AI lets domain experts turn knowledge into working software
Architecture, security and scale remain the work of engineers
Cheaper experiments let small firms test previously uneconomic ideas

In March 2025, Y Combinator, the accelerator from which Airbnb, Dropbox and Stripe started, revealed that in about a quarter of its winter batch startups, 95 percent of code had been written by artificial intelligence. This number says something that the market was slow to admit, that software creation has ceased to be a privilege for those who write code from scratch. In the same year, Collins Dictionary named “vibe coding” word of the year, while the Swedish platform Lovable, which converts simple descriptions into applications, exceeded $100 million in annual recurring revenue eight months after it reached its first million. Marc Andreessen wrote in 2011 that software eats the world. Fourteen years later artificial intelligence is changing who sits at the table of this meal and the change deserves a closer reading than the one reserved for it by the celebrations.

Creating Software Once Required Developers and Budgets

For most of the history of computing, an idea for an application was cheap and building it was expensive. A warehouse manager in Penang who wanted a simple tool to track his deliveries had to write specifications, find an external partner or hire developers, approve a budget and wait months until he saw the first working version, which often didn’t look like what he had in mind at all, because somewhere between the meeting with the vendor and the repository, his knowledge of the problem was lost in translation. Most such ideas never got there. They stayed on a spreadsheet or an email that no one answered.

The cost of this threshold exceeded the money. Knowledge of how an accounting department, doctor’s office, or oil palm plantation works was found in people who had no way to code it, while coding ability was concentrated in a limited and sought-after population of engineers. Gartner captured it in its own way when it predicted that the global market for low-code technologies would reach $26.9 billion in 2023, up 19.6 percent from 2022 and when it explicitly cited the high cost of tech talent as one of the reasons for this demand. Businesses were buying such tools primarily because the queue of requests to the IT department was growing faster than the department’s ability to empty it and much less because of some enthusiasm for the technology itself.

Generative AI Moves the Low-Code Frontier Even Further

The first crack in the threshold came with low-code and no-code platforms. Webflow, Glide and Bubble.io allowed people without technical backgrounds to set up websites, internal applications and simple databases by dragging ready-made components onto a screen. In Malaysia, the same pattern was first seen in commerce, where EasyParcel simplified shipments for small sellers, AVANA gave local businesses an online store without a single line of code and Shopee and Lazada’s seller tools did the same for millions of sellers. Gartner estimated that by 2026, users outside of official IT departments would make up at least 80 percent of the user base of low-code tools, up from 60 percent in 2021. However, the limit remained visible, since the user could only combine what the platform provider had already built.

Generative AI has moved this boundary much further. The interface is now everyday language and the model writes itself the building blocks that in low-code should have been ready-made. Andrej Karpathy, Tesla’s former director of artificial intelligence, coined the term “vibe coding” in February 2025 to describe a way of creating applications in which the developer almost forgets that code exists and within nine months the phrase had become word of the year for Collins Dictionary. The numbers that accompanied it moved just as quickly. Lovable announced in July 2025 that more than 10 million projects had been created on its platform, with about 100,000 new ones every day, while the company declared 2.3 million active users and 180,000 paid subscribers.

What changes in practice is the distance between the idea and the first prototype. A work that in the past would have entered a development cycle of several months can now take the form of a rough but functional version within a few days, provided that the person describing it knows exactly what he wants to solve, a condition that turns out to be much heavier than it sounds.

Field Experts Become Software Creators

For years, the distinction between user and software manufacturer was almost absolute. The Grab and Touch ‘n Go apps, which started with commuting and tolling and ended up offering payments, loans and insurance to millions of Malaysians, were designed by professional engineering teams for people who just used them. The user could complain, leave a review in the app store and wait for the next update. Today, the secretary of a doctor’s office who for years passed the appointments on an Excel sheet can describe to a model the reservation system she needs and try out a first version of it before the week ends.

The greatest value of this shift lies in the knowledge of the field. An accountant knows which exceptions delay the closing of the books each month, an agronomist knows which mark on the plant leaf precedes a disease and a distribution manager knows why the route suggested by the map fails every Friday afternoon. This knowledge rarely fit in its entirety in a specification document and the loss in translation from the field expert to the developer was a cost that no budget recorded. When the expert himself makes the first version, that loss is reduced and the resulting tool often fits the job better than a much more expensive system designed remotely. In a country with the cultural and linguistic diversity of Malaysia, where a trader in Kota Bharu and a nurse in Kuching are facing problems that a product team in Kuala Lumpur would probably never identify, the ability to build their own tools brings to the digital ecosystem solutions that reflect far more needs and far more people than previously fit on a software company’s client list.

The change is also affecting those already in the profession. Stack Overflow’s 2025 survey, with more than 49,000 participants in 177 countries, recorded that 84 percent of developers are using or planning to use AI tools, up from 76 percent a year earlier. The same technology that allows a non-expert to build a prototype is changing the way the expert works and the debate about who can be called a programmer is becoming more blurred than either side would like, mainly because the title has ceased to describe a single skill.

