Three years of AI headlines have been heavy on capability announcements and light on "here's what I actually use it for every Tuesday." The gap between AI as a concept and AI as a daily work tool is worth closing. So: what do practitioners - the people running it daily, not demoing it at conferences - actually find useful?
The tasks that stuck
The most consistent wins are in writing tasks that used to require disproportionate mental energy. Drafting a performance review that's accurate but diplomatically phrased. Writing a cold email that doesn't sound like a cold email. Turning bullet-point meeting notes into a coherent summary. These tasks used to take 30 minutes because getting the tone right was the hard part. Now they take five.
Code is a close second. For people who aren't software developers but occasionally write formulas, scripts, or automation logic, AI has been genuinely useful. Explaining a complex Excel formula in plain language. Writing a Python script to rename 200 files. Debugging why a Google Sheets VLOOKUP isn't working. You don't need to be a developer to describe what you want, and you don't need to understand every line of code it produces to get the job done.
Summarization is the third reliable bucket. Long PDFs, research papers, lengthy email threads - dropping them into ChatGPT or Claude and asking for a structured summary saves real time. The quality is usually good enough for initial orientation, even if you'd want to verify anything before sending it to a client.
Where it still disappoints
The use cases that sound impressive but rarely stick: full content generation from scratch (the output is generic without substantial editing), market research (AI confabulates data confidently and you have to verify everything anyway), and customer service automation (handles FAQ-level questions fine, falls apart on anything requiring judgment).
Research-heavy work is still a grind. AI can help you understand a topic faster, but it can't reliably replace source verification. Anyone who's caught a confidently stated hallucination in a client-facing document knows why you can't skip that step.
The structure that works
The best use cases share a pattern: there's a clear input (a document, a task description, a specific question), a clear output format (a summary, a draft, a list), and a human reviewing the result before it goes anywhere. The further you stray from that - asking for original insights, real-time data, or judgment calls - the less reliable it gets.
What works is using ChatGPT, Claude, or similar tools as a fast first-draft machine, not an autonomous agent. The people getting the most out of AI aren't the ones fully automating tasks. They're the ones who've found specific points in their workflow where a five-minute AI output saves two hours of from-scratch work.
That's a real return. It's just not the one that gets written up at conferences.