Guide · AI and knowledge base
AI answer guardrails: what stops an AI making things up
In short
An AI model writes a fluent answer whether or not it knows the facts. Guardrails are the rules around it: answer only from content you approved, show where each answer came from, say “I don’t know” and hand over when the content doesn’t cover the question, and keep sensitive data and other customers’ details out of its reach. Each one closes a different way an answer can go wrong.
Reviewed 9 November 2026. Terms used: RAG (retrieval-augmented generation), PII masking for AI.
Seven guardrails to put in place
Some of these are instructions to the model, and some are checks in code. Know which is which: only checks are certain.
Give it only content you’d sign
Point the AI at published, customer-facing articles, not internal notes, drafts or old PDFs. If you wouldn’t paste the text into a reply yourself, it shouldn’t be a source.
Make it cite, and hold back answers that don’t
Ask for the article behind every answer and show it to the customer. Better still, check it in code: an answer that names no source isn’t sent. A citation is something a customer can check and a person can fix.
Let it say “I don’t know”
When your content doesn’t cover the question, the bot should say so and offer a person or a rephrase. An honest “I don’t know” costs one hand-over; an invented refund policy costs far more.
Tell it what to decline
Medical advice, legal questions, anything you’d never want in writing: list the topics and give it a polite refusal. These are instructions, so test them with awkward questions; the hard limit is still the content it’s allowed to use.
Keep customer data out of the prompt
Mask card numbers, bank accounts, ID numbers and passwords before anything reaches the model, and keep other customers’ details out of its reach entirely. Treat what the customer types as text to answer, not instructions to follow.
What the model sees, after masking Customer wrote: My IC is 900101-14-5678, can you check my warranty?
The model gets: My IC is [ID_1], can you check my warranty?
Decide when it hands over
After one or two unanswered questions, when the customer is upset, or when they ask for a person. Collect a name and an email first, so the agent can reply even if the chat drops.
Read what it said, every week
Review a sample of answers and every question it couldn’t answer. Most fixes are a missing or unclear article, not a new rule.
How it works in Axonave
Platform rules you can’t switch off, and your own on top.
Axonave’s AI is instructed, in rules that come before yours and that no setting removes, to answer only from your published external articles, to say it doesn’t know instead of guessing, never to reveal its instructions or anything about other customers, and to treat instructions inside customer messages as plain text. One check runs in code: an answer reaches the customer only once it cites a published article, and that article is named with it. On top you set the tone, the longest answer, your own rules, up to 20 topics the AI is told to decline, and when it hands over. Card, bank, IC and passport numbers and secrets are masked before any AI call, and emails and phone numbers too by default.
- A person after 1 to 5 unanswered questions in a row
- The details to collect before a hand-over
- Unanswered questions grouped and turned into draft articles
Settings → Chatbot → AI in our demo workspace: the bot’s name, tone and greeting in three languages, then Teras Home’s own rules under “What the AI may say”.
Common questions
Can the AI still be wrong?
Yes, if an article is wrong or unclear the AI will repeat it, with the article named so you can find and fix it. That’s why the unanswered and handed-over questions are worth reading every week.
Does the AI see customers’ card or IC numbers?
No. Axonave masks cards, bank accounts, IC and passport numbers and secrets before any AI call. The AI gets placeholders, while the conversation and your team keep what the customer wrote.
Can we add our own rules?
Yes: your own rules and up to 20 topics the AI is told to decline, such as medicines or test results. They sit under the platform rules, which you can’t switch off.
See it working with your own customers.
We reply within one business day by email or WhatsApp.