CentriWebAI hallucinations are confident answers that are wrong or invented: a policy you do not have, a price you never set, a case that does not exist. You cannot switch them off, but you can cut them sharply by giving the AI your own facts to answer from, asking it to quote them, letting it say it does not know, and having a person check anything that matters before it leaves the building.
Vendors say the same thing. Anthropic’s own guidance for reducing hallucinations lists these techniques and then adds that they reduce the problem without eliminating it, so critical information should still be validated (Anthropic documentation, opened November 2026).
Why AI makes things up
A chat AI writes the most likely next words given what you asked and what it learned in training. It is not looking a fact up in a database and reporting back. When the answer is well covered in its training, the likely words are usually right. When it is thin, specific to your business, or recent, the likely words can be fluent and wrong, and they sound just as sure as the right ones.
That is why the same tool can describe the general rules of a trade well and then invent your opening hours. The court system has noticed too. The NSW Supreme Court’s practice note on generative AI warns lawyers about “hallucinations”, describing them as apparently plausible, authoritative and coherent responses that are in fact inaccurate or fictitious, with false citations and fabricated references as examples (Practice Note SC Gen 23, opened November 2026). It also notes that the data behind a tool may be out of date or not relevant to New South Wales or Australia, which is a common cause of wrong answers for Australian businesses.
What a wrong answer costs a small business
The damage depends on where the AI sits.
- Internal drafting. A wrong sentence in a draft is caught by whoever reads it. Cheap.
- Customer replies. A chatbot that invents a refund policy, a service you do not offer or a booking time you cannot meet creates a promise your team now has to honour or walk back.
- Advice and compliance. A wrong rule, deadline or figure that a client acts on is the expensive end, and your name is on it.
So the useful question is not whether the AI ever makes things up. It is how much a wrong answer costs in this particular job, and what stops it reaching a customer.
Ways to reduce AI hallucinations
These are practical changes you can make this week, whether you use a chat tool yourself or have an assistant built for your business.
1. Give it your facts and tell it to use only those
Paste in or connect your price list, service descriptions, policies and opening hours, then instruct the AI to answer only from that material. Anthropic’s guidance lists restricting the AI to the documents you provide, rather than its general knowledge, as a technique in its own right. This is the single biggest lever for a business assistant, because most invented answers happen when it has nothing of yours to draw on and fills the gap.
2. Ask it to quote before it answers
For a long document such as a contract, a supplier terms sheet or a policy manual, ask the AI to pull out the exact words that matter first, then answer using only those quotes. The same guidance describes this for long documents, and it works for a practical reason: you can read the quote against the source in ten seconds, where checking a paraphrase takes minutes.
3. Give it permission to say “I do not know”
By default an AI tends to try. Add a line such as “If the material does not answer this, say so and do not guess.” The documentation lists allowing the AI to admit uncertainty as a basic strategy that can drastically reduce false information. For a customer-facing assistant, pair it with an instruction on what to do instead: offer to take a message, or pass the conversation to a person.
4. Make it cite, then drop what it cannot support
Ask for a source quote against each claim in a draft, then tell it to remove any claim it cannot back up. This turns a block of text into something auditable. It does not prove the claims are true, because the source document could itself be out of date, but it catches the invented ones. One practical fix is to limit answers to your own files, as described in our guide to an AI knowledge base your staff can ask.
5. Keep the job narrow
An AI told to “handle customer questions” will wander. An AI told to answer questions about three services, take a booking or pass on a message has far less room to invent. Narrow jobs also make mistakes easier to spot in testing.
6. Ask twice and compare
The documentation also suggests running the same question several times and comparing the answers, since differences between runs can signal a made-up detail. It suits one-off research, not every customer message.
What a person should always check
Some content should never go out on the AI’s word alone, however well it is set up.
| Check by a person | Why |
|---|---|
| Prices, quotes, fees and discounts | A wrong figure becomes a commitment |
| Legal, tax, medical or financial statements | You or your staff carry the professional duty, not the tool |
| Names, dates, case numbers and citations | Fabricated references are the classic failure |
| Anything quoted from a document | Compare it to the original, not to the AI’s summary |
| Advertising claims and testimonials | Regulators hold the business responsible for what it publishes |
Checking is part of the job, not a sign the tool failed. The NSW practice note, for example, bars generative AI from generating the content of affidavits and witness statements, and allows uses such as chronologies, indexes and summaries of documents, with people responsible for the result.
Test your own AI before customers do
If you already have a chatbot, receptionist or email assistant, test it the way a difficult customer would.
- Write twenty real questions from your enquiry inbox, including ones that your information does not cover.
- Add ten traps: a service you do not offer, a discount you never gave, a question about a competitor, a request for legal or medical advice.
- Read every answer against your own facts and mark each as right, wrong, or sensibly declined.
- Fix wrong answers by changing the source material or the instructions, not by hoping.
- Repeat after any change to your prices or policies, and every few months regardless.
The ten traps matter most. A good assistant declines or hands over on all of them. If it cheerfully answers a question about a service you do not offer, it needs tighter instructions before it talks to the public.
When the AI talks to customers
For anything customer-facing, treat a handover to a person as part of the design. The assistant should answer from your information, say when it cannot, and pass the conversation across with the details already captured. CentriWeb’s AI work is built this way: it answers from the business’s own information, books into the real calendar and hands over to a person when it should. If you are weighing a website chat tool, this chatbot guide covers what to ask a supplier.
Also decide what your staff may paste into public tools. Wrong answers are one risk; confidential customer details leaving the business is another, covered in using ChatGPT safely in a small business and in an AI policy template.
A five-line rule for staff
- AI can draft. A person sends.
- Anything with a number, a date, a name or a rule gets checked against the original.
- If the AI cannot show where an answer came from, treat it as a guess.
- Customer-facing AI answers only from our own material and hands over when unsure.
- When it gets something wrong, tell whoever looks after the setup so the source material is fixed.
Hallucinations will not disappear, and no tool or supplier can promise they will. What you can control is how much of your business depends on a single unchecked answer.