pedagogical intent behind it. Teachers and students who used it heavily reported something consistent: better than a static textbook, frustratingly inconsistent in practice. It would give a careful Socratic hint on one problem and hand over the answer on the next near-identical one. It would explain a concept cleanly and then reach for different vocabulary in the follow-up, confusing rather than reinforcing.

For a student who knew how to work it, how to ask a second question, how to push back when an explanation wasn’t landing, how to notice confident nonsense, it was a real resource. For a student who just wanted the homework finished, it was arguably worse than the textbook, because it was engaging enough to feel like learning while functioning as an answer machine.

That gap isn’t a user-education problem you can fix with an onboarding video. It points at something the decks never mention: the benefit of an AI tutor scales with the sophistication of the user. The students who need the most help are exactly the students least equipped to extract it from a system that rewards knowing how to ask.

The venture model is also a bad fit for this sector specifically, and that explains more of the failure than any individual product decision.

Education runs on long timescales; the learning that matters shows up over years. Its beneficiaries are diffuse, because the entity that gains most from an educated population is society, which cannot be charged a subscription. And the variance across contexts is enormous. An intervention that works for a suburban student in Massachusetts can fail completely for a student in rural Mississippi, because teacher skill, family support, prior knowledge and social environment differ so much that no product abstracts over them.

Venture capital needs outsized returns inside seven years. Education doesn’t operate on that clock. The way you square it is to sell access instead of outcomes, charge for platform seats and completions and certificates, and report activity instead of learning. Which is precisely what three generations of EdTech did, and precisely why they underdelivered on value while successfully collecting revenue.

The current wave inherits the structure intact. The companies charging monthly for a student seat are measuring daily actives, lessons completed and retention. They’re not measuring whether a student who uses the product for a year learns more than a comparable student who doesn’t, because that study is slow, expensive, and might produce a number nobody can control. So nobody funds it.

Some of this has worked, and the cases that worked have something in common worth stealing.

Duolingo, before it started chasing engagement metrics hard, produced measurable outcomes in independent studies. The frequently cited result put around 34 hours of Spanish on the app roughly level with a university semester for reading and listening proficiency. That’s a real finding, and it happened because the pedagogy underneath was sound: spaced repetition implemented rigorously, failure made visible instead of smoothed over, exercises demanding active production rather than passive recognition.

None of which was ever the marketing. The marketing was streaks, leaderboards and a cartoon owl. The learning worked because of the boring parts.

The AI EdTech companies that do eventually work will solve the same problem, and it isn’t making learning feel less like work. Learning is effortful by definition, and a tool that reduces effort is not automatically a tool that improves learning. What works is directing effort toward the right difficulty at the right level with feedback that builds something. Most current products are much closer to the gamification model than the cognitive science one, and from the outside the two decks look almost identical.

There’s one more structural problem that AI doesn’t fix and may sharpen: the innovation happens in the private market while the students who most need better tools sit in public institutions that can’t pay private-market prices.

A monthly tutor subscription is a rounding error for a household in the top income quintile and a genuine decision for one in the bottom, especially with three children. The students with the least effective traditional instruction are the least likely to get the supplement, which means the market dynamics don’t merely fail to close the gap. They widen it.

Foundation grants and state programmes push against this and the sums are not close. A ten-million-dollar grant sounds enormous until you set it beside the several hundred billion American K-12 spends every year. Philanthropy in this sector is a rounding error on the system it’s trying to move.

So the companies worth watching are the ones building a sustainable way to reach the students who need it rather than the students whose parents will pay. The ones optimising for the profitable segment while claiming a broad social mission are following a script that is, by now, completely predictable. Look at where the current money is going, mostly consumer tutoring aimed at university students and comfortable families, and you can guess which kind this wave produces.

That isn’t an argument for pessimism about AI in education. It’s an argument for being clear about what a market will and won’t produce unprompted, and for being specific about what public money would have to do instead.