EdTech: what is actually working now?
In our EdTech market deck, you will find everything you need to understand the market
SUMMARY
EdTech is actually working now, but the clearest wins come from products that improve learning routines, teacher productivity, tutoring quality, feedback, or career outcomes rather than simply adding more AI.
AI tutoring can already teach well under controlled conditions. The harder problem is getting students to use that help deeply and repeatedly in ordinary classrooms, which is why school-wide gains remain useful but modest.
The strongest AI learning systems still make students do cognitive work. Guided questions, hints, retries and structured practice consistently look more valuable than giving a polished answer on demand.
Teacher-facing AI may be moving faster than student-facing AI because teachers can judge weak output and use the technology on repetitive work. Reported time savings are large enough that preparation and administration already look like one of EdTech's clearest near-term wins.
AI may be especially valuable when it upgrades an average human tutor instead of trying to replace the tutor. Tutor CoPilot's strongest gains appeared among weaker tutors, suggesting that cheap expert guidance can raise the floor of human instruction.
Traditional tutoring still sets a high bar. Human tutoring produces larger average academic effects than the school-wide AI tutoring evidence we have today, but attendance and cost remain major constraints that technology can help reduce.
Feedback is another area where AI is becoming genuinely useful. The evidence is strongest when AI gives fast, specific comments and the student still has to revise the work; the learning value gets much murkier once the model starts doing the rewrite.
Structured practice remains surprisingly hard to beat. Duolingo and Khan Academy both point to the same underlying advantage: deciding what the learner should do next, making repetition easy, and giving people a reason to come back matters more than a flashy conversation layer.
Consumer and career EdTech are working for a related reason: the value is obvious to the learner. Duolingo has built a daily habit at enormous scale, while Stride, Coursera and Udemy are seeing strong demand where learning is tied to a concrete skill, credential or employment goal.
The weakest models are the ones whose main value can be copied by a blank chatbot. Chegg's collapse shows how quickly paid answer access can lose its moat, while school procurement data shows that thousands of available tools mean very little if teachers and students actually use only a tiny core.
The broad pattern is simple: explanations, quizzes and content generation are getting cheap. The durable value in EdTech now sits in sequencing, practice, judgment, accountability, integration, habit formation and outcomes that people can see.
This market map, featured in our EdTech market deck, highlights top companies and startups in the EdTech market
Why are we asking whether EdTech actually works again?
EdTech is back under scrutiny because the pandemic boom collapsed, while AI has suddenly made digital teaching much more capable.
The first EdTech boom proved that education could move online quickly. It also created a lot of weak businesses. Global venture funding surged above $15 billion a year during the pandemic-era peak, then fell by roughly 85% from those highs as schools stopped buying software in emergency mode and investors became much harder to impress.
That correction changed what counts as success. A company can have millions of accounts, thousands of school customers or impressive AI demos and still make very little difference to learning.
AI has now reopened the whole debate. Explanations, practice questions, translations, lesson plans and basic feedback can be generated almost instantly. Students have access to something that looks remarkably close to a private tutor. Teachers can produce hours of preparation work in minutes.
So we now have a much better test of EdTech than we had during the pandemic: when access to technology is easy, which products actually make education better?
That question produces a surprisingly narrow set of winners.
What should count as EdTech “working” today?
EdTech is really working today when it improves learning, saves educators serious time, creates a learning habit people keep, or connects education to an outcome people are willing to pay for.
Downloads tell us very little. Registered users are only slightly more useful. Even school purchases can be misleading because schools routinely pay for software that barely gets opened.
For student products, we want evidence that knowledge, skills, retention or progression improve. For teacher products, several hours of saved work can be a real result even before test scores move. Consumer products also need repeated voluntary use. Career platforms have another useful test: does the course help somebody gain a skill, credential or job outcome valuable enough to justify the time and money?
Those tests produce very different results across EdTech.
Duolingo is proving that a structured learning habit can support tens of millions of daily users. Tutor CoPilot has shown that AI can improve the performance of human tutors. Khanmigo has produced measurable gains inside real middle schools, although they remain modest. Chegg has shown how quickly a once-successful model can fall apart when its main value becomes available elsewhere for free.
