From AI Tools To AI Development Systems: How To Build A Better Software Engineering Process
Roman Eloshvili is the Founder and CEO of XData Group, a B2B software development company with a focus on the European banking sector.
gettyAI coding tools have thoroughly changed the way engineers work, making it possible to generate code and build features much faster than we used to be able to do by hand alone. But now we face a bigger question: How do we build development processes around those tools?
I’ve already said in the past that simply writing code is not the same as actually putting reliable software into production. With AI becoming the fastest part of the engineering process, it can actually lead to entirely new problems if testing, security and human decision-making don’t keep up.
It’s not enough to simply plug an AI coding tool into an existing setup and expect everything else to stay the same. The process on the whole also needs to change. According to a report published by GitLab earlier in 2026, roughly 79% of developers claim that AI has greatly boosted code output, yet despite this greater productivity, the overall software delivery cycle has not become faster.
That’s why I believe the next logical step here is to move away from individual tools and toward a more complete AI development system.
The short and direct answer is: decisions. It’s crucial to understand that AI does not make engineers obsolete—it just changes where their focus needs to be. AI can handle a lion’s share of technical work, but it doesn’t decide what a business should actually build and why. That still belongs in the hands of the people leading that company.
At the same time, there are also things AI simply cannot do on its own because they involve human-to-human interactions in the real world. For example, a model can build a mobile application, but someone still needs to open it on an actual device and check if everything works as intended from the user’s perspective. This cannot be done through automation alone.
Even when a release is ready to go into production, it still needs to go through approvals and external reviews. A submission to place your app in App Store or Google Play, once again, can’t be handled by AI. Human oversight and engagement are still necessary at these stages.
The rest of the process, on the other hand, can become much more automated. Once the business requirements and the specifications for the necessary piece of software have been established, AI can generate the code, another AI will review it and then the result can be put through a smoke test to make sure all the core functions are in the clear.
The main point is this: The human engineer no longer needs to read every line of code—that would just defeat the point of using AI at this scale. So instead, their attention moves to decision-making, where it can yield the most value.
Humans decide what should be built, and check the finished product to determine whether it actually works and meets the established goals. Both the entry and exit points of the development process remain in the purview of trustworthy specialists, not just models.
Of course, we also have to talk about an obvious problem here. If AI can write 10 times more code in the same time span as a human would manually, that means it can also produce 10 times more bad code. And AI cannot be trusted to simply check itself.
A method my own team uses is to separate code generation from code review. Like I already mentioned earlier, one model writes the code, and a model from a different vendor reviews. This is done to achieve a clean context. If the same model handles both, it’s less likely to scrutinize past decisions, because those already made sense to it when the code was originally generated.
This is also why reviews should start with the original specifications, rather than the code itself. Looking for obvious mistakes and removing them is, naturally, a must. But the much more fundamental goal is to understand if the code actually matches with what you’re trying to build in the first place.
This changes how quality control works and at what stages it needs to happen. And as the use of AI-generated code becomes more commonplace, I think this change in approach will become increasingly necessary.
As we covered at the start of this article, when AI becomes the fastest part of development, it can lead to new problems.
Code stops being the main bottleneck, and instead that bottleneck is born from everything that still requires in-person intervention from the human side of things. It can be a review, a customer decision, testing or something else. In my own practice, one of the most common statuses for a task these days is: “Merged, waiting for device testing.”
That gives us a notion of how the development environment is changing around us. The real challenge now is not simply producing software faster; it’s how quickly we can go from an idea to a tested product that’s fully ready for release. And I think companies need to start paying more attention to measuring this difference.
The full impact of AI on the development process goes beyond just looking at how many more lines of code we can now write in the same amount of time. We also need to look at how many defects get discovered after a merge, and the amount of rework required as a result. And, of course, how long the queue of features is that are merged but still waiting for verification.
These, to my mind, are much better signals of whether AI is actually making your business faster, and by how much.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
