The risk is real, but manageable. An assistant's wrong answer usually comes from guessing about something it does not know. The solution is to draw a clear line from the start around what the assistant knows and does not know, and to have it hand over to a person when it does not know.
Companies thinking about putting an AI assistant on their site usually ask the same first question: “What if it tells a customer something wrong?” It is a fair question. AI models can produce a fluent, confident sentence even on a subject they know nothing about. In a chat app that is annoying; in an assistant speaking on behalf of your company it means a wrong price, a service you do not offer or the wrong opening hours.
The good news is that this risk varies greatly depending on how the assistant is set up. Below we explain where wrong answers come from and how the line is drawn.
Where do wrong answers come from?
A company assistant's wrong answer usually comes from one of three sources:
- The information was never given. If the assistant has not been told the delivery time and is allowed to guess, it can produce an answer like “usually 2-3 days” that sounds reasonable but is wrong for you.
- The information is out of date. Your opening hours have changed, but the assistant still knows the old ones. The model is working correctly; the information it has is wrong.
- The question is ambiguous. Which service is “How much do you charge?” about? If the assistant gives a figure without clarifying the question, it may match it to the wrong thing.
Drawing the line: three rules
1. Speak only from the company's information
The assistant's source of information is limited to what you give it: services, prices, opening hours, address, frequently asked questions. It does not answer with general information from the internet. This single rule keeps another company's price or an old news story from creeping into an answer. The AI Customer Assistant works this way: it answers only with the information you provide.
2. Say when it does not know
The most important behaviour is that the assistant does not guess when it does not know the answer. If the answer is not in the information provided, the assistant says “we will get back to you” and takes the customer's phone number, or points a customer who wants an appointment to the booking system. For the customer, “I don't know that, we'll call you” is always better than a wrong answer.
3. Log the conversations and read them
Without knowing what the assistant says, you cannot tell whether the line is working. Conversation summaries should reach you regularly. They serve two purposes: you spot a wrong or incomplete answer early, and you learn what your customers ask most. With the assistant we set up, summaries arrive in your inbox; when your information changes, we update the assistant.
An assistant is most trustworthy on the day its information was last updated. When a price, a promotion or your opening hours change, the assistant's information should change the same day. Make this one person's responsibility.
An example: the same question, two different assistants
Take a veterinary clinic. A customer writes late at night: “My dog has swallowed something, what should I do? And how much do you charge for neutering?”
An assistant with no line drawn can write a fluent answer to both questions: general first-aid advice and a figure resembling the average prices found online. Both are risky. The first is personal health advice; the second is not the clinic's real price.
An assistant with a clear line checks the clinic's own information first. If the clinic has provided an emergency phone number and hours, it gives those; it does not give advice. If the neutering fee is not in the clinic's price list, it does not make up a figure; it says “we will get back to you” and takes the phone number. In the morning the clinic sees this question in the conversation summary and learns two things: the customer needs a call, and this item needs adding to the price list.
The second assistant may look less “clever”, but when it is speaking on behalf of your company, this is exactly what you want. An assistant's value lies not in answering every question, but in every answer it gives being correct and in passing on to you what it does not know.
Which topics should not be left to the assistant?
Some topics should be handed to a person, however well the rules are written:
- Prices or discounts that vary from person to person
- Complaints, returns and disputes
- Personal advice on health, legal or safety matters
- Anything involving a contract or a commitment
On these topics the assistant's job is not to answer, but to take the customer's details and pass them to the right person.
Preparing the assistant's information
How accurate the assistant is depends directly on how clear the information it is given is. When preparing it, a few things make the job easier:
- Write each piece of information in one place, in one form. If the same service has two different prices in two different documents, the assistant cannot know which one to quote.
- Also write down the topics that need a “we don't know” answer. A sentence like “prices that vary by customer are discussed by phone”, for example, tells the assistant what to do with that question.
- Read the first week's conversations carefully. Missing information shows up most in the first few days; once each gap is filled, it does not come up again.
Ask yourself before setting it up
| Question | Why it matters |
|---|---|
| Which information does the assistant answer from? | An assistant that answers from general information should not speak on behalf of your company. |
| What does it do when it does not know? | Does it guess, or pass the question on to you? |
| Can you see the conversations? | You cannot fix a mistake you cannot see. |
| Who updates it when the information changes? | Outdated information is the most common cause of wrong answers. |
Where the information you give the assistant is stored, and whether an AI model learns from it, is a separate question; we cover it in Will an AI tool learn my company's data? and in the data security section.
In short
The risk of an AI assistant giving a wrong answer is not zero, but if the assistant speaks only from your information, passes on to you what it does not know and its conversations reach you regularly, the risk stays small and visible.
Frequently asked questions
If the assistant says something wrong about a price, is it binding?
That is a legal question, and the answer depends on the circumstances. The practical side is this: on sensitive topics such as prices, promotions and returns, the assistant should answer only with the up-to-date information you have given it, and pass the matter on to you if it has no information.
Can the assistant answer with general information from the internet?
For an assistant speaking on behalf of your company, that is not something you want. General information can bring in another company's price or outdated details. The assistant we set up answers only with the information the company provides.
How do I keep track of what the assistant is saying?
By reading the conversation summaries regularly. They show which questions came in, which were passed on to you and what information was missing.
Is there a job that keeps repeating at your company? Tell us about it and we will build the tool that does it.
Tell us what you needOther articles
Can a spreadsheet with thousands of rows be entered into a system in bulk?
In most cases, yes. The hard part is not the entry itself but getting the data into the shape the system expects and verifying that every row was really saved. If the system offers no bulk route, the work goes through the screen, and that depends on the rules of that system.
21 September 2026Read →Why does preparing a quote take so long?
Writing the quote does not take long; the time goes into hunting for the price, copying and correcting an old quote, and rebuilding the layout every time. Once the price list sits in one place and the rules are written down, a quote comes out in minutes.
21 September 2026Read →Which jobs are suited to automation, and which are not?
The job best suited to automation is one that repeats often, whose rules can be written down and whose data is already digital. Work that needs judgement, is full of exceptions or has its data on paper needs another step first.
18 September 2026Read →Is ranking in Google the same as being mentioned by AI?
Partly. Both draw on the same foundation, but they do not reward the same thing: Google ranks a page, while AI reads several sources, writes an answer and decides for itself who to mention in it.
17 September 2026Read →Why doesn't ChatGPT mention my company?
The reason is rarely “the AI does not like us”. Usually it is a technical block, scattered company information, or the company not being mentioned anywhere else online. All three can be fixed.
17 September 2026Read →What should your site have so AI describes you correctly?
When AI describes a company, it uses the clearest sentences it finds on the site. The list below goes page by page through where those sentences should sit, how they should be written and the markup machines will read.
17 September 2026Read →What is GEO? Being visible in AI answers
GEO (Generative Engine Optimization) is the work of making sure that when ChatGPT, Gemini, Perplexity and Google's AI Overviews answer a question, they find your business, describe it correctly and cite it as a source. It does not replace SEO; it builds on top of it.
17 September 2026Read →Will an AI tool learn my company's data?
Short answer: not in a tool running through a commercial API. But that answer depends on how the setup was done, so it is worth knowing what goes where.
16 September 2026Read →When does a spreadsheet stop being enough?
Excel is not a bad tool; it handles most jobs and more. The problem is carrying on with it for months after the point where it stopped being the right fit.
16 September 2026Read →Off-the-shelf software or custom-built?
The right answer depends on how specific the work is to you. Work that everyone does the same way is cheaper off the shelf; work you do your own way is cheaper built for you.
16 September 2026Read →