“EdTech” therefore covers businesses with radically different levels of actual educational value.
| What we are measuring | A useful sign that EdTech is working |
|---|---|
| Student learning | Better mastery, retention, scores or skills |
| Teacher productivity | Hours of real work removed or improved |
| Consumer learning | Frequent voluntary practice over time |
| Career learning | Skills, credentials or employment value |
| School software | Repeated use after procurement |
| Business durability | Customers keep paying even when cheaper alternatives exist |
As this chart shows, and as featured in our EdTech market deck, online search interest in online learning has grown significantly
Are AI tutors actually helping students learn now?
AI tutors are producing real learning gains now, although the best large-scale evidence is much less spectacular than the “personal tutor for everyone” pitch.
One of the strongest current tests comes from an NBER study released in August 2026. Researchers Philip Oreopoulos and Nina Low ran a two-year randomized experiment across 18 Tennessee middle schools using Khan Academy and its AI tutor, Khanmigo.
Students assigned to Khanmigo gained about 1.3 national percentile ranks in mathematics per term. Across a school year, that works out to roughly 0.06–0.08 standard deviations. Students who actively participated for the equivalent of a full year showed an estimated gain of around 0.14 standard deviations.
Those are useful gains. They are also close to what previous studies have found from structured Khan Academy practice without the AI layer.
Usage explains a lot of the gap between the hype and the outcome. Some 96% of students tried Khanmigo, yet the median student messaged it on only around one-third of practice days. Even after making a mistake, students opened a substantive exchange with the tutor in just 17% of exercise sessions.
Smaller experiments show that the technology itself can do much better. A randomized Harvard physics study involving 194 undergraduates found that a carefully designed AI tutor produced more learning in less time than an active-learning classroom condition.
Put those findings together and the bottleneck becomes obvious. We can already build AI tutoring interactions that teach well. Getting ordinary students to use them deeply and repeatedly inside ordinary schools is proving much harder.
Does AI tutoring work better when students still have to think?
AI tutoring works best when the student has to reason, answer, explain and correct mistakes instead of simply requesting the finished solution.
The strongest systems increasingly share this design.
Khanmigo is configured to guide students toward an answer. Stanford's Tutor CoPilot pushes tutors toward probing questions. Google's LearnLM experiments have focused heavily on guided instruction. The Harvard physics tutor was built around learning-science principles rather than an unrestricted chatbot conversation.
The reason is simple. Learning requires effort at exactly the moments AI is very good at removing it.
A student who struggles to retrieve an answer, explains a misconception and tries again is doing useful cognitive work. A student who pastes the assignment into a chatbot and receives an elegant response may finish faster while learning less.
We can see the same pattern in older intelligent tutoring research. A recent meta-analysis covering 30 studies found positive overall effects from intelligent tutoring systems, with worked examples among the features associated with significant gains.
The new language models have made tutoring conversations far more natural. The educational logic underneath them remains familiar: ask, attempt, diagnose, hint, retry.
That constraint looks increasingly important. The best AI tutor knows when giving less help will teach more.
This chart, featured in our EdTech market deck, shows annual VC investment in EdTech startups
Are students using AI to learn, or mainly to avoid doing the work?
Students are using AI for both, and the scale of AI use has moved much faster than universities' ability to redesign learning around it.
HEPI's 2026 survey of 1,054 UK undergraduates found that 95% used AI in some form and 94% used generative AI to help with assessed work. Two years earlier, only 53% reported using generative AI to help prepare assessments.
That is an enormous behavioural shift in a very short period.
Much of the usage can genuinely help learning. Students use AI to explain concepts, summarise material, generate ideas and get immediate help when they are stuck. In the same HEPI study, 49% said AI had improved their student experience.
The uncomfortable part is growing too. Twelve percent said they had directly included AI-generated text in assessed work, compared with 3% two years earlier.
Universities are already reacting. HEPI found that 65% of students felt assessment had changed significantly because of AI.
We should expect that change to continue. Take-home essays and generic written assignments are becoming much weaker evidence that a student can independently reason, research or write. Oral defence, supervised work, project evidence and assessment of the student's process become more useful when generating polished text is trivial.
AI has solved the access-to-help problem faster than education has solved the what-did-the-student-actually-learn problem.
If you want more recent data on this point, please see our latest EdTech market report.
