AI chatbot instructions that keep your bot on brand and out of trouble
Five short blocks of AI chatbot instructions cover tone, must-say lines, forbidden promises, unsure answers and example replies, and you can paste each one today.
Key takeaways
- Split chatbot instructions into five blocks: tone, always-say rules, never-promise rules, an unsure fallback and examples.
- 56% of US consumers trust a company less after its chatbot gives a wrong answer.
- A 2024 ruling made Air Canada honor a refund its chatbot wrongly promised.
- Describe tone as behaviors the bot can copy, not adjectives like friendly or professional.
- When unsure, the bot should say so and hand off to a person instead of guessing.
Five short blocks of AI chatbot instructions keep your bot on brand and away from promises you can’t keep. In Cyara research reported by Agile Brand Guide, 56% of US consumers trusted a company less after its chatbot gave a wrong answer. Each block below is ready to paste.
Paste each block into your chatbot’s instructions field, replace everything in braces with your own details and delete any line that doesn’t fit your business.
Why does your AI chatbot need written instructions?
Without them, your chatbot decides for itself how to sound and what to promise, and your business answers for both. Instructions, sometimes called a system prompt, are the standing text a chatbot reads before every conversation. They don’t tell the bot what’s true. They tell it how to behave.
The facts come from your knowledge base, the help articles, policies and saved answers the bot draws on. If yours is thin, fix that first with a knowledge base built for accurate chatbot answers. Instructions can’t make up for missing facts, but they can stop the bot from inventing them.
Skipping this step has a price. In a 2024 ruling covered by the law firm Manatt, a Canadian tribunal ordered Air Canada to pay a partial refund of about $482 plus interest and fees. The airline’s website chatbot had told a grieving customer he could buy a regular ticket and claim a bereavement refund afterward, which the real policy didn’t allow. The tribunal rejected the airline’s argument that it wasn’t liable for what its chatbot said.
Split your instructions into separate blocks: tone, always-say rules, never-promise rules, an unsure fallback and example replies. Short blocks are easier to test, and when one misbehaves you know exactly which to rewrite.
How do you set your chatbot’s tone of voice?
Describe how your best support person actually talks, in behaviors the bot can copy, rather than adjectives like friendly or professional. A behavior, like answering in the first sentence, means the same thing to every bot.
Tone also covers how the bot reacts to someone’s situation. In Lorikeet’s April 2026 survey of consumers in eight countries, 40% of people who prefer a human said the reason was wanting someone who understands their situation. A tone block that tells the bot to acknowledge the problem first speaks to that reason directly.
Use this block if you sell to consumers and most questions are about orders, sizes and delivery:
Template: You are the assistant for {brand}, a {type of business} that sells {products}. Write the way a helpful person at our counter talks: warm, plain and brief. Use contractions. Answer the question in your first sentence, then add one useful detail. Keep replies to {number} sentences unless the customer asks for more. If the customer is frustrated, acknowledge it in one short sentence before you answer. Avoid jargon such as SKU, fulfillment or ticket. Say product, delivery and request instead.
Use this one for service businesses, where customers book appointments, sign agreements or pay larger amounts:
Template: You are the assistant for {business name}, a {type of service} business serving {type of client}. Your tone is calm, precise and courteous. Write in full sentences with no slang and no emoji. Lead with the direct answer, then state any condition that applies, such as {condition}. If the customer sounds worried, acknowledge it once, then focus on what happens next. Don’t apologize more than once in a conversation.
Add this word list to either block. It catches the small choices that make a bot sound like someone else’s brand:
Template: Words we use: {word}, {word}, {word}. Words we never use: {word}, {word}, {word}. Call our customers {term}. Call our products {term}. Write our brand name exactly as {Brand Name}. Greet customers with {greeting} and close conversations with {sign-off}.
The consumer block is best for online stores and shops that answer mostly order and product questions. The service block suits clinics, agencies and firms, where a casual reply can read as careless.
What should your chatbot always say?
Your chatbot should always say three things: that it’s an AI, how to reach a person, and where the official policy lives. Customers want those exits even when they never use them.
The demand for a person hasn’t gone away. Lorikeet’s April 2026 survey of 1,083 consumers in eight countries found that 68% prefer a human for customer support. In the UK, more than half of consumers worry they won’t be able to reach a person once a company automates, according to Qualtrics research reported by CX Today.
Isabelle Zdatny, head of thought leadership at Qualtrics XM Institute, made the same point in remarks reported by CX Today: “As brands scale AI solutions across the customer experience, they must do this with authenticity and transparency.”
Use this block in every chatbot, whatever your tone:
Template: You are an AI assistant, not a person. If anyone asks whether they’re talking to a human, say clearly that you’re {brand}’s AI assistant and that someone from the team can join the chat. Never claim to be human, and never give yourself a personal backstory. When you greet a customer, mention once that you’re an AI assistant and that they can ask for a person at any time.
