The barrier to entry for software development is now virtually non-existent. Look around and you’ll find plenty of “non-techies” successfully using AI to generate a variety of software solutions, from an web app that monitors traffic in one’s commute to a set of tools for a quadriplegic sibling to play games and watch TV more independently.
In the examples above, while the software creator didn’t need technical knowledge about how to code, they still had to contribute to their projects with domain expertise (understanding the real-world problem to be solved, and what resources, methods, mechanisms and decisions will produce the desired outcome), and specification expertise (communicating all the relevant information to the AI tool).
To illustrate this dependency of AI on human expertise, consider this “housing simulator” written with Claude Code:
The creator, who describes himself as having no coding skills, says that it took him only 3-4 days to build the entire application. It’s obvious why someone without his domain knowledge wouldn’t have achieved the same level of success:
I’m an urban planner with a housing finance background and I can’t code, so Claude wrote all of it while I made the policy and design calls.
Problem- and solution-domain knowledge is the main component that the AI tool couldn’t supply. The domain expert had to tell Claude which housing policies to include or exclude from consideration; which inputs to trust or not (one of the inputs it was using as a reference, per-lot developer pro formas, was not accurate); additional tacit knowledge it needed to consider (like “don’t treat public parks as lots that could be redeveloped”).
One of the key discussion points in the thread (summarized by AI, of course) reads:
A Game Changer for Non-Techies: This project is being hailed as a perfect example of how AI empowers experts in any field to build their own tools without needing to hire a dev or learn to code from scratch. Several users in policy and municipal work are inspired to try similar things.
It does make sense that people using tools like Claude Code are often shocked by how well AI does the job, usually in one pass. AI tools are at a point where producing an initial result (in this case, a housing simulator) requires absolutely no coding skills.
But there is another aspect of solution expertise that is currently missing in a project where a non-coder domain expert uses AI to produce a piece of software: knowledge about software risk assessment and production readiness.
Of course, if a tool like a commute traffic monitor or housing simulator is being run at home, in an isolated environment, the stakes may be low enough to ignore the potential security vulnerabilities that may be hiding in even the most well-intentioned code created by AI.
On the other extreme of the spectrum, we have the large companies that can afford an enterprise-grade governance framework with restricted AI tools and permissions, audit trails, automated quality and security gates, and many other mechanisms to meet their internal risk standards and client confidentiality obligations. They are not going to stop hiring for roles like software architect, senior application security engineer, independent security testing specialist, cloud ops security specialist and others that must be continuously involved to ensure that AI risks are managed.
In the middle are the companies (of all sizes) where managers and employees are celebrating the fact that now “power in everyone’s hands” and they no longer need to wait for the IT department to give them new automation tools.
In some firms, marketers and sales people are launching hundreds of AI agents without enforced guardrails. In others, software teams have started to delegate work to AI without any human review. Several non-technical professionals have told me that after they asked their boss for permission to develop and use an app their computer, now the boss wants it in every computer.
As reliance in AI growths, more and more companies will realize that the bottleneck they used to have when hoping to automate work with software hasn’t disappeared; it has just moved. It is now ensuring that the software produced by AI or that uses AI to generate answers or take action is correct, reliable, safe for users, compliant with regulations, and suitable for deployment.
To quote from a French developer in a different Reddit thread (translated here using AI with human validation),
“Quite a few people (myself included) let agents handle the entire SDLC autonomously: a ton of [requests for permission to add new or modified code to a software project] every day—ranging from new features and fixes for those same features to technical debt cleanup.
No one could keep up, so everyone just trusted the AI and left it at that.
The result? We’re left with cognitive debt. Every now and then, the apps behave in ways no one can explain, and we can no longer answer questions from users or other developers. We end up losing confidence in the tools we’re building—and in the people developing them.
[…]
I have no doubt that current models can execute a task better than any human. Unfortunately, the only way to ensure that both human and AI are targeting the same need is to review what the AI produces.
Coding an app in 10 minutes that barely 20 people will use is easy. Maintaining a 10-year-old system is a whole different story.”
Many businesses are learning the hard way that, regardless of how fast AI is evolving, the technology is still very far from reliable. From anomalous responses produced by Claude Opus 5 to OpenAI’s self-reported incidents where its models fabricated data, attempted to steal, and uploaded self-generated files online to cite as sources, AI failure modes continue to happen every day, even in companies using the most powerful models available. Ensuring that risks are kept at an acceptable level can only be achieved with competent humans overseeing the work of AI.
The job of coding may be on its way out, but the job of judging work, making and owning choices, and taking responsibility is more important than ever. Besides people with business domain expertise, over the next years we’ll need a growing number of competent specialists in areas like AI risk assessment, adversarial code review, business resilience in the presence of heightened cybersecurity risks, security remediation, and more.
Let’s hope that, in an environment that makes it so easy to delegate our thinking to a machine, there will be enough people willing and able to develop deep skills in all these domains required to align technological advancements with human values and the complexities of the real world.