What AI is good at, and what it isn't

Three ways to see it

1. 1

   AI in 2026 is genuinely strong at a specific cluster of tasks. Drafting routine text, translating between English and Urdu when the topic is everyday, summarising long documents, extracting structured fields from messy ones, writing code, classifying text by sentiment or category, generating images from a brief, transcribing audio, polishing a CV, brainstorming options. If your task lives inside this cluster, AI saves real hours, often half a working day per person per week.

2. 2

   AI is unreliable, often dangerously so, on a second cluster. Multi-step arithmetic, accurate citations, anything that requires up-to-the-minute facts, anything where the right answer depends on a long chain of strict reasoning, anything involving a person you actually know, anything where the data is sparse for Pakistan. The Zakat calculation belongs here, as does the question 'what is the latest case law on Section 491 CrPC?' and 'is my cousin's flu serious?' Treating these tasks like the easy cluster is how serious mistakes happen.

3. 3

   There is a third cluster: tasks AI can do only if you wire it up with the right external tool. AI plus a calculator plug-in does arithmetic well. AI plus a court database does citations well. AI plus a current-news search does facts well. AI alone, sealed in its training cut-off, does none of these reliably. The skilful user knows when to attach a tool and when to ask a person.

The why-tree

Try this with Claude

AI-edge prompt: 'I am about to ask you for [task]. Before answering, tell me whether you are reliable for this kind of task, what could go wrong, and what I should double-check. Then answer.' This single habit prevents the majority of professional AI mistakes.