In 2022, a man bought a full-price Air Canada ticket for his grandmother's funeral after the airline's chatbot told him he could claim a partial bereavement refund within 90 days of flying. That policy didn't exist. When he applied, Air Canada refused, then argued in tribunal that the chatbot was a separate entity responsible for its own words. The tribunal called that argument remarkable, found the airline liable for negligent misrepresentation, and ordered it to pay.
That case is why a lot of small business owners are nervous about AI support, and the fear is reasonable: a bot that invents a policy, promises a refund you don't offer, or makes up a feature isn't a convenience, it's a liability.
But the thing that failed there isn't what most modern AI support tools actually do. The fear is aimed at one specific kind of AI. Once you understand the difference, "will it make things up?" stops being a yes or no question and becomes something you control.
In this guide, you'll learn:
- What an AI "hallucination" actually is
- Why general-purpose AI invents answers and grounded AI doesn't
- The mechanisms that keep a support bot honest
- Where AI can still get it wrong
- A practical checklist to keep your bot accurate
What a "Hallucination" Actually Is
A hallucination is when an AI produces an answer that sounds completely confident and fluent but is factually wrong. Not a typo, not a "sorry, I don't know," a clean, well-written, totally made-up answer.
It happens because of how language models work: they're trained to produce the most plausible-sounding next words. They're brilliant at sounding right, with no built-in sense of whether they are right, unless you give them a way to check. An AI left to answer from general training fills gaps with plausible fiction. An AI forced to answer only from an approved source tells you when that source comes up empty. Same technology, completely different behavior.
Why General AI Makes Things Up (and Grounded AI Doesn't)
Picture two new employees. The first never read your docs. Asked "Do you ship to Norway?", they don't want to look unhelpful, so they guess: "Yes, 3 to 5 days." Confident. Wrong. The second has your shipping page open and can only answer from it. Same question: "I don't see Norway in our shipping zones, let me get someone to confirm." Less flashy. Correct.
That's the difference between general-purpose AI and a grounded one. A grounded support chatbot doesn't answer from the open internet or its imagination. It answers from your content, the pages it crawled from your site plus any documents you uploaded, retrieving the relevant passages and building its answer from them. If the answer isn't there, a well-built bot says so instead of inventing one.
| General-purpose AI | Grounded support AI | |
|---|---|---|
| Source of answers | Everything it was trained on | Only your website + your documents |
| When it doesn't know | Guesses fluently | Says it doesn't know, hands off |
| Who controls the facts | The model | You |
| Risk of inventing policy | High | Low, by design |
| Best for | Open-ended conversation | Answering for your business |
A support AI grounded in a single source of truth, the one you maintain, can't invent a bereavement policy, because the only policy it can describe is the one you wrote down.
The Mechanisms That Keep a Support Bot Honest
Grounding is the foundation; a trustworthy bot stacks a few safeguards on top:
- Answers only from your knowledge base. Its entire world is the content you gave it. Ask Namiru's agent about quantum physics and it won't bite; ask about your return window and it answers from your actual return policy.
- Says "I don't know" instead of guessing. The single most important behavior a support bot can have. When it can't answer from your content, it tells the visitor honestly and points them to a human.
- You own and edit the source of truth. After the crawl you see exactly what it learned, entry by entry. Fix an entry and you fix the bot, with no hidden conversation trees drifting out of sync.
- Escalates to a human when it matters. Can't answer, or the visitor asks for a person? The conversation lands in your inbox with full context.
- Flags its own gaps. When the dashboard shows "14 visitors asked about Switzerland shipping and the AI couldn't answer," that's a precise instruction for what to add next. Together these turn accuracy from a hope into a process: the bot stays inside the lines you draw, tells you where they're too narrow, and hands off whatever it shouldn't decide alone.
Where AI Can Still Get It Wrong
No serious tool should promise perfection, and anyone who does is selling you something. Grounding dramatically reduces hallucinations; it doesn't make a bot omniscient.
