Everyone talks about where AI works. Far fewer people ask where it breaks down.
Over the past year, software engineering discussions have been dominated by success stories. AI generates code. AI accelerates delivery. AI helps engineers move faster. These observations are real, and our survey of approximately 250 engineers confirmed them.
But alongside the productivity gains, engineers also described situations where AI consistently struggles.
What’s interesting is that these answers were remarkably consistent. Across hundreds of comments, engineers repeatedly pointed to the same categories of work where AI becomes unreliable, expensive to validate, or simply not worth using.
The pattern was so clear that it reveals something important about the current state of AI-assisted engineering. AI does not fail randomly. AI fails predictably.
The most frequently mentioned challenges fell into several categories:
At first glance, these may seem like unrelated issues. In reality, they all share a common characteristic. They require understanding.
One engineer summarized the issue directly: AI struggles with complex business logic and broad context; rewriting makes usage not worth it.
Another observed: Generated code will never be understood by the programmer at the same level as if they had written it themselves.
These comments point toward a distinction that appears repeatedly throughout the survey. AI is very good at generating solutions. It is much less reliable when deep understanding becomes the primary requirement.
When requirements evolve during implementation, when architecture emerges while work is underway, or when business logic contains significant nuance, engineers spend increasing amounts of time validating, correcting, and rewriting generated output.
In these situations, generation speed stops being the bottleneck. Understanding becomes the bottleneck.
One of the strongest themes in the survey involved large AI-generated pull requests. Engineers repeatedly described situations where AI produced large amounts of code quickly, but where the resulting changes became difficult to review and maintain.
The problem isn’t simply code volume. The problem is human comprehension, as reviewers still need to:
AI can accelerate generation dramatically. Human understanding does not scale at the same rate. This is why many engineers reported faster coding without proportional improvements in delivery.
The most important finding isn’t any individual failure mode. It’s the pattern connecting all of them.
AI struggles when:
Conversely, AI succeeds when:
This suggests that the boundary between successful and unsuccessful AI usage is not technical. It’s structural. The more a task depends on understanding, judgment, and evolving context, the less effective today’s AI workflows become.
The survey did not reveal an AI that works everywhere. Nor did it reveal an AI that fails everywhere. Instead, it revealed something more useful. Engineers consistently identified the same situations where AI creates leverage and the same situations where it creates friction.
The lesson is not that organizations should avoid AI. The lesson is that understanding where AI struggles may be just as important as understanding where it succeeds.
Because the most effective teams are not the ones using AI everywhere. They are the ones using it where it actually works.
