Startups in the AI era: why "surface" products vanish while training, consulting and communities thrive

Three years after the AI boom, an observation widely shared among Vietnamese startup builders is taking shape: most "surface" AI products — chatbots, content tools, automation, expert simulators — no longer exist, while the healthiest businesses are training, consulting agencies, long-term advisory and communities. In his original post, author Tuan Le also lists six structural difficulties that virtually every tech startup in the AI era now faces.

Quick answer

• The central observation: according to author Tuan Le, roughly 80% of the "surface" AI products launched in the boom three years ago no longer exist today (this is an estimate from the author's personal observation, not official statistics).
• The survivors: training, consulting agencies, ready-to-use VPS services, account provisioning, long-term advisory and communities — models that "spread and amplify AI" instead of investing heavily in infrastructure and features.
• Not a new story: the same script played out with forums, Vietnamese social networks, domestic e-commerce and the sharing economy — except that in the AI era, a wrong direction kills within months.
• Six big difficulties for AI-era startups: ideas crushed by native AI platforms, users unwilling to pay, hiring undermined by the one-person-company wave, investors demanding early numbers, legal-compliance risk and security risk.
• SaaS is not exempt either: AI boosts productivity, but a single platform update can wipe out a software company's advantage.

What is a "surface" AI product?

A "surface" product is an AI product that lives only in a thin layer of interface and features on top of someone else's AI models: chatbots, content-writing tools, video generators, workflow automation, token managers, SEO tools, "virtual experts", knowledge wikis. What they share is speed — quick to build, quick to launch — while their core value sits inside a third-party model that anyone can call through an API, and that the model provider itself can absorb as a built-in feature at any moment.

The "healthy" models in the author's observation do the opposite: they do not sell features, they sell the spread of AI and human accompaniment — teaching others to use AI (training), doing it for businesses (consulting agencies), staying alongside for the long term (advisory), packaging infrastructure ready to go (pre-installed VPS, accounts), and gathering people (communities).

Two groups of models: the vanishing and the thriving

Criteria	"Surface" products	Spread / service models
Examples	Chatbots, content/video tools, automation workflows, token managers, SEO tools, virtual experts, knowledge wikis	Training, consulting agencies, advisory, pre-installed VPS, account provisioning, communities
What is sold	Software features	Knowledge, effort, accompaniment, trust and a human network
Competitive moat	Almost none: everyone calls the same APIs	Real: personal reputation, client relationships, a community of one's own
Platform dependence	Very high: one update from the model provider can erase the product	Lower: the stronger AI gets, the more valuable teaching and done-for-you services become
Fate after 3 years (per author)	Around 80% no longer exist	The "healthiest" group

Why do surface products die so fast?

Three mechanisms, visible throughout the original post, squeeze this group at the same time:

• The technical barrier has disappeared. When a chatbot or content tool can be vibe-coded in days, being first is worth almost nothing — competitors, and the AI platform itself, can copy it instantly.
• Users have been spoiled. Cheap tokens and the ability to build their own tools with AI make users deeply reluctant to pay for anything they feel they could "just ask AI to make".
• Life cycles have collapsed. The author recalls earlier eras — the decline of forums, the death of Vietnamese social networks, domestic e-commerce dead ends, sharing-economy bankruptcies — to stress that a wrong model dies in any era; in the AI era it dies in months instead of years.

Six difficulties of tech startups in the AI era

This is the backbone of the original post — six pains that, according to the author, almost every tech startup now runs into:

1. Ideas get crushed by native AI. Whatever a startup manages to think up and prototype, the big AI platforms can integrate immediately; there is almost nothing left "that AI hasn't thought of".
2. Users will not pay. Spoiled by cheap tokens and DIY vibe coding, users default to expecting everything to be cheap or free.
3. You cannot gather a team. The vibe-coding and "one-person company" movement means a skilled individual can earn well alone with a fleet of AI agents — extremely hard to recruit them into a startup to "endure" the early phase together.
4. Investors changed the rules. Funds still value AI capability, but now demand early revenue numbers; the era of long money-burning growth runs is over for most startups.
5. Legal compliance is a standing risk. Regulation is moving deep into copyright, data and information safety — precisely the areas where AI products most easily get caught.
6. Security is an oversized danger. Many AI startup founders do not come from security, so they easily stumble into incidents they cannot take responsibility for or properly handle.

Not just startups: SaaS hurts too

The author extends the observation to established SaaS and software companies: AI helps their teams become more productive and build things once only imagined — but it is also ready to "take away their livelihood in a single update", when the AI platform or operating system integrates the very feature the company sells. Platform risk, in other words, is no longer a young startup's problem alone.

So what should one do?

The original post deliberately gives no answer — the author closes by saying the answer exists, "but not in this post". Still, the observations themselves point to a few consistent directions readers can draw:

• Sell outcomes and services before products: training, consulting, done-for-you work and advisory are where clients still pay, because they are buying time, responsibility and trust — things AI does not yet replace.
• Build moats beyond features: community, personal reputation, client relationships and vertical domain data are assets a platform update cannot copy.
• Treat legal and security as day-one work: for AI startups these are no longer "later" items but conditions of survival as regulation and risk tighten.
• Accept short cycles: validate a direction with real revenue within months instead of burning years of resources on a product the platform can replicate at any time.

Frequently asked questions

What exactly is a "surface" AI product?

A product with only a thin feature layer on top of existing AI models — chatbots, content tools, automation, virtual experts — owning no core technology, proprietary data or distribution channel, so it is easily copied and easily replaced by the platform itself.

Why do AI training and consulting outlive product startups?

Because clients pay for knowledge, effort, responsibility and accompaniment — values that live in people and trust. The stronger AI becomes, the more demand there is for being taught, having it done for you and being advised, while pure software features get cheaper and cheaper.

How does vibe coding affect startup hiring?

It creates the "one-person company" generation: one skilled person with a team of AI agents can earn a good income solo, so startups struggle to hire them into a low-pay, high-risk early phase the way they once could.

What do investors demand differently in the AI era?

Per the author, funds still prize AI capability but simultaneously require early business numbers; the long burn-money-for-growth model is effectively closed for most startups.

Is the "80% of products vanished" figure accurate?

It is an estimate from the author's personal observation after three years of watching Vietnam's AI startup community, not official statistics. Its value lies in the trend picture — thin products disappearing en masse while service models endure — rather than in the absolute number.

Conclusion

"Startups in the AI era" is worth reading because it argues against the usual optimism: advantage no longer lies in shipping features fastest, but in owning what AI cannot copy — people, trust, community and the ability to take responsibility. For anyone about to build with AI, the six difficulties it lists make a practical checklist to review before writing the first line of code.

Source: the post "Startup thời AI" by Tuan Le on Facebook (2 Nov 2026), edited and structured by Wotech Electronics.