It seems that many software development companies these days are having trouble scaling AI adoption. Some say that, after a lot of initial excitement, they’re pulling back.
Most companies rolling out AI coding assistants are actually chasing the wrong metric. They expect more code, produced faster, and for a while that is exactly what happens. The reality is that in many companies, especially those where management doesn’t have this kind of hands-on experience, people don’t really understand the psychological and motivational aspects of software development at all.
The first wave of adoption is straightforward: give developers an AI assistant, and output goes up. It holds together right up until autonomous and multi-agent systems enter the picture. At that point the developer stops writing code line by line and starts decomposing problems, defining boundaries, setting constraints, delegating work to agents, evaluating what they produce, resolving conflicts between them, and making architectural calls.
That is the job of an architect or an engineering manager. The developer was usually hired and trained to do something else.
Those skills come from years spent on complex systems, from architectural trade-offs made under real constraints, from failures and their consequences, from learning how to direct the work of other people. Not every developer has had the chance to build that kind of experience, and that has nothing to do with how good they are at writing software. I’ll add that this kind of work isn’t even what a lot of developers want to do.
AI is removing the part of the job where most developers are most comfortable, and at the same time it is raising the value of skills that only a smaller group of senior engineers currently hold.
As AI gets better at writing software, writing software matters less as the core engineering skill. What becomes scarce is judgment.
- What to build.
- How to break a problem apart.
- Which constraints cannot be violated.
- Which decisions are reversible.
- How to tell whether a set of agents is heading in the right direction.
- How to verify that thousands of generated lines actually hold together as one coherent system.
Companies are starting to realize that “adopting AI” and “becoming AI-native” are two completely different things. The first one means: “We give Claude Code/Codex/Cursor to our developers.” The second one means rethinking how software is designed, built, tested, deployed, and maintained when a significant part of the coding is done by agents.
The real work ahead is teaching developers how to lead AI, investing in the judgment of senior engineers rather than rolling out another license. Which makes the next phase of AI adoption much less about tooling, and much more about engineering leadership.
That is a much bigger organizational shift than most companies have budgeted for.