# Check AI replies in Kannada, Hindi or Tamil before customers see them Version 1.1 ยท Kramaniti Kosh > How to use this file: fill in the Working template section below. It opens in any text editor, or you can paste the whole file into ChatGPT, Claude or Gemini and ask it to help you fill it in for your business. ## Intended outcome A simple way to know whether the AI replies your customers get in Kannada, Hindi, Tamil or any other language are correct, polite and true to your business. For each language, a person who reads and writes it well checks the replies against a short list, writes down every mistake, and decides when that language is ready for wider use. Until then, no reply in that language reaches a customer without a person seeing it first. ## Before you begin - The languages your customers actually write to you in, including mixed ones like Kannada typed in English letters - One fluent checker for each language, who knows your products and how you speak to customers - Your approved facts (prices, timings, policies) so the checker can spot a wrong answer, not just a wrong word - A place to log mistakes, such as a shared sheet beside the chat tool ## How to use 1. Fill in the languages, checkers and checklist below once. Start with the language you are least sure about, not the one you use most. 2. While a language is on trial, the assistant drafts and the checker reads every reply before it is sent, marks it against the checklist, and fixes or blocks it. Each mistake goes in the log. 3. On a fixed day, the owner and checker look at the log together. If the same kind of mistake keeps coming back, fix the source (the fact sheet, the instructions, the words list) before checking again. The owner decides when a language moves from "every reply checked" to "a sample checked each week", and can move it back at any time. ## Working template ## Business and people Business: What you sell and where customers write to you (one line): Owner who decides when a language is ready: Person who keeps the mistake log: ## Languages and checkers Copy this block for each language. - Language (and how customers type it, for example Kannada script, Kannada in English letters, mixed with English): - Fluent checker: - Backup checker: - Stage (on trial, every reply checked / ready, sample checked weekly / paused): - How many replies the checker looks at each week once it is ready: - Date of next review: ## Words and names to keep fixed List the product names, brand names, place names and terms that must never be translated or changed, and any word the assistant tends to get wrong. - Keep as written: - Always use this word for: - Never use: ## Reply checklist The checker marks each reply against these questions. - Is the meaning right? Does it answer what the customer actually asked? - Are the facts right? Price, timing, stock, delivery and policy match the approved fact sheet. - Is it polite in the way your customers expect? Right level of respect, no word that sounds rude, cold or too familiar in this language. - Is it the language the customer used? A customer who wrote in Tamil should not get Hindi back. - Are names and numbers intact? Product names, amounts, dates and phone numbers are not changed or translated. - Does it promise anything it should not? No discounts, health claims, delivery dates or refunds that the business has not approved. - Would a fluent customer find it natural, or does it read like a word-for-word translation? ## Mistake log One row per mistake. - Date: - Language: - What the customer asked (short, no phone numbers or personal details): - What the assistant wrote: - What was wrong (meaning, fact, tone, wrong language, name or number, promise, unnatural): - What the checker sent instead, or "blocked": - Fix at the source (fact sheet, instructions, words list, nothing yet): ## When a language is ready Write the rule before the trial starts, not after. - The checker has looked at replies across the usual kinds of questions, not just greetings: - No mistake about facts, promises or wrong language in the recent checks: - Tone mistakes are rare and getting fewer: - Owner's decision and date: ## What the assistant may and may not do The assistant may: - Draft replies in the customer's language from the approved fact sheet - Say it will pass the question to a person when it is not sure The assistant may not: - Send a reply in a language that is still on trial without the checker seeing it - Translate or change product names, prices or numbers - Make health, medical or results claims about any product - Promise discounts, refunds or delivery dates that are not on the fact sheet ## Stop rules - If the checker for a language is away and there is no backup, replies in that language go to a person to write, not to the assistant. - Any mistake about a fact, a promise or a health claim pauses that language until the owner and checker have fixed the source. - A complaint, a refund request or an upset customer goes to a named person, whatever the language. - Do not paste customer phone numbers, addresses or ID details into the mistake log or an AI tool. ## Owner and review Checklist owner: Look at this checklist again after the first month, whenever you add a language, and whenever the chat tool or its settings change. ## Demonstration This is an illustrative scenario, not a client case study or a measured result. Do not reuse its details as facts about your work. A sample Ayurvedic wellness products shop in Bengaluru takes orders on WhatsApp. Most customers write in English or Kannada, and a growing number write in Tamil. The owner has switched on an AI assistant for replies and wants to know whether the Kannada and Tamil replies can be trusted before letting them go out on their own. ### Sample inputs Demonstration inputs only. Languages: Kannada (checker: the shop manager), Tamil (checker: a part-time staff member who is a native speaker). Both on trial, every reply checked. Words to keep fixed: the shop's product names and the brand name. Fact sheet: approved prices, delivery areas and the return policy. ### Example output #### Mistake log entries [Fact within this demonstration] Tamil: a customer asked whether a hair oil was in stock. The draft answered in Hindi. Marked "wrong language", blocked, and the checker replied in Tamil. [Fact within this demonstration] Kannada: a customer asked what a herbal powder is for. The draft said it "cures" acidity. Marked "promise", blocked, and Kannada replies went back to a person to write until the instructions were fixed. #### What the review found [Inference] Both mistakes came from the instructions, not the language itself: the assistant was not told to always answer in the customer's language, and it was not told that product benefits must come word for word from the approved descriptions. #### Next step [Recommendation] Add both rules to the assistant's instructions, add approved one-line product descriptions to the fact sheet, and keep every reply checked in both languages until the owner is satisfied. This recommendation is not the business's decision. ## Quality check Every language has a named fluent checker and a backup; replies in a language on trial were all seen by the checker before sending; every mistake is in the log with its type and the fix at the source; no reply changed a product name, price or number; no health claims or unapproved promises went out; the owner, not the assistant, decided when a language was ready. ## Limits and human review The template helps a fluent person check AI replies in your customers' languages; it does not make the assistant fluent, it cannot catch mistakes in replies nobody checks, and it is not a translation or language quality certificate. It does not tell you which health or product claims the law allows; take that to a qualified adviser. Keep complaints, refunds and any reply that commits the business with a named person. ## Edition notes New in Kosh at version 1.1. Provider-neutral Markdown instructions; no installation or platform compatibility is implied. Runtime behaviour has not been verified across AI providers or languages. Reuse terms have not yet been published. Background reading: [AI Observability Is Becoming the Trust Layer](/insights/ai-observability-is-becoming-the-trust-layer/), on being able to see what an AI system actually said and why, and [The Exception Log Is a System Design Tool](/insights/the-exception-log-is-a-system-design-tool/), on turning repeated corrections into fixes. Keeping an assistant checked, adjusted and trusted after it goes live is part of Kramaniti's [Complete Lifecycle Retainer](/retainer/).