AI can now write most code, but humans still can't automate away the messy parts—and that's creating a new bottleneck

Anthropic CEO Dario Amodei has predicted that AI is already writing roughly 90% of all code in the world. Instead of creating ten times as many finished apps, most of this AI-generated code is staying private—people are using it to build personal software for their own needs rather than products for public use.

The real work now happens outside the code itself: quality assurance (testing), security audits, documentation, data handling, and keeping apps running reliably at scale. While LLMs can offer advice on these tasks, they cannot fully replace human judgment because they only predict the next word and don't understand how real users will actually behave with the software or what hidden business needs might emerge later.

For anyone shipping software—whether you're a startup founder, a developer, or a business leader—this means writing code is now the easy part, but maintaining and evolving it over time is where your real effort lands. Software needs to be built so it can adapt when circumstances change, and that requires human developers who listen to users, spot inconsistencies, and make predictions about what might break. Trying to automate that step away forces you to spend more time testing and fixing what breaks later instead.