Is AI more useful for teachers than for students right now?
Teacher AI is currently one of the clearest parts of EdTech because the productivity gain appears large and immediate.
A Gallup-Walton Family Foundation survey of 2,232 U.S. public-school teachers found that 60% had used AI for work during the 2024–25 school year and 32% were using it at least weekly.
Frequent users estimated saving 5.9 hours per week.
Across Gallup's average 37.4-week school year, that comes to roughly 221 hours. The researchers described it as the equivalent of almost six working weeks.
Teachers were using AI mainly for practical jobs: preparing lessons, creating worksheets, adapting materials, doing administrative work and building assessments. Depending on the task, 60% to 84% of users said AI saved them time. Very few said the task took longer.
The figure is self-reported, so we should not pretend that every teacher literally gained 221 perfectly measurable hours. Even with that caveat, the reported effect is large.
There is also a useful reason teacher AI may work sooner than student AI. Teachers already know what a good lesson, worksheet or explanation should look like. They can reject weak output, repair mistakes and decide what fits their class.
Students often need the technology precisely because they lack that judgment.
The clearest short-term EdTech win from generative AI may be fairly mundane: removing preparation and administrative work from teachers.
This chart, featured in our EdTech market deck, shows why Duolingo is winning in EdTech
Can AI make an average tutor much better?
AI can already make weaker tutors noticeably better, which may be more valuable than building a fully autonomous tutor.
Stanford's Tutor CoPilot is one of the strongest examples we found.
In a randomized trial involving more than 700 tutors and around 1,000 students from underserved communities, students whose tutors could use Tutor CoPilot were four percentage points more likely to master their mathematics topic.
The biggest improvement appeared among weaker tutors. Students assigned to lower-rated tutors gained as much as nine percentage points in topic mastery when the tutor received AI assistance.
Researchers also analysed more than 350,000 tutor messages. Tutors using the AI were more likely to ask students to explain their thinking and less likely to rely on generic encouragement.
The software cost was estimated at roughly $20 per tutor per year.
A separate Stanford analysis of LearnLM gives us another useful comparison. Students tutored with LearnLM achieved a 66% success rate on later, harder topics, compared with 61% for students working with human tutors and 56% for those receiving static hints. Human supervisors approved 76.4% of LearnLM responses with little or no editing.
Those two experiments point in the same direction. AI can cheaply distribute some of the behaviours of strong tutors, and the biggest payoff may come from raising the performance of ordinary instructors rather than chasing a completely teacher-free classroom.
If you want more recent data on this point, please see our latest EdTech market report.
Is human tutoring still better than most EdTech?
High-quality tutoring is still one of the strongest educational interventions we know, and the most useful EdTech often helps schools deliver tutoring more consistently or cheaply.
A large review of 89 randomized tutoring studies found an average effect of roughly 0.29 standard deviations on academic achievement. Effects were around 0.29 in literacy and 0.27 in mathematics.
That is substantially larger than the school-wide Khanmigo effect we currently have.
Frequency is part of the explanation. Strong tutoring works because a student repeatedly receives attention, correction and accountability. Software can reproduce parts of that process, though participation remains a problem.
Recent virtual-tutoring research makes this painfully clear. In one NBER experiment involving struggling Grade 4–8 students in Toronto, fewer than half of students initially attended the virtual tutoring they were offered. Changes to enrollment later pushed initial participation above 80%, yet weekly attendance still remained below 50%.
Other programs show positive results when students actually receive the intervention. A recent randomized virtual reading-tutoring trial involving 1,550 Kansas City elementary pupils found a gain of about 0.08 standard deviations among children who started furthest below grade level.
So the hard question for EdTech is increasingly practical: can technology make effective tutoring cheap enough and convenient enough that students receive it frequently?
A brilliant tutor that nobody opens does very little.
This chart, featured in our EdTech market deck, shows annual funding in EdTech startups
Is AI feedback on student work actually good now?
AI feedback is good enough to improve student work today, especially when the learner has to act on the feedback rather than letting AI rewrite the assignment.
The average effect across studies is becoming meaningful.
A recent meta-analysis covering 19 studies and 11,875 students found a positive effect from AI-assisted feedback in second-language writing, with a Cohen's d of 0.27. Another review of algorithmic writing feedback across 33 studies found an overall effect around 0.36.
