Imagine a company created an AI system that could do everything a professional software engineer does.
Not just write snippets of Python or JavaScript, but understand vague customer requirements, design complex architectures, build secure enterprise applications, deploy them to production, maintain them for years, and adapt them as business needs changed.
That wouldn’t simply be another software product.
It would arguably be the greatest competitive advantage in modern business.
The global software development industry is worth hundreds of billions of dollars annually. Companies spend enormous sums employing engineers to build financial platforms, healthcare systems, cybersecurity products, logistics software, mobile applications, cloud infrastructure, and internal business tools.
If one company possessed AI capable of replacing most of those developers, there would appear to be a far more profitable strategy than selling subscriptions.
Instead of licensing the technology to everyone else, it could quietly become the world’s largest software consultancy. It could complete projects in a fraction of the time, charge less than every competitor, win contracts across every industry, and generate extraordinary profits while everyone else struggled to compete.
That is the central logic behind the skeptical argument.
Revolutionary technology is usually exploited before it is shared.
So Why Are AI Companies Selling It?
The answer is that reality is considerably more complicated than the argument suggests.
One reason is that today’s AI simply isn’t capable of replacing professional software engineering in the way many headlines imply.
Modern coding models are exceptionally good at generating code, explaining unfamiliar programming concepts, writing documentation, fixing bugs, producing unit tests, and accelerating repetitive development work.
But software engineering is not simply writing code.
Professional developers spend much of their time understanding customer requirements, discussing trade-offs, designing architectures, reviewing code, coordinating with colleagues, ensuring security, complying with regulations, planning long-term maintenance, and making countless judgement calls that depend on experience rather than syntax.
These are areas where AI still struggles.
Even the companies building frontier AI systems are generally careful about this distinction.
For example, OpenAI describes its coding models as tools that help developers become more productive rather than fully autonomous replacements for engineering teams. The company’s own research frequently highlights ongoing work in reasoning, reliability, and long-term planning rather than claiming these problems have already been solved. You can read more in OpenAI’s Research at https://openai.com/research.
Similarly, Anthropic regularly discusses both the strengths and current limitations of large language models in its published research, particularly around reliability and safe deployment. Its research library is available at https://www.anthropic.com/research.
Google DeepMind also continues publishing papers showing impressive advances in AI reasoning while acknowledging that today’s systems still require significant human oversight in many real-world applications. Its latest publications can be found at https://deepmind.google/research/.
Software Development Is More Than Coding
One reason discussions around AI coding often become misleading is that many people assume programming consists primarily of typing code into an editor.
Ask almost any experienced software engineer, however, and they’ll tell you that writing code often occupies only part of their working day.
The harder work frequently involves understanding what clients actually want, balancing conflicting business priorities, making architectural decisions that will affect products years into the future, reviewing other developers’ work, identifying security risks, and communicating with multiple teams across an organisation.
Generating thousands of lines of code in seconds is certainly impressive.
Generating the right software that solves the correct business problem remains a much harder challenge.
That distinction explains why AI coding assistants have become valuable productivity tools without eliminating engineering teams.
Why Selling AI May Actually Be the Better Business
Even if AI eventually becomes dramatically more capable, selling it may still make economic sense.
Running frontier AI models is extraordinarily expensive.
Training modern large language models requires enormous computing clusters containing tens of thousands of GPUs. Operating those models requires continuous spending on electricity, networking, cooling systems, and cloud infrastructure.
The companies developing these systems need enormous recurring revenue simply to continue improving them.
Selling AI through subscriptions and APIs creates exactly that.
There is another advantage as well.
Every developer who uses an AI coding assistant helps improve the next generation of models.
Millions of users expose AI systems to bugs, edge cases, unusual programming languages, complex architectures, and unexpected problems that internal testing could never replicate.
That constant stream of feedback is one of the reasons models continue improving so rapidly.
If AI companies kept their technology completely private, they would lose access to one of their greatest assets: millions of real-world users testing their systems every day.
The Marketing Has Probably Gone Too Far
Where the skeptics do have a point is in questioning some of the marketing surrounding AI.
Over the past two years, there has been no shortage of predictions claiming software engineers are about to disappear.
Some startups promote autonomous coding agents that supposedly build entire applications with little or no human involvement. Social media is filled with stories of one-person billion-dollar companies powered entirely by AI.
Most professional software organisations simply do not operate this way today.
Large software projects remain collaborative efforts involving engineers, designers, product managers, security specialists, legal teams, quality assurance professionals, and customers themselves.
AI can assist many of these activities.
It cannot yet replace all of them.
AI Still Makes Plenty of Mistakes
Another reason companies have not quietly replaced software engineers is that AI-generated code continues to require careful review.
Developers regularly encounter hallucinated APIs, subtle security vulnerabilities, incorrect assumptions about business logic, inefficient algorithms, and code that appears convincing but fails under real-world conditions.
Research published by the Association for Computing Machinery (ACM) has repeatedly highlighted that while AI coding assistants can significantly increase developer productivity, they also introduce new risks when developers accept outputs without sufficient verification. The ACM regularly publishes research into software engineering and AI at https://cacm.acm.org/.
In other words, AI often behaves less like a fully autonomous engineer and more like a very fast junior developer who still requires experienced supervision.
That is incredibly useful.
It is not the same as replacing an engineering department.
Productivity Is Not the Same as Replacement
History suggests that transformative technologies rarely eliminate professions overnight.
Calculators made mathematicians more efficient.
Spreadsheets transformed accounting.
Computer-aided design revolutionised architecture.
Search engines changed legal research.
None of those innovations removed the need for skilled professionals.
Instead, they shifted where those professionals created value.
AI appears to be following a similar path.
The best developers increasingly use AI to automate repetitive coding tasks while focusing more of their time on architecture, product design, debugging, security, and customer needs.
That represents a significant evolution of software engineering rather than its disappearance.
So Does the Original Argument Hold Up?
The claim that “if AI were really that good, companies would keep it to themselves” is an effective challenge to some of the more exaggerated claims surrounding artificial intelligence.
It reminds people to distinguish between marketing and measurable capability.
But it is probably too simplistic as a complete explanation.
Technology companies have often become more valuable by building platforms than by competing directly in every industry those platforms serve. Microsoft became enormously valuable by selling operating systems rather than becoming a manufacturing company. Amazon built cloud infrastructure instead of limiting it to its own retail business. NVIDIA became one of the world’s most valuable companies by selling GPUs rather than keeping them exclusively for internal projects.
AI companies appear to be following much the same strategy.
The evidence today suggests that AI coding is neither useless nor magical.
It is a remarkably powerful productivity tool that allows skilled developers to work faster and tackle larger problems. It continues to improve at an extraordinary pace, but it still depends heavily on human judgement, experience, and oversight.
If one day AI truly becomes capable of autonomously replacing entire engineering teams, the business landscape may change dramatically.
Until then, the fact that AI companies are eager to sell their coding tools tells us less about some hidden conspiracy and more about where the technology actually stands today: immensely useful, commercially valuable, but still a long way from becoming an unlimited digital workforce.