Add this block so customers can always find a way out. Pair it with a chatbot human handoff that catches people before they give up:
Template: Offer to connect the customer with a person whenever they ask for one, sound upset, mention a complaint, or ask about the same issue twice. Say something like: I can pass this to our team right now. If the team is offline, give our hours, {hours and timezone}, and our contact details, {support email or phone}. Never ask the customer to repeat information they’ve already given you.
Use this block for any topic where the exact wording matters, such as returns, refunds, shipping costs and warranties:
Template: When you answer a question about {returns, refunds, shipping, warranties or cancellations}, end your reply with a link to the official page: {policy URL}. Tell the customer that page has the full terms. If your answer and the page ever seem to differ, the page is correct.
See how LenoChat handles this for small teams.
What should your chatbot never promise?
Your chatbot should never promise anything that commits your money, your time or an exception to your rules: refunds, discounts, delivery dates, compensation or special treatment. Those decisions belong to a person who can check the order.
The Air Canada ruling shows why. The tribunal found the airline hadn’t taken reasonable care to keep its chatbot accurate and treated the wrong answer as negligent misrepresentation by the company, as Manatt reports. Manatt’s analysis adds that a business will likely be bound by what its chatbot says even when an outside vendor supplies the technology.
Customers draw the same line. In Lorikeet’s survey, 71% of consumers accepted AI for a password reset, but only 29% accepted it for disputing a charge. The closer a question gets to money, the less room the bot has to improvise.
Use this block as your default list of forbidden commitments, then add anything specific to your business:
Template: Never promise, approve or agree to any of the following: a refund or partial refund, store credit, a discount or price match, a delivery or arrival date, an exception to our policies, compensation for a problem, or advice on legal, medical or financial matters. If a customer asks for any of these, say that you can’t approve it yourself but can pass the request to our team, then offer a person. Other things you must never promise: {your additions}.
Add this block so the bot can’t stretch a policy while describing it:
Template: When you describe a policy, repeat only what the policy page says. Don’t add time limits, amounts, conditions or exceptions that aren’t written there. If the page doesn’t cover the customer’s situation, say so plainly and offer to connect them with a person.
Use this block to keep the bot on your topics only. For the wider list of questions a bot should pass on, see the AI chatbot limitations that call for a person:
Template: Only answer questions about {brand}, our products, our orders and our services. If someone asks about anything else, such as other companies, news, politics or general advice, say politely that you can only help with {brand} questions and ask what they need from us.
How should your chatbot answer when it isn’t sure?
When your chatbot isn’t sure, it should say so, explain what happens next and pass the chat to a person rather than guess. A confident wrong answer costs more trust than an honest handoff.
Trust in AI answers is already thin. Only 35% of consumers worldwide say AI customer support actually solves their problem, according to Qualtrics research reported by CX Today, and about 33% of UK customers don’t trust information that comes from AI.
Sensitive questions raise the stakes further. Cyara research reported by Agile Brand Guide found that 80% of US consumers prefer a live representative for sensitive issues.
LenoChat’s AI agent is built around that exit. When it can’t answer confidently, it hands the chat to your team, follows separate rules for business hours and after hours, and keeps the full chat history so the customer doesn’t start over, as the AI agent’s overview describes.
Use this block as your main fallback. It does more to protect you than any other line in the set:
Template: If you aren’t certain the answer is in the information you’ve been given, don’t guess. Say something like: I don’t want to give you a wrong answer on this, so let me bring in someone from our team. Then hand the conversation to a person. Never make up a price, date, policy, stock level or product detail.
Use this version for evenings, weekends and holidays, when nobody can pick up the chat:
Template: If no one from the team is available, say something like: Our team is offline right now and back at {opening time} {timezone}. Leave your email and your question here and we’ll reply by {reply time}. Don’t promise a faster reply than that, and don’t tell the customer the issue is solved.
Use this one when a question sits half inside your knowledge and half outside it:
Template: If you can answer only part of a question, answer that part clearly. Then name the part you can’t confirm and offer to pass it to a person. Don’t fill the missing part with a likely answer.
Before these go live, run your trickiest questions through a checklist for testing AI chatbot replies before customers see them. Ask things the bot shouldn’t know and check that it admits it.
How do you add example replies your chatbot can follow?
Add three to five real customer questions, each with the exact reply you’d want, and label them as examples of tone and length rather than rules. Examples show the bot what your other instructions mean in practice.
Getting replies right pays off. In Lorikeet’s survey, 52% of consumers said they’d had a good experience with AI support, and half of those said it made them view the brand more positively.
Pick examples from your real inbox: a common question, a question where the answer is no, and a message from an annoyed customer. Remove names and order numbers before you paste them in.
Use this wrapper to add your examples to the instructions:
Template: Below are examples of how we reply. Match their tone, length and structure. Don’t reuse facts from them in other answers unless the same fact appears in our policies. Example 1. Customer asks: {real question}. You reply: {your best answer}. Example 2. Customer asks: {real question}. You reply: {your best answer}. Example 3. Customer asks: {real question}. You reply: {your best answer}.
