What is an AI hallucination?
Sooner or later, every owner who uses AI sees it write something that is simply not true: a price you never charged, a policy you never had, a delivery date nobody agreed. People call this a hallucination. It sounds alarming, but it is predictable, and most of it can be prevented before a customer ever sees it. Here is what a hallucination is, why it happens, and the guards that work for small businesses.
In one sentence
An AI hallucination is when an AI agent states something as fact that is not true, such as an invented price, policy or opening time, usually because it was asked something its notes did not cover and filled the gap with a plausible guess. You guard against it by giving the agent the real facts, telling it what to do when it does not know, and holding the riskiest replies for a person.
In more detail
A language model writes by choosing what is most likely to come next. Most of the time, the most likely next words are also the true ones, especially when the facts are right there in its notes. When the facts are missing, it still writes the most likely-sounding answer. If most shops offer free returns within 30 days, it may tell your customer you do too, whether or not that is your policy.
The word "hallucination" makes it sound as if the AI is seeing things. It is closer to a confident new member of staff who would rather give an answer than admit they do not know. The answer is fluent, friendly and wrong, and that is what makes it risky. A customer has no way to tell the difference.
Hallucinations are not random. They cluster around a few kinds of question: anything with a number, anything about rules or policies, anything about the future, and anything that sounds like it should have a standard answer. Knowing that tells you where to put your effort.
Better models hallucinate less, and none of them hallucinate never. The aim is not to find a perfect model. It is to build a set-up where a guess is unlikely, an honest "I will check" is the easy path, and the replies that could cost you money or trust are seen by a person before they go.
The example below shows the same refund question before and after a few simple guards were added. Names and details are examples.
An example
| Who | Message |
|---|---|
| Customer (Ben, on Messenger) | I bought a jacket two weeks ago and it doesn't fit. Can I get a refund? |
| Before: no refund rule in the notes | Hi Ben, of course! We offer full refunds within 60 days, no questions asked. Just send it back and we will refund you straight away. |
| After: rule says refunds go to a person | Hi Ben, sorry it does not fit. I have passed this to Laura, who handles returns, and she will reply today with the next steps. Could you send your order number in the meantime? |
The kinds of hallucination that cause trouble
| Kind | What it looks like | Why it is risky |
|---|---|---|
| Invented prices | "A full service is about £120." | The customer holds you to it, or feels misled when the bill is different |
| Invented policies | "You can cancel any time up to an hour before." | It becomes a promise you have to keep or argue about |
| Invented availability | "Yes, we have that in stock in a size 10." | A wasted trip, a cancelled order and a lost customer |
| Invented facts about the customer | "I can see your order shipped yesterday." | The agent cannot see orders unless it has a tool for them |
| Invented links or numbers | A booking link or phone number that does not exist | The customer gets stuck, or reaches someone else |
| Invented commitments | "I will make sure the manager calls you tonight." | Nobody knows the promise was made |
Look at your own messages for a week and you will probably find most questions fall into a handful of these. Those are the ones to cover first.
Five guards that work
- Write the facts down. Hours, prices, services, areas, policies and links. Every fact in the notes is one fewer gap to fill.
- Give it an honest way out. Tell it plainly: if the answer is not in your notes, say you will check and a person will reply. Agents follow this rule well when it is written clearly.
- List what it must never promise. Refunds, discounts, delivery dates, medical or legal answers, anything you decide case by case.
- Hand the hard ones to a person. Complaints, refunds and anything unusual should leave the agent's hands quickly, with a short holding reply.
- Hold the risky replies as drafts. For prices and anything involving money, have the agent write the reply and wait for you to approve it, at least until you trust it.
Sample guard lines for your notes
These lines can be copied into the notes your agent reads before replying. Adjust the details to your business. Names and details are examples.
Only state prices, times, policies and links that appear in these notes.
If a customer asks something these notes do not answer, reply:
"Good question, let me check that for you. Someone from the team will
reply shortly." Then hand the chat to a person.
Never promise: refunds, discounts, delivery dates, call-backs at a set
time, or anything about health.
Never say you can see an order, booking or payment unless a tool has just
shown it to you.
