Editor’s note: In recent years, Vietnam was celebrated as Southeast Asia’s rising star—a high-growth market fueled by a young, mobile-first population, strong developer talent, and rapid consumer adoption. Today, that story is shifting from potential to maturity.
Over the coming weeks, TNGlobal will publish an exclusive series of conversations with venture capitalists, ecosystem builders, startups founders. Whether you are an investor looking for the next growth driver in ASEAN, a founder scaling in the region, or an industry observer, this series offers a grounded, ground-level look into Vietnam’s digital economy.
When Chegg lost 99 percent of its market value in just over three years, it became the clearest warning yet to the edtech industry: a business built on selling academic answers has no future once answers are free with AI. MAAS, a Vietnamese education company that has spent a decade supporting Vietnamese students at foreign universities, has drawn the opposite lesson — pushing toward deep expertise a general-purpose model cannot replicate, from research and publication advisory to tools like Grade Check that estimate a student’s grade.
In an interview with TNGlobal, Do Mai Anh, founder of MAAS, shared insights on where genuine value in academic support sits, why AI has raised the penalty for not knowing your field rather than removing it, how Vietnam improving its business conditions, and a vision to 2030 growth, among others.
Below are the edited excerpts:
Chegg lost 99 percent of its market capitalization, approximately $14.7 billion, from its peak in February 2021 as students moved to free AI tools. What does that tell you about where the real value in academic support lies, and what are the lessons for MAAS?
Chegg sold access to answers. Answers are now free, fluent and infinite, so a business whose entire value sat inside what a general model does well was always going to be repriced towards zero.
Worth noting that the market turned on Chegg in 2021, more than a year before ChatGPT launched. AI finished the company but did not start the decline, so the vulnerability was in the product all along and the technology merely found it.
MAAS is acting on two factors. The first is to move towards work that needs deep expertise, where a general model cannot follow. Publishing advisory is the clearest case: a model can polish a sentence in a manuscript, but it cannot tell an author that their contribution is real but framed against the wrong literature, nor it cannot carry the responsibility when that judgement is wrong.
The second is to build tools that sit where free models fail. Grade Check, a tool built by MAAS, takes a student’s own draft, along with the brief and the marking rubric, and returns an estimated grade per criterion, with feedback on what would move each one up a band. What makes it work is what it was built from: real assignment briefs, real submitted work, and the grades and written feedback lecturers actually returned on that work.
Teachers and senior professors in each field review how the model marks, and correct it, so the model keeps improving; they are not reading individual submissions. The result reads a rubric the way a marker does rather than the way the rubric reads on paper.
The use we did not design for now matters most. Where a university permits AI use with declaration, a student still has no way to judge whether the result is good. A chatbot says the text is fine, because fluency is what it optimizes for, and fluency is not the same as meeting a rubric. MAAS estimate is for study purposes only, while the lecturer sets the real mark and MAAS does not write students’ work for them.
As AI becomes equally accessible to students and educators, is academic success shifting from what a student knows to how well they can direct and interrogate AI? Is that a problem, or has education become a contest of utilizing AI?
Partly yes, but the framing hides something. Directing AI is not a skill that sits on top of knowing nothing. You cannot interrogate a model in a field you do not understand, because you cannot tell when it is wrong, and it is wrong often with total confidence. So AI has not replaced domain knowledge. It has raised the penalty for not having it, because a plausible wrong answer is now free and beautifully written.
What decides whether this becomes a contest of tool use is how universities respond, and two shifts are already visible. First, disclosure rather than prohibition: a growing number of institutions permit AI use in preparing work provided the student declares it, and have written that into assessment policy. That is more honest than a ban, because a ban is unenforceable and everybody knows it.
Second, assessment is moving towards group work, oral examinations, in-class tasks and supervision meetings where a student explains their own choices. Harder to run and more expensive to staff, and the only version I can see surviving this decade, because it measures the thinking rather than the presented results.
Where I would push back is on detection, which many institutions reached for first. A Stanford study in Patterns in 2023 ran TOEFL essays by non-native English speakers through seven commercial detectors: 61 per cent were flagged as machine-written on average, against roughly 5 per cent for native speakers, because second-language writing is more statistically predictable.
To be clear: undeclared AI use in assessment is a genuine problem, and universities are right to act on it. My argument is narrower, these tools are too unreliable to carry the weight of a misconduct case, and the students they misjudge are disproportionately the ones writing in a second language.
Vietnam has been tightening enforcement around tax declaration and IP rights covering software and equipment. Is stricter compliance ultimately healthy for edtech, or does it add friction that falls disproportionately on smaller local companies?
It is healthy, and I would say that even on the days it costs us. In a market where a competitor undercuts you by not declaring revenue or by running on software it has not paid for, being the better company does not win. Their price is not a real price. It is a subsidy taken from the tax base and from the people who built the software.
Enforcement removes that gap, leaving a market where transparent companies can grow and substandard ones cannot hide, which matters more in a sector that sells trust to parents.
The direction has also moved fast. Resolution 68 set the direction in May last year, the lump-sum tax method for household businesses ended on 1 January, Decree 70/2025 pushed electronic invoicing much further down the size scale, the amended Intellectual Property Law took effect on 1 April, and the authorities ran a nationwide inspection of unlicensed software in May.
The friction is real, because the cost of compliance is largely fixed rather than proportional. Analysis in the Vietnamese press has put the annual cost of the tax transition for a small household business in the tens of millions of dong, which can exceed what that household previously paid in tax altogether. A twenty-person company carries the same filing and licensing burden as a five-hundred-person one, but without a finance or legal team, so the founder absorbs it out of hours that would have gone into the product.
So the number I watch is not the enforcement target but the reform plan’s own aim of cutting compliance costs by at least 30 per cent. Smaller companies do not need lighter rules. They need a lower cost of following rules and enforcement that is steady rather than arriving in campaigns. Predictability is worth more to us than leniency.
Vietnam has set ambitious 2030 targets, including doubling GDP per capita and growing the digital economy. How well does that align with where you are taking MAAS, and what would the company need to look like by 2030 for you to call it a success?
Where we are taking MAAS is based on three moves: up the expertise curve into research and publication support, out of pure labor services into tools, and inward into automating our own operations. Each of those meets a different part of the 2030 agenda.
Doubling GDP per capita is arithmetic about output per person, and a company whose work is delivered by people produces more without adding headcount once it automates its own operating process. That is the least glamorous reading of the digital economy target, which tends to be taken to mean founding new technology companies when the larger share of available gain sits in existing companies becoming more productive.
We are ten years old and have automated much of the operation: the same target, counted differently.
Moving up the expertise curve meets the research half of the agenda. Academic advancement here runs through publication. Appointment as a professor or associate professor requires papers in reputable international journals, and doctoral regulations require published output. The ambition is right, but it has left a great many capable Vietnamese researchers holding good work with no route through review. That gap is closed by expertise rather than software, and it is the part of our work that AI is least able to take.
As for 2030, MAAS measures success with three criteria, none of them related to finance. MAAS aims to helpe measurably more Vietnamese researchers publish in credible journals; MAAS tools serve students at a scale our mentors never could, well enough to beat a free chatbot; and MAAS can run the same whether I, as the founder, am there or not.