The Core of Engineering Still Belongs to Engineers

There’s a recent precedent worth remembering before declaring everyone a programmer. A few years ago, data science was also considered democratized because anyone could load a CSV file into Python’s pandas library and output graphs and models in minutes. The result was a wave of analyses that looked professional and relied on correlations that read as causality, samples that no one checked and models that were never validated, because the library had been made accessible to everyone while the math and statistics beneath it had stayed exactly where they were. For those who had spent years on these foundations, the experience was more than annoying.

Data on AI-generated code shows that the risk of repetition is real. Veracode, in its 2025 generative AI code security report, looked at 80 programming tasks in more than 100 language models and found that the generated code introduced security vulnerabilities in 45 percent of cases, with the failure rate in Java exceeding 70 percent. In the Stack Overflow survey, 46 percent of developers said they don’t trust the accuracy of tools, compared to 33 percent who trust it, while 66 percent cited solutions that are almost correct but not completely correct as the main annoyance. A randomized study by the organization METR additionally found that experienced developers in large open-source projects took 19 percent longer when working with AI.

None of this cancels out the value of the tools, but it does show where the prototype ends and the system begins. The architecture that will withstand tens of thousands of concurrent users, the protection of payment data, the reliability of a service that is not allowed to go down, the connection to the old systems of a bank and the very artificial intelligence models on which all this new creativity rests remain the work of people with deep technical training. Aonic, a start-up from Penang that analyzes images of crops from drones, is a good example, since the farmer’s knowledge defines what is worth measuring while image analysis and flight systems are high-level engineering. Even the leadership of Y Combinator warned that the code written by artificial intelligence can encounter problems at scale and that the founders need classic programming skills to maintain what they built.

Cheaper Experiments Change the Economics of Software Creation

The most tangible financial consequence lies in the cost of experimentation. When an internal tool cost a developer’s salary for a quarter, a small business in Ipoh or Kuala Lumpur had to be almost certain to pay off before commissioning. With costs falling to the level of a monthly subscription and a few afternoons of work, trials are being done that before didn’t even make it to the management table and most will fail, which is perfectly acceptable as long as failure costs little. Y Combinator reported that the 2025 winter batch was growing overall at a rate of 10 percent per week, a speed linked to small teams being able to test, reject and rebuild products much faster than previous generations of founders.

The obvious objection is that, with a 45 percent vulnerability rate, this democratization simply shifts the cost from development to data leaks and fixes. The objection would be valid if every prototype went straight into production. Veracode’s metrics are for code produced without explicit security requirements and without auditing and METR’s finding is for the most experienced engineers in huge, mature repositories with over a million lines of code, which is exactly where the expert’s value is greatest. At the other extreme, where an idea just needs to be tested on ten clients, the cost of one mistake is small and the value of speed is great.

For technology managers and SME owners, the conclusion is practical. Prototyping can be done by those who know the problem, while anything that touches customer data, payments, or critical operations has to go through engineers before it reaches real users and the budget has to anticipate this scrutiny from the start instead of discovering it after the first incident. Engineers’ time is thus shifting from writing forms and management pages to architecture, code review and security, that is, to the part of their work that no model has yet proven they can do on their own.

Table 1: Division of Software Creation Between Domain Experts and Engineers

A quarter of Y Combinator companies with 95 percent machine-written code is a number that’s easy to read the wrong way. It shows that software creation has opened up to people who would never have gotten to a functional product before and this is in line with an idea that often recurs in the debate about Malaysia’s digital trajectory, that the future will be made more by technology creators than by its consumers. A creator in this deeper sense is anyone who turns their own knowledge of a problem into a tool, even if they didn’t write a single line themselves and this definition leaves the core of systems where it belongs, to engineers. If software eats the world, artificial intelligence enlarges the table and more and more people are now sitting at it as cooks, as well as guests. How much of what is cooked will eventually be served remains open. Lovable reported 100,000 new projects a day in the summer of 2025 and no available data yet indicates how many of them reached actual customers or how many first passed through the hands of an engineer.

The views expressed in this article are those of the author(s) and do not necessarily reflect the official position of The Economy or its affiliates.

References

Andreessen, Marc (2011) 'Why software is eating the world', The Wall Street Journal, 20 August.
Becker, Joel, Rush, Nate, Barnes, Elizabeth and Rein, David (2025) Measuring the impact of early-2025 AI on experienced open-source developer productivity. Berkeley: METR.
CNBC (2025) 'Y Combinator startups are fastest growing, most profitable in fund history because of AI', CNBC, March.
Collins Dictionary (2025) Collins Word of the Year 2025: vibe coding. Glasgow: HarperCollins Publishers.
Gartner (2022) Gartner forecasts worldwide low-code development technologies market to grow 20% in 2023. Press release, 13 December. Stamford: Gartner, Inc.
Heim, Anna (2025) 'Eight months in, Swedish unicorn Lovable crosses the $100M ARR milestone', TechCrunch, 23 July.
Lovable (2025) $100M ARR and Lovable Agent. Stockholm: Lovable, 23 July.
Mehta, Ivan (2025) 'A quarter of startups in YC's current cohort have codebases that are almost entirely AI-generated', TechCrunch, 6 March.
Stack Overflow (2025) 2025 Stack Overflow Developer Survey. New York: Stack Overflow.
Veracode (2025) 2025 GenAI Code Security Report. Burlington, MA: Veracode.