Twelve months ago, AI coding assistants were the most-discussed new developer tools in years. The coverage was breathless. The claims were large. Now that the novelty has worn off and usage data has accumulated, it is worth looking at what actually happened.
What Improved
Code completion for boilerplate is genuinely better. Autocomplete suggestions for standard patterns, common library calls, and repetitive code structures are faster and more accurate than they were a year ago. This is real time saved, and it is distributed broadly across developers regardless of experience level.
Documentation generation has become useful. The quality of auto-generated docstrings and comments has reached the point where many developers are comfortable using it as a starting point rather than writing from scratch. Editing generated documentation is still faster than creating it from nothing.
What Did Not Change Much
Code review quality from AI assistants remains uneven. The suggestions are frequently technically correct but contextually shallow. They catch style issues and obvious bugs more reliably than they catch design problems or architectural mistakes. A senior developer reviewing code is still doing something qualitatively different from what an AI assistant can do.
The harder problems in software development, the ones that require understanding business requirements, user needs, and system-wide implications, have not been substantially addressed. AI assistants are good at applying known patterns to new situations. They are less good at recognizing when no known pattern fits.
The Honest Picture
Productivity data from teams that track it carefully suggests modest but real improvements. Not the two-times or ten-times productivity claims that some vendors made, but somewhere between ten and twenty percent on typical development tasks. The gains are larger for junior developers and smaller for senior engineers, which is the opposite of some early predictions.
The tools are useful. They are not transformative in the way that was promised. That is fine. Useful tools that do what they say are worth having. The disappointment comes from confusing marketing claims with actual capabilities.
