Don't Build Critical Logic on a Guess
LLMs guess the next word. That is fine for drafts, risky for invoices, loan math, and anything that must be exact every time.
If you are bringing AI into a product, hallucination is not a bug you prompt-engineer away. It is built into how LLMs work. They predict the next token. For many prompts there are several valid continuations.
Words like table or chair have a small set of reasonable options. Numbers do not. The price of a chair could be almost anything, so the model picks one that sounds plausible. That is why salary figures, EMI math, and billing totals need deterministic code behind them, not a longer system prompt.
Prompt engineering helps with tone and structure. It does not make an LLM reliable for business-critical numbers. The fix is architecture: use AI where fuzzy is OK, and use code, schemas, and validation everywhere else.
That is how I built EMI Analyzer. AI reads messy payslip text. FOIR and eligibility math run in plain TypeScript. Same inputs, same outputs, every time.