An AI hallucination is when an AI tool produces information that sounds right but isn't: an invented statistic, a policy that doesn't exist, a made-up citation, a wrong price. It happens because language models generate likely-sounding text rather than looking facts up the way a database does.
Newer models hallucinate less, but none are immune. For a business, the question isn't whether your AI will ever be wrong. It's whether wrong answers reach customers.
Why it matters for businesses
There are real cases where businesses were held to promises their chatbots made, including a well-known Canadian tribunal decision in 2024 that required an airline to honour a refund policy its chatbot invented. The principle is simple: what your AI says on your behalf is treated as what you said.
| Risk | Example |
|---|---|
| Wrong prices or discounts | Bot offers a discount that doesn't exist |
| Invented policies | Bot promises refunds outside your terms |
| Legal or compliance errors | AI-written content misstates tax rules |
| Fake sources | Report cites studies that don't exist |
| Reputation damage | Screenshots of a bad answer shared online |
Why AI makes things up
- It predicts text, not truth. When it lacks information, it fills gaps with plausible words.
- Vague questions invite confident guesses.
- Missing context: if it doesn't have your price list, it may guess one.
- Pressure to answer: many systems are tuned to be helpful rather than to say "I don't know".
- Out-of-date knowledge: models trained months ago don't know recent changes.
Seven ways to reduce hallucinations
1. Ground answers in your own documents
Instead of letting the AI answer from general knowledge, connect it to your approved sources (help articles, policies, price lists, product data) and instruct it to answer only from them. This approach is called retrieval-augmented generation. Read RAG explained for business.
2. Tell it to say "I don't know"
Clear instructions help: "If the answer isn't in the provided documents, say you don't know and offer to connect the customer with the team." Then test that it actually does.
3. Limit the scope
A support bot that only handles orders, delivery and returns is far safer than one that will chat about anything. Refuse off-topic questions politely.
4. Keep humans on high-stakes outputs
Anything involving money, legal commitments, health, safety or public statements should be reviewed by a person before it goes out. Use AI to draft, not to decide.
5. Show sources
When AI answers from your documents, have it cite the article or policy it used. Customers trust it more, and you can check it quickly.
6. Keep your knowledge base current
Most "hallucinations" in business bots are actually outdated or missing information. Update prices, policies and product details the day they change.
7. Test and monitor continuously
- Build a test list of 50 real customer questions with correct answers and run it after every change.
- Review a sample of real conversations weekly.
- Track handover rates and complaints.
- Ask staff to flag any wrong answer they see.
For AI-written content
If you use AI to draft articles, reports or marketing:
- Check every fact, figure and name against a reliable source.
- Never publish AI-generated citations without opening each one. Invented references are common.
- Have an expert review anything technical, legal, financial or medical.
- Be specific in prompts and provide source material rather than asking the AI to "research" from memory. See prompt writing for business teams.
Our AI-assisted content workflow builds these checks in.
A quick risk matrix
| Task | Hallucination risk | Safeguard |
|---|---|---|
| Summarising your own meeting | Low | Quick read-through |
| Drafting a social post | Low–medium | Human edit |
| Answering FAQs from your help centre | Medium | Grounding + "don't know" rule + monitoring |
| Writing a blog post with statistics | Medium–high | Fact-check every claim |
| Quoting prices to customers | High | Pull from live price data only, never generate |
| Legal, tax or medical guidance | High | Expert review, or don't use AI for it |
Choosing tools with fewer hallucinations
When you evaluate AI products, ask:
- Does it answer only from my content, or from general knowledge too?
- Can it cite sources for each answer?
- Can I control what it refuses?
- Does it hand over to a person, and when?
- Can I review conversations and correct answers?
Zelvorm Assist, for example, answers from your own website and documents, shows sources and hands over to your team when it isn't sure.
Disclosure helps too
Telling customers they're talking to an AI sets expectations and, for EU customers, is a legal requirement under the EU AI Act's transparency rules. It doesn't remove your responsibility for wrong answers, though.
The short version
AI hallucinations happen because models generate likely text, not verified facts. You're responsible for what your AI tells customers. Ground answers in your own documents, instruct it to admit uncertainty, narrow its scope, keep humans on high-stakes outputs, show sources, keep information current and test continuously.
Building something customer-facing? Our AI automation team designs AI systems with these safeguards from day one.