Individual experiments reveal where the value comes from.
A randomized study involving 459 secondary-school students found that theory-based GPT feedback improved text revision by around 0.19 standard deviations and motivation by 0.36. Another experiment with 269 trainee teachers found that adaptive ChatGPT feedback improved the quality of students' written justifications.
The result is less consistent when the task demands expert disciplinary judgment. In one higher-education experiment involving 90 students, teacher feedback produced the strongest gains in scientific argumentation while LLM feedback produced the smallest.
So AI feedback already looks useful for rapid correction, suggestions, language work and repeated revision. Teachers still have a clear edge when feedback requires subtle judgment about the quality of an argument or deep subject understanding.
The dividing line is also educationally sensible. Feedback helps when it sends the work back to the learner. Once AI starts doing the rewrite itself, we are measuring the model's performance more than the student's.
| Type of AI feedback | What the evidence currently suggests |
|---|---|
| Quick corrective feedback | Useful |
| Feedback followed by student revision | Increasingly convincing |
| High-volume personalized comments | Strong practical use case |
| Complex disciplinary judgment | Human feedback still has an edge |
| AI producing the finished response | Much weaker as a learning activity |
Is boring practice software still beating flashy AI?
Structured practice remains one of EdTech's safest bets because learning still depends heavily on repetition, retrieval, difficulty and timely feedback.
This sounds almost too simple next to generative AI, yet the newest evidence keeps pointing back to it.
The large Khanmigo trial produced useful mathematics gains, while the researchers found those gains similar to previous Khan Academy results without generative AI. The extra conversational capability was available to almost every student; students simply did not use it very much.
That tells us something important about personalization.
A chatbot can instantly produce a custom explanation. A mature learning system needs to know something harder: which skill the learner should practise next, which old skill is about to be forgotten, whether a wrong answer is a careless mistake or a misconception, and when assistance should stop.
Duolingo works from the same basic logic. Learners move through a structured path, practise repeatedly, receive immediate correction and encounter material again over time. AI can now create more exercises and richer conversations around that loop, while the loop itself remains the product.
This is why the strongest EdTech often feels less magical than the demos.
The difficult work sits in deciding what the learner should do next.
If you want more recent data on this point, please see our latest EdTech market report.
This chart, featured in our EdTech market deck, compares the main business model options for online course platforms
Is Duolingo proving that consumer EdTech can really work?
Duolingo is currently the clearest proof that consumer EdTech can become a huge daily habit rather than an app people download and forget.
Its latest reported numbers are difficult to explain away. Duolingo reached 58.7 million daily active users in the second quarter of 2026, up 23% from a year earlier. Monthly active users reached 140.6 million, while paid subscribers increased 17% to 12.7 million.
Revenue for the quarter reached $298.5 million, up 18%.
The most interesting number may be the relationship between daily and monthly usage. Roughly 42% of Duolingo's monthly users were using the product on an average day. A year earlier, that ratio was about 37%.
That is an unusually intense habit for an education product.
The model works because several pieces reinforce each other. Most learning is free, so users can form a habit before paying. Sessions are short. Streaks create an obvious reason to return tomorrow. The curriculum decides what comes next. Errors produce immediate feedback. Progress is visible.
AI is now accelerating the machinery behind that experience.
Duolingo previously said AI-assisted production had increased the number of curriculum skills it could create dramatically, and it is expanding conversational features such as Video Call. The important part is where that AI sits: inside a product that already knows how to make people practise.
Duolingo's latest growth gives us a fairly sharp lesson. Consumer learning can work at enormous scale when the product wins the daily habit first.
Are online courses still useful now that AI can teach almost anything?
Online courses are still working, especially for people trying to gain a specific professional skill or credential, and AI appears to be increasing demand for some courses rather than killing them.
Coursera's current business looks very different from the original MOOC experiment.
Coursera completed its combination with Udemy in May 2026. In the first full reported quarter after the transaction, the combined company generated $298.6 million of revenue. More than 85% of revenue now comes from recurring subscription streams across consumer and enterprise products.
Across Coursera and Udemy, the platforms have reached more than 300 million registered learners and more than 12,000 enterprise customers.
