What AI is, and is not, at work

What you are actually using

When people in India say “AI” at work in 2026, they usually mean a chat app: ChatGPT, Gemini, Claude, Grok, or Sarvam. You type a request. It writes a reply. That is a generative tool: it predicts the next words from patterns in its training, not from your shop’s books or your college’s syllabus.

Older “AI” still runs in the background: a bank flagging a UPI pattern, a hospital reading an X-ray, Aadhaar face match. Those systems are trained for one job. They are not the same as a chat box that will happily invent a circular number if you ask it to summarise policy.

Hold this sentence: the model is a junior intern who types fast, has never visited your city, and will not admit when it is guessing.

Why US-only examples fail here

A US tutorial will show you a neat English email and a clean PDF. Your week looks different: a WhatsApp voice note in Hinglish, a GST invoice with CGST/SGST, a college assignment that must stay on the syllabus, a customer who will pay by UPI only if the amount is exact.

Indian languages also cost more tokens than English: the same paragraph in Hindi can be “longer” to the model. Code-mixed messages (“kal evening tak bhej dena”) are normal here and awkward for models trained on tidy essays. If the output sounds like a US brochure (“unleash”, “revolutionise”), it is wrong for your customer even if the grammar is fine.

Three jobs the tool may draft, three it must not decide

Draft is allowed: a WhatsApp reply you will edit, a first cut of notes from a lecture you attended, a table extracted from an invoice you can see.

Decide is not allowed: the tax you will file, the grade you will submit, the delivery date you promise, the eligibility for a scheme, a medical or legal conclusion. Those need a portal, a teacher, a CA, or a human who can be held to account.

If you cannot name the official page you would check, you are not ready to send.