Not all companies who have begun using LLMs for code development have seen any productivity increases. The "20/80" divide between the 20% of companies who have been more productive using AI vs. those who have seen no productivity increase is a gap in architectural integrity. Evidence from early 2026 indicates that the gains reported by the top 20% are, in many cases, essentially a loan taken out against the future stability of their products. We are currently seeing the emergence of a phenomenon called Comprehension Debt, which is far more dangerous than traditional technical debt.

Traditional technical debt happens when a human takes a shortcut (e.g., "I'll skip the documentation so I can ship today"). The human still understands why they took the shortcut and might raise a lower priority ticket to address it. In the "productive" 20% of companies, AI is generating up to 41% of new code. Developers are approving AI pull requests that "work" and pass all tests, but they don't actually understand the underlying logic. Research from March 2026 (RocketDevs) shows that after about 18 months of AI-heavy development, teams hit a wall where they can no longer safely modify their own systems because no one on the team understands the "why" behind the AI's architectural choices.

The 20% of companies seeing gains are often the ones who have automated their testing most aggressively. We are seeing a rise in the Project Managers' Green Ticks == warm fuzzy feeling approach where all automated tests pass so the code is shipped, but the system fails in production and or becomes impossible to scale. AI is ok at writing code that satisfies specific test cases but terrible at any kind of architectural judgment. It will often implement the same logic in a dozen different ways across the same project, and this leads to a fragmented, insane codebase that eventually becomes too heavy to do anything with.

Data from various 2026 audits suggests a predictable lifecycle for AI-driven productivity whereby in the early months velocity increases by 50% or more. Features ship more quickly and everyone thinks they've found a silver bullet. During the next 6 months integration becomes harder. Weird bugs begin to appear that take longer to solve. In the 6 months after that delivery cycles grind to a halt. The team spends 90% of their time debugging the AI's previously generated code rather than building new features.

Unlike a financial debt that you can pay off with some sort of cleanup sprint, Comprehension Debt requires a total re-learning of the system, like undoing the work of a team of rogue developers. So, the 20% of the companies that have gone all in on AI (because they essentially have limitless funds) are building their reinvented business models on codebases that no human understands. The real price of AI will hit these companies twice: first in the form of higher LLM API costs, and secondly in the form of massive remediation costs when their AI-generated systems inevitably become too brittle to update.

Microsoft's recent 30-40% LLM written code for the Windows codesbase and the equally recent calamitous Windows 11 updates is not a coincidence.

AI Creates new forms of tech debt

AI Tech Debt Crisis