If you are unsure whether something is true, say so. A short honest answer
is always better than a confident guess.How to catch hallucinations early
- Read the agent's replies every day for the first few weeks. Most hallucinations show up in the first handful of unusual questions.
- Test it yourself. Message your own account with awkward questions: a service you do not offer, a price you never set, a policy you do not have. See what it says.
- Search the conversations for numbers and promises, such as "£", "refund", "guarantee" and "tomorrow". These are where invented answers hide.
- Every time you find one, fix the cause in the notes, not just the single reply. The same gap will come up again.
Common mistakes
- Assuming a newer model fixes it. Newer models help, and the gaps in your notes still need filling.
- Writing "be accurate" in the notes. It sounds sensible, but it does not tell the agent what to do when it lacks a fact. Tell it exactly what to say instead.
- Letting the agent answer from your website by memory. If a fact matters, put it in the notes the agent reads.
- Correcting the customer but not the notes. The next customer gets the same wrong answer.
- Trusting a reply because it sounds confident. Confidence is not a sign of accuracy in an AI agent.
Why it matters for your business
- A single invented price or policy can cost more than months of saved time, in refunds, arguments or reviews.
- Customers cannot tell a guess from a fact, so the responsibility for accuracy sits with your set-up.
- Most hallucinations come from a few predictable gaps, so a small amount of work in the notes removes most of the risk.
- An honest "let me check" keeps trust far better than a fast wrong answer.
- Holding money and policy replies for a person lets you use AI for speed without betting the business on every word.
How Spaces handles it
Spaces is built so the guards live in one place. The business brief in your workspace holds your facts, common questions and answers, tone, a list of things never to promise and a rule for when to hand over. Your agent is told to read it before replying, and every agent you connect reads the same brief, so a fix reaches all of them at once.
When the agent hands a conversation to a person, it stays out of that chat until the chat is handed back: no replies, scheduled messages or routines there. That stops it adding a second guess after you have taken over.
For replies you want to see first, ask your AI agent to hold replies as drafts, and nothing goes out until you approve it. You can approve a draft as written, send an edited version, or reject it with a reason. If you ever need everything to stop while you sort something out, Pause everything in Settings stops every send at once while reading carries on.
In short
A hallucination is a confident guess dressed as a fact. It happens when the agent is asked something its notes do not cover.
Fill the gaps that matter, give it an honest way out, hand the hard ones to a person and hold money replies as drafts. With those four habits, hallucinations become rare and harmless rather than a reason not to use AI at all.
FAQ
Why does AI make things up?
Because a language model writes the most likely-sounding answer, and when the real fact is missing, the most likely answer may not be true. It does not know it is guessing. Giving it the facts and a rule for what to do when it does not know prevents most of it.
Can hallucinations be stopped completely?
Not completely. They can be made rare and caught before they reach a customer. Clear notes, a "let me check" rule, hand-over for hard questions and drafts for risky replies together remove most of the risk.
Which questions are most likely to cause a hallucination?
Questions with numbers, such as prices and times, questions about policies, questions about the future, such as delivery dates, and questions about things the agent cannot see, such as a specific order. Cover those first in your notes.
Is a hallucination the same as the AI lying?
No. Lying means knowing the truth and saying something else. The agent is not aware it is wrong. That is why the fix is better information and rules, not telling it to be honest.
What should I do if a hallucination reaches a customer?
Correct it quickly and kindly in the same chat, ideally yourself, then fix the gap in the notes so it cannot happen again. Our guide on what to do when your AI agent gives a wrong answer walks through it step by step.
Will holding replies as drafts slow my customers down?
It can, because a held reply only goes out when someone approves it. Many owners hold only the risky kinds, such as prices and refunds, and let simple answers go straight out.
Does telling the agent to "be accurate" help?
Very little on its own. It is much better to tell the agent exactly what to do when it lacks a fact, such as saying it will check and handing the chat to a person.
Does Spaces check the facts in my agent's replies?
No. Spaces does not judge what your agent writes. It gives the agent your business brief to read, lets it hand chats to a person, holds replies as drafts when you ask, and keeps every send inside the account limits.