The clearest thing to watch is what people are choosing to learn.
Management said the two platforms now offer almost 10,000 generative-AI courses. Learners were enrolling in generative-AI content at a combined rate above 45 enrollments per minute in 2026, compared with roughly 25 per minute the previous year.
That is an 80% increase in the enrollment rate.
People with access to ChatGPT are therefore still paying for organized instruction about AI. They want sequences, expert selection, exercises, credentials and a clear path through a subject that is changing quickly.
The same logic applies beyond AI. A professional preparing for an AWS certification or learning Python for work has a concrete reason to finish. Casual enrollment in a broad course carries much less pressure.
Online courses remain useful. The strongest demand these days is concentrated where the learner knows exactly why the skill is worth acquiring.
This chart, featured in our EdTech market deck, shows how revenue is distributed across customer segments in the EdTech market
Is career-focused EdTech winning over general education?
Career-focused EdTech is growing faster in several important businesses because the value of the learning is easier for students, employers and governments to see.
Stride gives us an unusually clean comparison inside one company.
For its 2026 fiscal year, Stride generated $2.52 billion of revenue and averaged 243,900 enrollments. Overall enrollment grew 4.2%.
Career Learning enrollment grew 13.9% to 109,700 students. General Education enrollment went the other direction, falling 2.5%.
Career Learning revenue increased 15%.
The contrast is hard to miss: within the same operator, using much of the same technology and infrastructure, career-oriented education is growing far faster.
Coursera and Udemy are leaning heavily toward workplace skills, professional certificates and enterprise learning as well. Chegg, after the collapse of its traditional study-help business, has also made skilling and employability central to what remains.
The economics are easier when learning has a visible payoff. A career course can potentially be funded by the learner, an employer or a public workforce program. Its outcome can also be tied to a credential, promotion, job switch or concrete skill.
General school software has a much longer chain between purchase and proof that anything improved.
| Recent example | What is growing |
|---|---|
| Stride | Career Learning enrollment +13.9%; General Education -2.5% |
| Coursera + Udemy | Subscriptions, enterprise learning and professional AI courses |
| Chegg | Skilling revenue growing while traditional academic revenue shrinks sharply |
| Duolingo | Daily consumer practice tied to a clear skill |
If you want more recent data on this point, please see our latest EdTech market report.
Did ChatGPT destroy the old homework-help business?
Generative AI has already wrecked much of the old paid homework-answer business because students can get similar help instantly from general-purpose AI.
Chegg is the clearest case.
Chegg generated $163.1 million of revenue in the second quarter of 2024. By the first quarter of 2026, quarterly revenue had dropped to $63.3 million, down 48% from the previous year.
The following quarter fell again to around $52 million.
That means Chegg lost roughly two-thirds of its quarterly revenue in about two years.
The company's skilling business moved in the opposite direction. Chegg reported first-quarter 2026 skilling revenue of $17.6 million, up 9%. The old academic-help engine was collapsing while career-oriented learning continued growing.
Chegg has repeatedly pointed to AI and declining Google referral traffic as major pressures on the business.
The old customer journey explains why the damage has been so severe. A student once searched a homework question, found a Chegg result and paid to unlock the explanation. These days, the student can paste the same question into ChatGPT, Gemini or Claude and receive an immediate response.
Quizlet reached a different version of the same conclusion. It launched Q-Chat as an early high-profile AI tutor, then discontinued the standalone product in 2025 while continuing to weave AI into its broader study experience.
The defensible layer in EdTech has moved away from simply generating an answer. Curriculum, assessment, credentials, trusted content, learning history, school integration and strong habits are much harder to reproduce with a blank chatbot window.
This chart, featured in our EdTech market deck, shows how AI conversational tutor technology has evolved over time
Are schools really using all the EdTech they buy?
Schools currently have access to vastly more EdTech than teachers and students actually use.
Instructure's 2026 EdTech Top 40 gives us one of the clearest measurements because the company analysed actual Canvas integration activity rather than relying on website visits.
The dataset covered more than 12.6 million people across U.S. K–12 institutions, including more than 11 million students and 618,000 educators.
Districts had access to an average of 3,001 unique digital tools.
The average student used four integrated tools during the year. The average educator also used four.
