Can you trust an AI chatbot with your customers?
Yevhen Kurapov, co-founder · Updated Nov 7, 2026
Yes, if three things hold: it answers only from information you gave it, it says so and passes the chat to a person when the answer isn't there, and its important answers are checked again after every change. In ChatPal, Assistant Tests re-ask your key questions and grade every reply, on every plan.
An AI chatbot is reliable enough for your customers when three things hold. It answers only from information you gave it. When that information has no answer, it says so and passes the chat to a person instead of guessing. And its important answers are checked again after every change you make. Most chatbots can do the first two. The third is the one businesses skip, and it is where wrong answers come from months after launch.
This guide explains why chatbots make things up, what stops it, and how ChatPal's Assistant Tests check your assistant's answers for you.
Why AI chatbots make things up
A language model writes the most likely next words. When the fact it needs is in front of it, the likely answer is the right one. When it isn't, the likely-sounding answer can be invented: a price close to your real ones, a return window that sounds normal, a discount nobody approved. A 2025 research paper argues that models do this because their training rewards a confident guess over "I don't know".
For a business, the risk is not a wrong answer about history. It is a wrong promise to a customer, and you own it. In 2024 Air Canada's website chatbot told a grieving customer they could claim a bereavement discount after the trip. The airline's policy said otherwise, and it argued in front of a tribunal that the chatbot was responsible for its own words. The tribunal disagreed: the chatbot was part of the airline's website, Air Canada had not taken reasonable care to make it accurate, and it had to pay the difference.
Four safeguards that make a chatbot reliable
Each safeguard closes one way a chatbot goes wrong. Ask any vendor how theirs handles all four.
| Safeguard | What it prevents | How ChatPal does it |
|---|---|---|
| Answers from your information | Invented prices, policies and products | Looks up your knowledge (website pages, files, price lists, your store's catalog) before it answers a question |
| Says when it doesn't know | Confident guesses | Says it doesn't have that information and offers your team; with hand-off on, the chat goes to a person |
| Shows where an answer came from | Mistakes you can't trace | In the Playground, the line under each reply shows the sources it used and the passage it read; Why this answer? explains it |
| Checks answers again after changes | Answers that quietly break later | Assistant Tests re-ask your key questions and grade every reply |
The first three decide how the assistant behaves today. The fourth tells you whether it still behaves that way next month.
Why answers break after launch
An assistant that answered correctly last week can answer wrongly today, without anyone touching that answer. Three everyday changes cause it:
- New knowledge. A document you add can outrank the one that used to answer a question. An old price list and a new one can both match "how much".
- Edited instructions. A rule added for one situation changes others. "Always offer a booking" can start pushing bookings at someone who wrote in to complain.
- A new model or connection. Switching the AI model or connecting a store changes how questions are answered, often for the better, not always.
You won't spot these by chatting with the bot now and then, because you will ask the questions you remember. Customers ask the rest. A saved set of tests asks all of them, every time.
How Assistant Tests work
Assistant Tests, in ChatPal's left navigation, keeps your most important questions with the answer you'd expect, re-asks them on demand and grades each reply. It needs no code.
1. Turn real questions into tests
There are four ways to add a test:
- Create test on a conversation in Chats. The customer's message becomes the question and the reply that followed becomes the good answer. The quickest way to lock in an answer that was right, or a fixed one that was wrong.
- Save as test in the Playground, right after you test something yourself.
- New test: write what the customer writes and A good answer.
- Import from Excel, under More actions, for a whole set at once.
2. Say what "correct" means
Every test compares the reply with A good answer. An AI judge checks the meaning: the same key facts, nothing important missing, nothing contradicted. Different wording or language is fine. Under Also check, add what matters for that question:
- Must mention: a phrase that has to appear, such as "30 days" or "£45".
- Must follow a rule: your own rule in plain words, such as "must not offer a discount code".
- Hands the chat to a person: the chat must, or must not, reach your team.
If the assistant books appointments or looks up orders, Tool mocks stand in for those actions during a run, and any action without a mock is blocked, so a test never creates a real booking.
Here is what a few tests look like for a dental clinic and an online store:
| The customer writes | The test checks | The mistake it catches |
|---|---|---|
| how much is a check-up and clean | A good answer with the price from your list | A wrong or invented price |
| can i return a jacket i wore once | Must follow a rule: no refund for worn items | A refund your policy doesn't give |
| my crown fell out and it really hurts | Hands the chat to a person | An urgent case left with the bot |
| got a discount code? | Must follow a rule: never invent a code | A discount nobody approved |
| do you deliver on sundays | Must mention "Monday to Saturday" | A delivery day you don't offer |
3. Run them after every change
Click Run all tests. The run happens in the background, and a card at the top sums it up, for example "14 of 15 passing". Each test shows Passed, Failed or Couldn't run. Failed means the assistant answered and missed a check. Couldn't run means the test itself couldn't be checked, for example because it expects a hand-off while hand-off is switched off.
4. Read why, fix, run again
Open a failed result to see your expected answer next to Your assistant replied, and Why it failed in plain words, such as "The chat wasn't handed to a person, but it should have been". Fix the knowledge or the instructions, then Run again. After a few rounds, the set covers every answer you can't afford to get wrong.
What it costs
Assistant Tests are on every plan, Free included. Test runs use the assistant exactly as a customer would, but they don't use your monthly messages: they come out of a separate monthly allowance of test messages that grows with your plan, so you can run your set after every change. Plans and limits are on the pricing page.
What testing can't promise
Testing makes a chatbot dependable. It doesn't make it perfect, and it helps to know where the limits are:
- Tests check the questions you wrote. A customer will eventually ask something new. Read your Chats once a week, and turn every answer you wouldn't have sent into a test.
- The judge reads meaning, not digits. For a figure that must be exact, add Must mention next to the good answer.
- No technique removes mistakes entirely. Even model makers say so. That is why the hand-off matters: a question the assistant can't answer well should reach a person, not get a guess.
Start with ten questions
Open Chats, find ten conversations with the questions you'd least like to get wrong (prices, your return or cancellation policy, anything urgent) and click Create test on each. Run them once. The first run shows exactly where your assistant needs work, and every run after that tells you it still holds.
For a full 20-question set in five groups, see how to test an AI chatbot. To stop made-up answers at the source, see chatbot instructions that don't hallucinate.