Is AI actually making software better, or is it mostly a marketing checkbox?
It's a question the entire software industry is dancing around. The optimistic case - sometimes called "deshittification" - argues that AI lets teams ship faster, automate the tedious parts, and deliver features that genuinely weren't possible before. Smart autocomplete in code editors, AI-powered search that understands intent instead of just keywords, automatic summarization of long documents. These are real improvements that save real time.
But for every useful AI feature, there are a dozen that feel bolted on. AI-generated email replies that nobody asked for. Chatbots stuffed into apps where a simple search bar worked fine. "AI-powered" labels slapped on features that amount to basic filtering with a language model wrapper. Companies are racing to add AI to everything because investors expect it, not because users need it.
The honest answer is that we don't have great data yet. Most companies measure AI feature adoption by whether users click the button, not whether the output was actually useful. A chatbot that gives wrong answers 30% of the time still shows strong "engagement metrics" because people keep trying it.
Where AI Clearly Helps vs. Where It Doesn't
The pattern so far: AI works best when it handles a specific, well-defined task where mistakes are easy to catch. Code completion in Cursor or GitHub Copilot, transcription in Otter, grammar checking in Grammarly. These tools do one thing, do it visibly, and let you correct errors instantly.
AI works worst when it's shoved into a general-purpose role with no clear success metric. The generic "Ask AI" button that appears in every SaaS dashboard now rarely produces better results than the features it replaced.
The companies getting AI integration right tend to share one trait: they started with a specific user problem, not with "we need an AI feature." That distinction matters more than the underlying model.