The popular story — that Byju’s collapse spawned dozens of new test-prep apps — doesn’t survive the numbers: Kota lost more than half its students and roughly half its revenue, and the market is consolidating, not multiplying. The documented problem is not a black market but legal misleading advertising, addressed by India’s consumer authority in November 2024. And “AI personalisation” is now the least verifiable sales claim in the sector.
Kota, a city in the Indian state of Rajasthan, exists almost entirely to prepare teenagers for two exams: the JEE, which grants entry to the Indian Institutes of Technology, and the NEET, the national medical entrance test. Sixteen-year-olds arrive from across the country, live in hostels, and sit in lecture halls that hold a thousand people. In December 2024, local operators quoted by the Indian press estimated that student numbers had fallen to 85,000–100,000, against a usual 200,000–250,000, and that annual district revenue had dropped to around ₹3,500 crore from ₹6,500–7,000 crore. The decline was attributed to negative publicity following student suicides. That figure is where any honest analysis has to start, because it inverts the standard narrative: India’s test-prep sector is not exploding. It is shrinking and concentrating.
Inside that contraction, one thing is growing: the number of digital products promising to replace the human tutor with a recommendation engine. The artificial intelligence component involved is almost always the same, and it deserves to be named properly: an adaptive learning engine — a machine learning model trained on the history of student responses to large question banks, which estimates each learner’s probability of error on each topic and decides which exercise to serve next. Allen Digital, the digital arm of one of Kota’s oldest coaching brands, launched in 2022, says it uses AI to personalise lessons, tests, marking, practice and feedback, and in January 2024 released an app built around a knowledge graph, a map of dependencies between concepts. That is a plausible description of a real architecture. It is also a company description: independent evaluations of how well these systems work in Indian test-prep apps are simply absent from the public record.
What actually happened to Byju’s
The version told abroad is that Byju’s went bankrupt and dozens of competitors rushed into the vacuum. The procedural reality is slower and duller. Think & Learn, the Byju’s parent company, is in an open insolvency process. On 29 January 2025 India’s National Company Law Tribunal annulled the reconstituted committee of creditors and restored the original one, ruling that the interim resolution professional holds no decision-making power and cannot reclassify creditors. On 4 May 2026 the Supreme Court dismissed founder Byju Raveendran’s appeal against that restoration. On 23 July 2026 the Bengaluru NCLT suspended the bidding process until 31 August, after the founders challenged the claim asserted by GLAS Trust, which on the strength of that claim controls 99% of the voting rights on the committee. The insolvency itself began with a suit by the Indian cricket board, BCCI, over a ₹158 crore sponsorship contract.
This is not a liquidated company; it is a company stuck. And Kota’s crisis came earlier, with causes of its own. Lining the two up as cause and effect is storytelling, not data.
The real winner went public
If a wave of new entrants existed, capital would show it. What capital actually showed was one very large IPO. PhysicsWallah raised ₹3,480 crore in an offering open from 11 to 13 November 2025, priced in a band of ₹103–109 per share, with ₹1,562.85 crore raised from anchor investors on 10 November. It listed on 18 November and closed 44% up. Its FY25 accounts show revenue of ₹28.9 billion, up 49%, and net loss narrowed to roughly ₹2.4 billion (₹243 crore) from ₹11.31 billion. Financial coverage read that debut as an exception to the broader slowdown in Indian edtech, not as the start of a new spring.
Here is the economic implication that matters, and it is counterintuitive: a serious personalisation engine has enormous fixed costs and almost zero marginal ones. You need an item bank calibrated on tens of millions of responses, engineers to maintain the model, and a continuous flow of students answering questions so the data does not go stale. The first student costs a fortune; the millionth costs nothing. A cost structure like that does not produce dozens of small competitors. It produces two or three very large operators and a long tail of vendors selling interfaces on top of content bought elsewhere.
The problem isn’t illegality. It’s what’s legal to say
On 13 November 2024, India’s Central Consumer Protection Authority issued the Guidelines for the Prevention of Misleading Advertisements in the Coaching Sector, 2024: they prohibit false or misleading claims, including guarantees of success, and require transparency about candidates’ actual ranks. The enforcement record as of November 2024 stood at 45 notices and ₹54.6 lakh in fines against 18 institutes; an official statement of 17 April 2025 put the totals at 49 notices and ₹77.6 lakh.
Those are modest sums, but they point to where the weak spot is: in advertising, not in any underground trade. The old abuse was putting a top scorer’s face on a hoarding when the student had attended the course for three weeks. The new one is subtler, and the guidelines do not name it because it arrived later: claiming that an artificial intelligence system “personalises each student’s path.” That is not a promise of a result — so formally it is not a guarantee of success — and it is practically impossible to falsify from outside. No parent can open the model and check whether there is a knowledge graph with Bayesian mastery estimates behind it, or a hand-written rule tree alternating easy and hard questions. On screen, both produce the same sentence: we’re adapting your study plan.
The difference between algorithmic personalisation and a fixed sequence is invisible to the user, and no regulation currently requires anyone to document it.
Rajasthan is regulating buildings, not software
On 19 March 2025 the Rajasthan Coaching Centres (Control and Regulation) Bill, no. 11/2025, was introduced. It requires every centre and every branch to register separately, mandates a counselling system as a condition of registration, requires “fair and reasonable” fees with fee and refund information made public, and applies to centres with 50 or more students. Allen, through director Naveen Maheshwari, welcomed it; smaller centres called it a death sentence. Both reactions are rational: compliance is a fixed cost, and fixed costs are absorbed by whoever is already large.
But the text reasons in terms of buildings, classrooms and headcounts. A recommendation engine has no branches to register and no students to count in a room. India’s toughest attempt yet to regulate coaching leaves outside its perimeter precisely the product the sector is migrating toward.
The questions nobody asks
Anyone assessing one of these apps — a family, a journalist, a regulator — has verifiable questions available that almost never get asked:
- on how many responses was the item bank calibrated, and when was it last recalibrated;
- how many exercises must a student complete before the system stops serving a default sequence and genuinely starts adapting;
- do the paths of two students with different error profiles diverge, and after how long;
- does an internal measure of improvement exist, and has it ever been compared against a group that did not use the app.
None of these requires opening the source code. They are metrics any company with a working model already holds, because it uses them internally to decide whether the model works. The fact that they never appear in sales material is the most useful information available: in a shrinking market, the only defensible asset is the opacity of what you are selling.