AI has changed the economics of software development, not just its speed

Faster code. Smarter review. Better software. (Image: KRS)

---

While AI-accelerated development has made writing code faster, according to Cape Town-based custom software development company KRS, the bigger shift has been in the economics of building the software itself, not just the speed.

“Three months ago, we wrote about what AI-accelerated development looked like at our company,” says Ayesha Bagus, Head of People and HR Director at KRS.

“At the time, three ideas were central to how the team worked: review code, use test-driven development (TDD) and keep AI away from infrastructure. But three months is a surprisingly long time in AI development, and one of the things we’re learning is that we have to adapt at a much faster pace, constantly looking at our approaches, building in feedback loops and changing how we work as we learn.” 

Data from the wider industry is showing proof of that. According to Google’s 2025 DORA (DevOps Research and Assessment) report, 90% of tech professionals surveyed are now using AI in their work, with more than 80% saying it has improved their productivity. The research describes AI as an amplifier: AI doesn’t fix a team, but amplifies what’s already there.

“That idea really resonates with what we’ve experienced at KRS,” adds Bagus. "As AI increases the pace at which we can build, it has also made the quality of our engineering practices, feedback loops and decision-making even more critical."

Faros AI's 2026 research also paints an interesting picture. Looking at two years of telemetry from 22 000 developers across more than 4 000 teams, it found that AI adoption is producing real increases in output: epics completed per developer are up 66%, task throughput is up 33.7% and pull request (PR) merge rates are up 16.2%.

But the same data shows incidents-to-PR up 242.7%, bugs per developer up 54% and median time in review up more than 440%.

"Those numbers capture something we had been experiencing and helped us to reconsider our strategy," says Bagus. "AI makes development considerably faster, but if the way you work does not change, the extra output can simply create pressure somewhere else."

Bagus adds that this is where the real opportunity lies.

"For us, the most interesting part of AI has been the reduction in the cost of trying something. We can take an idea, explore an architecture, build a prototype, test an approach, decide that it is not the right direction and get to try something else much faster than we could before.

“These are exciting times for our industry as the boundaries shift,” she continues. “That gives us more room to experiment with our clients and explore their ideas that might have been dismissed before because the cost of finding out whether they would work was too high."

That is a significant change in the economics of software development.

Bagus adds: "If you can imagine something, there is now a much better chance that we can explore it with you, get something tangible in front of you, test it and learn from it before committing to months of development. AI has made it faster to turn an idea into a prototype and engage with clients on their vision and needs. The challenge now is making sure that the increase in speed does not simply create an increase in risk."

The speed of writing code, while an advantage, has created a bottleneck, as there is now more code to review.

"Applying our QA process to a much larger volume of AI-generated code took an already clumsy process bottleneck and made it far worse. Our developers became faster, but our seniors became increasingly overloaded trying to do PRs," shares Bagus.

Faros saw something similar on a much larger scale. Its data shows the median time to first PR review increasing by 156.6%, the average time spent in code review increasing by 199.6%, and the median time in review increasing by 441.5%. It also found that 31.3% more PRs were being merged without review as developers simply gave up.

That prompted a rethink of what code review should actually achieve at KRS.

"Instead of asking a developer to look at more code, we looked for ways to create more evidence that the software is working before it reached the human reviewer," says Bagus. "Test-driven development has always been part of our engineering practice at KRS, but AI has made us much more deliberate about it. When AI is producing a lot of code, we need to make sure it is producing what is expected."

That search for supporting tools led the team to Playwright. "TDD gave us a way of defining expected behaviour as part of the development process, and Playwright allows us to test that behaviour through the application itself,” adds Bagus. “This includes what is rendered in the browser and how the system responds to real engagement. As AI matures in working with these tests, the feedback loop will get shorter again."

KRS has also introduced a second layer of scrutiny: using one AI model to challenge the output of another, with one model producing the work, and another independently reviewing and testing it.

“If Claude generates something, we ask GitHub Copilot to challenge it, or the other way around," says Bagus. "We do not see this as a replacement for human review, but another layer of challenge before it needs a human to spend time on the change, and it has helped us to really think about where that human attention is most valuable."

But not every change carries the same risk, and KRS's process reflects that. "Clearly, a small, well-tested UI change does not carry the same level of risk as authentication logic, payments, customer data or a database migration," says Bagus. "We feel confident moving faster when the risk is low but put considerably more engineering judgment around changes where the consequences of getting something wrong are high."

That same caution shapes how KRS uses AI around infrastructure and data. “At first, we were not comfortable allowing AI access to our infrastructure," says Bagus.“Today, we are using AI to help write an infrastructure script or draft a deployment script. We remain cautious about its ability to execute. AI can help prepare the change, but we do not allow it to decide when a change is executed in production."

Databases remain an area where KRS is particularly conservative. Bagus adds: "While AI is improving quickly, we see real weaknesses in its ability to reason about schemas, migrations and the consequences of changes to real data. We still believe those decisions need experienced human oversight."

For KRS, all of this points to a change in what it means to be a software engineer, as the value engineers bring has shifted. "AI is very good at producing possible answers and exploring alternatives,” says Bagus, “but it still needs an engineer who understands the system and the user requirements well enough to know which questions matter."

That shift is central to how KRS is developing its next generation of engineers through the company's AI-First Graduate Engineering Programme, for which applications are now open for 2027.

"We want our developers to become highly capable users of AI, but we also want them to develop the judgment to know when to question the output, when to test it more deeply, and when to step away from the tools and think about the people who use our systems," says Bagus. "The ability to adapt quickly is becoming part of the engineering discipline itself. The technology is moving quickly, so our feedback loops need to be fast enough to tell us when something that worked yesterday is no longer the best way of working today."

KRS says the value of AI should extend beyond producing software faster, instead allowing room to explore ideas, test possibilities, learn earlier and get to something useful sooner.

“We have different kinds of conversations about what is possible when the cost of experimentation falls,” adds Bagus. “A client can bring us a difficult problem or an ambitious idea and, rather than immediately asking how many months it will take to build, we are starting to ask: what can we explore, how quickly can we test it and how fast can we learn?

"That is the opportunity we see in AI-accelerated development at KRS. We are learning how to use the speed and capability of these tools while applying the engineering judgment needed to build software that works in the real world. If you can imagine it, let's explore what we can build. And as technology continues to change, so will the way we work with it," Bagus concludes.