- Garbage in, garbage out. If your site is vague, outdated, or contradicts itself, the AI inherits all of it. One page says 14-day returns, another says 30, and the bot might quote either. The fix is a clean source of truth, which is why the post-crawl review matters.
- Missing information. The bot can't answer what you never wrote down. If your shipping page doesn't mention Norway, the best it can do is admit it doesn't know, correct behavior, but only as useful as your willingness to then add Norway. For an online store, those silent gaps are the kind that quietly cost you sales, a visitor with one unanswered question just closes the tab.
- Ambiguous or adversarial questions. Vague phrasing gets vague answers, and a few people will try to trick a bot into saying something silly, the way one dealership's bot was talked into "selling" a car for a dollar. Grounding limits the blast radius, but public bots should be monitored.
- High-stakes content. Where exact wording carries legal or financial weight, keep a human review step. AI handles the 80 to 90 percent of routine questions that make up most support volume; the rest is what the handoff is for. The honest summary: a grounded support AI is far more trustworthy than a general chatbot, and it's a tool you're responsible for, not an oracle. The businesses that get burned deployed and walked away. The ones that do well treat the first few weeks as a tuning period.
How to Keep Your Bot Accurate: The Checklist
None of this is hard, and most of it is a one-time setup plus light upkeep.
| Step | What to do | Why it matters |
|---|---|---|
| 1. Clean your source | Make sure pricing, shipping, returns, and policies are clear and current on your site | The bot can only be as accurate as what it reads |
| 2. Review the crawl | After the AI builds its knowledge base, read through the entries | Catch anything misread, outdated, or contradictory before customers do |
| 3. Test real questions | Ask your bot the 15-20 questions you actually get most | Reveals gaps and wrong answers while it's still private |
| 4. Watch unanswered questions | Check the dashboard for what the AI couldn't handle | Each one is a precise gap to fill |
| 5. Set up handoff | Confirm escalation goes to an inbox you check | The safety net for everything outside the bot's scope |
| 6. Re-crawl on changes | New product, new price, new policy? Re-crawl | Keeps the single source of truth in sync |
| 7. Remove contradictions | One answer per question across your whole site | Eliminates the most common cause of inconsistent replies |
Do steps 1–5 before you go live and 6–7 as normal upkeep, and you've removed the large majority of ways a support bot can embarrass you.
How Namiru Is Built to Prevent This
Namiru.ai was designed around grounding from the start. The agent answers only from the knowledge base it builds from your website and uploaded files, never the open web. When it doesn't have an answer, it says so and routes the visitor to you by email. You can edit, add, or delete knowledge entries any time, so the source of truth stays yours and current with one re-crawl. And because every conversation is analyzed and tagged, the same conversation intelligence that powers pain-point detection and lead capture doubles as your accuracy monitor, surfacing the questions it couldn't answer so accuracy compounds over time instead of decaying.
The Bottom Line
"Will it make things up?" is the right question to ask, and the Air Canada case proved the stakes: your business is responsible for what your bot says, full stop. But the answer isn't to avoid AI support. It's to use the kind that's grounded in your own content, honest about what it doesn't know, editable when it's wrong, and backed by a human handoff for anything that matters.
A general chatbot guesses. A grounded support agent checks, answers from your facts, and admits the gaps so you can close them. Set it up properly, watch the dashboard for the first few weeks, and you get accurate 24/7 support without the risk that keeps everyone up at night.
Deploy a Support Agent You Can Actually Trust
Namiru.ai gives small businesses AI support built for accuracy, not improvisation.
- Paste your URL and get a trained AI agent in 30 seconds
- Answers grounded only in your website and documents, never the open web
- Says "I don't know" and hands off to you instead of inventing answers
- Editable knowledge base, so the source of truth is always yours
- A dashboard that flags every question the AI couldn't answer
- 200+ languages with automatic detection
- Flat pricing, no per-seat or per-resolution fees
- EU-hosted infrastructure
- Free plan with 50 conversations per month, no credit card required Your bot should never have to guess. Give it your facts, and it won't.
