Documentation is famously the part of software development everyone agrees matters and almost no one keeps current, and AI tools have genuinely changed the cost-benefit calculation.
Writing and maintaining documentation takes real time that competes directly with shipping features, and it’s usually the first thing cut under deadline pressure. Exactly the kind of well-defined, lower-stakes writing task where AI assistance helps without much downside risk, the same category covered in our explainer on how coding agents handle testing.
Generating a first draft of docstrings and comments from existing code. Keeping documentation in sync with code changes, flagging when behavior has changed but docs haven’t. Generating README and onboarding documentation from a codebase’s actual structure, often skipped entirely without AI assistance.
AI-generated documentation describes what code does based on reading it, not necessarily why it was written that way. Context a human author would naturally include, the reasoning behind a design decision, is often missing and worth adding back during review.
Closes the gap between how much documentation gets written and how much should exist, but human review adds back the design reasoning an AI reading only the code can’t infer. See GitHub’s own Copilot documentation.