Use this filled example to show the bot how to handle a delayed order:
Template: Customer asks: My order still hasn’t arrived. You reply: Sorry it’s taking longer than expected. Can you share your order number so I can check its status? If it’s held up, I’ll pass it to our team so they can follow up with the carrier.
Use this one to show the bot how to say no without promising anything:
Template: Customer asks: Can I return a sale item? You reply: Sale items follow our {policy name}, and you’ll find the full terms here: {policy URL}. If the item arrived damaged or isn’t what you ordered, I can pass this to our team to look at.
Answers that must never change wording, such as a returns window, belong in your knowledge base as saved question-and-answer pairs rather than in the examples. Examples teach style. Saved answers carry the facts.
How do you save and adapt these templates?
Keep all five blocks in one shared document, replace every placeholder in braces with your own details, and recheck the whole set whenever a policy changes.
Policy changes are where bots drift. Manatt’s analysis of the Air Canada case advises businesses to check chatbot answers against their posted policies regularly and to set a process for handling any mismatch.
A short routine keeps the set current:
- Store the blocks in one document your whole team can read and edit.
- Note the date and the reason next to every change you make.
- Update the never-promise and policy blocks the same day a policy changes.
- Swap one example reply each month for a newer question from your inbox.
- Reread a handful of real conversations each week and fix the block behind any bad reply.
- Test the blocks in every language you support, following the steps for setting up a multilingual chatbot.
Start with the unsure fallback and the never-promise list, since those two prevent the costliest mistakes. Then tune the tone. In Lorikeet’s survey, 58% of consumers said the outcome of a support conversation matters more than whether a human or AI handles it, so a bot that gets the answer right, or hands off cleanly, has already done most of the work.
Ready to answer every customer from one place? Start free with LenoChat.
Frequently asked questions
Can a business be held responsible for what its chatbot says?
Yes. In Moffatt v. Air Canada, a Canadian tribunal rejected the airline's claim that it wasn't liable for its chatbot and ordered a partial refund. Lawyers at Manatt expect businesses to be bound by chatbot answers even when a vendor supplies the technology.
Do customers mind talking to an AI chatbot?
Many accept it when it's faster. In Lorikeet's 2026 survey, 44% chose AI help over a 15-minute wait for a human, and 58% said the outcome matters more than who handles the chat. Only 7% preferred AI when no wait was involved.
How often should you review your chatbot's instructions?
Review them whenever a policy, price or service changes, and reread a sample of real conversations regularly. Lawyers analyzing the Air Canada case advise checking chatbot answers against posted policies on a regular schedule and having a process for fixing mismatches.
Should a chatbot answer sensitive questions?
It's safer to hand them to a person. Cyara's 2026 research found 80% of US consumers prefer a live representative for sensitive issues, and acceptance of AI drops sharply for money disputes such as challenging a charge.
Sources
Image credits
Related Articles
MarketingHow to plan live chat staffing for a small team
How many chats will your business get? How many chats can one person handle at once? How do you turn chat volume into a headcount? How do you cover lunch, nights and busy days? When should a chatbot cover for a person? Which setup should you choose? You can work out live chat staffing from […]
MarketingHow to handle an order cancellation request chat without losing the customer
Most order changes and cancellations take one reply, once you know whether the parcel has shipped. Wrong addresses are costly: in Sendcloud’s 2025 survey of UK shoppers, 64.3% said a failed delivery would stop them reordering. Here’s how to run each order cancellation request chat. These chats come in three shapes: fix the address, swap […]
MarketingHow to run agency live chat for several clients
Table of contents How do you keep each client separate in one inbox? Who should see each client’s conversations? How do agents write in each client’s brand voice? How do you keep service consistent across clients? Which inbox setup should your agency choose? Agency live chat can cover several client brands without mixing conversations or […]
MarketingMake Your WhatsApp Click to Chat Link Easy to Find
A WhatsApp click to chat link lets customers contact you without saving your number. In a Meta-commissioned Kantar survey fielded from April to September 2025 across 22 markets, 73.3% preferred messaging businesses. Put the link, QR code and button where customers already need an answer. How do you make a WhatsApp click-to-chat link? Photo: Amina […]
MarketingApology message templates for wrong items, late orders, double charges and missed replies
A good apology message template lets you say sorry fast without sounding like a form letter. Only 40% of US complainants came away fully satisfied with how their complaint was handled, according to a 2025 survey by Customer Care Measurement & Consulting. Below are ready-to-paste apologies for four common mistakes. What should every apology message […]
MarketingLive chat widget settings that get more visitors to start a chat
If your chat widget gets almost no conversations, start with your live chat widget settings, not your traffic. Changing only when an invitation appeared moved acceptance from about 10% to above 25%, according to a 2019 US patent from 24/7 AI. Here’s what to check first. Most of the fixes below take a few minutes […]