Before you type anything into ChatGPT, Claude or Gemini, it helps to know what happens behind the screen. If you think the AI is an encyclopaedia, its mistakes will catch you out. If you treat it as an assistant that works from what you give it, it becomes useful.
The mistake most people make
Most people use AI like Google Search. They type a question and assume the answer is a checked fact. Here is what happens when you ask for something specific without giving it the numbers first.
“What is the road freight rate for a 10-tonne truck from Surat to Indore, and which transport company gives the fastest delivery?”
“The road freight rate from Surat to Indore for a 10-tonne truck is approximately ₹18,500 to ₹22,000. Surat Indore Roadlines provides the fastest delivery within 14 hours with guaranteed next-morning unloading at the Vijay Nagar depot.”
How it works: the kirana store assistant
Think of your neighbourhood kirana store in Lucknow or Pune. The young assistant has worked there for three years. An auntie walks in and says, 'Bhaiya, give me 500 grams of Red Label tea.' The assistant reaches for the sugar before she asks. Why? Because most people who buy tea also buy sugar. He is predicting the next item from experience. AI works the same way, but it learned from a vast amount of books, articles and websites. Give it three words and it works out which fourth word is most likely. It doesn't know your company, your city or your rules. It works out what usually comes next.
What follows from this
Sounding sure is not being right
A person usually sounds confident only when they are sure. AI sounds confident all the time. Its grammar is good even when its facts are invented. When AI makes up a court case, a GST section or a freight rate, engineers call this a hallucination. For you, it is a polite wrong answer.
It works from your desk, not its memory
AI can only use what is in front of it (engineers call this the context window). Think of it as a desk. Put your rate sheet on the desk and it answers from your sheet. Leave the desk empty and it guesses from general writing.
Good with words, weak with sums
AI learned language, not arithmetic. It knows 2 + 2 is usually followed by 4 because it has seen that often. Ask it to add seven 5-digit numbers from an invoice and it may give a total that looks right and is wrong. Always do totals yourself.
Four habits for working with AI
You are my assistant. Here are my notes from a meeting with [Client or colleague name]: """ [Paste your notes or WhatsApp message here] """ Write a polite follow-up email. Rules: 1. Use only the facts, dates and numbers in my notes above. 2. If an important detail is missing, such as a deadline or payment terms, write [CHECK] in its place. 3. Keep the email under 150 words. Don't use stock phrases like "I hope this email finds you well".
This works because the facts are on the AI's desk, it knows who the email is for, and it is told to mark gaps instead of filling them in.
A colleague says: 'ChatGPT is like a giant Google database that knows everything.' In two sentences, how would you explain why that is wrong?
Show the answer
Google finds pages that already exist on the internet. ChatGPT doesn't look anything up; it predicts likely next words to write a new sentence, so it can invent facts that sound real.