We should be careful with that ratio because “available tools” and LTI launches are different measurements. Even so, the scale of the gap is extraordinary.
Schools have accumulated websites, apps, subscriptions, extensions and integrations over years of digital expansion. Every extra vendor can create procurement work, privacy reviews, security checks, rostering issues, training requirements and support requests.
The latest Instructure data suggests districts are starting to consolidate around a smaller core of products that can prove regular use.
That is healthy for the sector. A school contract has always been a weak definition of product-market fit when teachers rarely open the software.
The tougher question for a new EdTech company today is much better: can the product become one of the four tools an educator actually uses?
So what is actually working in EdTech now?
EdTech is clearly working now, although the evidence points to a fairly specific formula: structure the learning, keep people practising, improve feedback, make teachers more effective, or connect the skill to an outcome people genuinely care about.
The strongest evidence comes from several directions that fit together surprisingly well.
Structured practice continues to produce learning. Human tutoring remains powerful. AI can improve weaker tutors and provide useful direct tutoring when the interaction is well designed. Teachers are already reporting large time savings from AI. Feedback systems can help students revise more effectively. Duolingo has shown that an education product can become a genuine daily consumer habit. Career learning continues growing because the payoff is easy to understand.
The weaker models also share something.
Chegg's collapse shows how fragile paid answer access has become. The Khanmigo experiment shows that putting a capable tutor one click away does little if students rarely have substantive conversations with it. Instructure's data shows how little school software survives the jump from procurement to everyday use. Universities are discovering that unlimited AI assistance can make traditional assignments easier to complete without making the student more capable.
As seen above, the technology itself is rarely the scarce part anymore.
A good explanation is cheap. A new quiz is cheap. Generating a lesson plan is cheap. Producing thousands of exercises is getting cheap too.
Knowing what a learner should do next, making the learner actually do it, spotting when the learner is confused and creating a reason to return tomorrow remain much harder.
That is where EdTech is working today.
If you want more recent data on this point, please see our latest EdTech market report.
In our EdTech market deck, we identify pain points entrepreneurs should prioritize
OUR METHODOLOGY
This analysis asks a simple question with several very different answers: EdTech: what is actually working now? We broke that question into separate dimensions including student learning, teacher productivity, tutoring, feedback, consumer habits, career outcomes, school usage and business durability, then tested each one with the most direct evidence we could find.
Learning claims were weighted toward randomized trials and meta-analyses. Questions about how teachers and students actually behave were assessed through large surveys and observed usage data. Company growth, adoption and business durability were checked against official results, SEC filings and first-party operating data.
We also separated what a technology can achieve under carefully designed conditions from what happens when the same type of technology reaches ordinary classrooms and everyday users. That distinction is especially important for AI tutoring, where strong experimental performance can coexist with weak day-to-day participation.
Comparisons were chosen to isolate what was actually changing. Where possible, we favored like-for-like comparisons, established educational interventions as reference points, and within-study or within-company contrasts that reduced unnecessary differences between populations, products or business models.
Calculations derived from reported figures were used only when they made scale, intensity or change easier to understand, such as converting weekly teacher time savings into an approximate school-year total or comparing daily users with monthly users.
Individual studies and company results were treated as pieces of evidence rather than standalone answers. We assessed the strongest recent findings point by point, then looked for patterns that repeated across different settings, products and types of evidence. Conclusions were given more weight when independent evidence converged in the same direction.
Key sources used for this analysis include NBER's two-year Khanmigo school experiment, the Harvard randomized AI tutoring study in Scientific Reports, HEPI's Student Generative AI Survey 2026, Gallup and the Walton Family Foundation on teacher AI use and time savings, Stanford's Tutor CoPilot research, Stanford's meta-analysis of 89 randomized tutoring studies, Duolingo's Q2 2026 SEC filing, Coursera's Q2 2026 results, Stride's FY2026 Form 10-K, Chegg's Q2 2026 results, and Instructure's EdTech Top 40 usage dataset.
This chart, featured in our EdTech market deck, shows revenue breakdown by region across Europe, Asia, North America, Africa, and South America in the EdTech market
Related blog posts
- EdTech: what is getting real adoption now?
- The latest funding news in the EdTech market
- Is the EdTech Market growing now?
- EdTech: what are the top startups?
