AI has moved the work, not reduced it: KRS developer on the hidden cost of AI-assisted coding

Khanyisa Real Systems (KRS) senior developer Miengha Behardien says AI-assisted software development has shifted developer time from writing code to reviewing, testing and validating it, changing what businesses should ask about AI-generated code.

---

AI-accelerated software development has made writing code significantly faster, but it hasn't reduced the total work, according to Miengha Behardien, a senior developer at KRS, a custom software development company in Cape Town.

The hours once spent writing code haven't disappeared. They've shifted to reviewing, testing and validating what AI produces before it reaches production. "AI has made output a lot faster, but the time I saved writing code has shifted to verifying the code it produced," says Behardien.

For businesses investing in AI-accelerated software development, that's an important distinction. Faster code generation doesn't mean faster software delivery if the output still needs careful engineering review.

Like many developers in 2026, Behardien uses AI throughout her working day. "I use AI to help me with my everyday tasks, from planning and debugging to actually writing code and designing solutions."

Her primary tool is Claude Code, which she says has become her preferred AI coding companion. How much work she delegates to it depends entirely on the task.

"This is always shifting. Predefined tasks are easy to hand over to the AI with clear instructions. More nebulous discovery-type tasks require a more hands-on approach with the AI agent, guiding it in the direction you want."

Rather than replacing software engineering, AI has become another tool within the development process. The developer remains responsible for deciding what should be built, evaluating whether the output solves the problem correctly, and ensuring the final code meets engineering standards.

At KRS, that means asking a different set of questions before code ships:

1. Should I let AI do this?
2. Did it misunderstand?
3. Is this over-engineered?
4. Is this maintainable?
5. What should I test?
6. What don't I need to test?

For Behardien, the biggest challenge isn't that AI occasionally gets things wrong. It's that the code it produces isn't always the simplest solution.

"My main issue I have is code bloat, where it seems to take the longest route to get to the solution rather than the simplest."

Some of the time saved generating code goes into reviewing it, simplifying it and confirming it behaves as intended.

This experience isn't unique to KRS. According to Sonar's 2026 State of Code Developer Survey, 40% of developers say AI has increased technical debt by generating unnecessary or duplicate code.

Behardien believes part of the reason lies in the data AI models were trained on. "AI has a wealth of knowledge, but it was trained on human code with human errors, which can be inadvertently perpetuated."

Because AI-generated code can appear convincing even when it contains flaws, Behardien has developed her own way of working with it.

Rather than only correcting the code, she revisits the instructions she originally gave the AI.

"Generally, when this happens, I go back and re-engineer my prompt to the agent, explain why what it did was wrong, and note that issue for future."

The goal is to improve future output while still treating every generated response as something that requires human review.

Behardien also gives her AI agents explicit rules. The first is straightforward.

"My agents are never allowed to commit or push code without my approval. This is to prevent committing code that could look right on the surface, but actually not be addressing the issue or solving the problem at hand."

Her second rule is that the AI should ask clarifying questions rather than make assumptions whenever requirements are unclear. Those safeguards sit alongside KRS's existing engineering processes.

One area where Behardien believes AI has genuinely improved her work is during the early stages of problem-solving.

Before bringing a technical challenge to the wider development team, she now uses AI to first explore possible approaches.

"Before, we would bring the whole team together and brainstorm ideas. Now I ask Claude to suggest solutions, outline pros and cons, highlight any gaps it sees, etc. I then bring those findings to the team, and we can build from there."

Instead of replacing collaboration, AI gives the team a starting point for discussion.

As AI becomes a standard part of software development, businesses are asking how they can trust code that wasn't written entirely by a person.

For Behardien, the answer lies in the engineering process rather than the AI tool itself.

"I would tell clients that all code is checked and vetted, not only by myself, but by the rest of my team as well. We apply targeted review and testing where it provides confidence. Our engineering process focuses on validating what matters most, rather than trying to test everything."

More broadly, KRS's process also includes adversarial QA: using a different AI model to review code than the one that generated it, so that pitting models against each other catches issues a single model working alone might miss.

For organisations evaluating a software development partner, questions around review processes, testing and quality assurance are just as important as understanding how AI is being used.

Although AI has changed how Behardien approaches many aspects of software development, it hasn't changed what she believes all developers should know.

"An understanding of the basics of programming is invaluable to be able to use AI to its fullest potential. But a course on prompt engineering and how to do agentic development would definitely help on top of the basics."

She also believes AI has changed how she approaches difficult work.

"AI work is exciting. It has given me the confidence to tackle problems that felt overwhelming on my own."

For KRS, that's the role AI plays in the business: a tool that helps experienced developers work more effectively, but one that still depends on human judgment, review and accountability to deliver software that clients can trust.