# How do you stop a support AI from making things up? > Why support chatbots invent answers, and the guardrails that make it rare: grounding, hand-over rules, limits on promises and weekly reviews. Source: https://morfalabs.com/en/blog/stop-ai-chatbot-hallucinations _AI agents · 3 min read · Published October 30, 2026_ Every company that has put an AI in front of customers seems to have a story about it confidently inventing a refund policy. You can't make a language model perfect. You can make made-up answers rare, and make sure the ones that slip through do no damage. ## Short answer Ground every answer in your own approved content, let the AI say "I don't know" and hand over when that content doesn't cover the question, limit what it's allowed to promise, log the source behind every reply, and read a sample of real conversations every week. Test with real past questions before launch and after every change. ## Why support bots invent answers A language model writes the most likely next words. Ask it about your return policy when it has never seen your return policy, and it will write one. A plausible one. Usually a generous one. So most hallucinations in support aren't mysterious. The bot was asked something it had no facts for, and nothing told it to stop. A generic bot with a friendly prompt and no access to your documents is set up to fail in exactly this way. **Example: the same question, with and without grounding** > Can I return a sale item after 30 days? - **Ungrounded bot:** Of course! All items can be returned within 60 days for a full refund. _(Invented policy, no source)_ - **Grounded agent:** Sale items can be returned within 14 days for store credit. Your order is from 32 days ago, so I've passed this to our team and they'll reply today. _(Source: returns policy, section 2. Handed to a person.)_ ## Five guardrails that work None of these is exotic. Together they turn a bot that improvises into one that looks things up and knows when to stop. **The guardrails, in the order we set them up** 1. **Ground the answers:** The agent answers from your approved policies, catalog and order data, not from general knowledge. 2. **Allow "I don't know":** When the content doesn't cover a question, it says so and hands over instead of guessing. 3. **Limit promises:** Refunds, discounts and delivery dates only within rules you set. Anything else needs a person. 4. **Keep the source:** Each reply is logged with the document or record it came from, so you can check it. 5. **Review every week:** Read a sample of real chats, fix the gaps in your content and add the hard cases to your tests. ## Test it before your customers do Collect a few hundred real questions from your inbox or chat history, including the awkward ones: angry customers, messages that are half English and half Arabic, questions about products you stopped selling. Write down what a good answer looks like for each. Run the agent against that set before launch and again after every change to its content or rules. It's a boring step. It also catches most of the embarrassing answers before a customer ever sees them. ## When it still gets one wrong It will, now and then. Treat each wrong answer as a bug report about your content: find the document that was missing or unclear, fix it, and add the question to your test set so it stays fixed. Then tell the customer the right answer. People forgive a mistake that gets corrected quickly far more easily than one that gets defended. ## Questions people also ask ### Can AI hallucinations be removed completely? No. A language model can always produce a wrong sentence. Grounding, hand-over rules and testing make it rare, and limits on what the agent may promise keep the rare mistakes cheap. ### Do we need our own AI model to fix this? Almost never. The problem is usually missing facts and missing rules, not the model. A leading model connected to your own content, with clear limits, is the usual answer. ### How often should someone review the conversations? Weekly for the first couple of months, while you're still finding gaps in the content. After that, a monthly review plus an alert for conversations a customer rated badly is usually enough. ## Related service [AI Agents & Assistants](https://morfalabs.com/en/services/ai-agents.md): Custom AI agents and chatbots for WhatsApp, web chat and staff. They answer from your own documents in English and Arabic and hand tricky cases to people. ## Worried about what your support bot is telling customers? - Email: [info@morfalabs.com](mailto:info@morfalabs.com) - Web: https://morfalabs.com ## More articles - [AI and blockchain by 2030](https://morfalabs.com/en/blog/ai-and-blockchain-future.md) - [AI in business by 2030](https://morfalabs.com/en/blog/future-of-ai-in-business.md) - [Are AI agents worth it?](https://morfalabs.com/en/blog/are-ai-agents-worth-it.md) - [Zapier, n8n or a custom build?](https://morfalabs.com/en/blog/zapier-n8n-or-custom-automation.md) - [Invoice entry without templates](https://morfalabs.com/en/blog/automate-invoice-entry-without-templates.md) - [Choosing a software agency](https://morfalabs.com/en/blog/choose-a-software-agency.md) - [Getting recommended by ChatGPT](https://morfalabs.com/en/blog/get-recommended-by-chatgpt.md) - [Do you need a blockchain?](https://morfalabs.com/en/blog/does-your-business-need-a-blockchain.md) - [Is company data safe in ChatGPT?](https://morfalabs.com/en/blog/is-company-data-safe-in-chatgpt.md) - [New ERP or better integrations?](https://morfalabs.com/en/blog/new-erp-or-better-integrations.md) - [Excel or a live dashboard?](https://morfalabs.com/en/blog/excel-reports-vs-live-dashboard.md) - [Will this automation pay off?](https://morfalabs.com/en/blog/automation-roi-before-you-build.md) - [Is your data ready for AI?](https://morfalabs.com/en/blog/is-your-data-ready-for-ai.md) - [Native, cross-platform or web?](https://morfalabs.com/en/blog/native-cross-platform-or-web-app.md) - [Agency or in-house team?](https://morfalabs.com/en/blog/agency-or-in-house-team.md) - [Writing a software brief](https://morfalabs.com/en/blog/write-a-software-brief.md) - [SEO vs GEO, side by side](https://morfalabs.com/en/blog/seo-vs-geo.md) ## Pages - [Insights](https://morfalabs.com/en/blog.md) ## Other languages - [العربية](https://morfalabs.com/ar/blog/stop-ai-chatbot-hallucinations.md) - [Français](https://morfalabs.com/fr/blog/stop-ai-chatbot-hallucinations.md) - [Español](https://morfalabs.com/es/blog/stop-ai-chatbot-hallucinations.md) - [Deutsch](https://morfalabs.com/de/blog/stop-ai-chatbot-hallucinations.md)