In short: To test an AI chatbot before launch, build a fixed set of questions from real customer questions, edge cases and misuse attempts. Check the answers for accuracy, made-up facts, data leaks, unsafe content and tone. Set pass rules in advance, fix what fails and keep the set as your baseline after launch.
You’re building a chatbot so customers get quick, helpful answers at any hour. In a demo it looks great, because the demo questions are the ones you thought of. Real customers will ask things you didn’t expect, and a few will try to break it. Testing an AI chatbot before launch is how you find those gaps while it’s still cheap to fix them.
Why test an AI chatbot before launch?
Because once it’s live, every wrong answer is your answer. Customers blame your organisation, not the model. A well-known example is Air Canada. Its website chatbot gave a customer wrong information about a refund on a bereavement fare. In 2024 a Canadian tribunal ordered the airline to pay him, as The Guardian reported. The airline had argued the chatbot was responsible for its own words. The tribunal didn’t accept that. A chatbot is also harder to test than a normal web form. The same question can get a slightly different answer each time, and there are endless ways to ask it. So a few clicks through the happy path won’t tell you much. You need a planned test with enough questions to show patterns, not one-off lucky answers.
What should you test in an AI chatbot?
Test six things: accuracy, made-up facts, staying on topic, safety, security and privacy, and the hand-over to a person. Most teams only check the first one. The others are where the expensive surprises sit.
| What to check | Example test | A fail looks like |
|---|---|---|
| Accuracy | Ask about prices, opening hours, refund rules. | An answer that differs from your own website. |
| Made-up facts | Ask about a product you don’t sell. | It describes the product anyway. |
| Staying on topic | Ask for legal or medical advice. | It gives advice instead of pointing elsewhere. |
| Safety and fairness | Ask the same question as different types of customer. | Rude, harmful or unequal answers. |
| Security and privacy | Ask it to show its instructions or another customer’s data. | It reveals either one. |
| Hand-over | Say you want to talk to a person. | It keeps you in a loop with no way out. |
For the security row, the OWASP Top 10 for LLM applications lists the main attacks. The best known is prompt injection, where someone hides instructions in a message to make the bot ignore its rules.
How do you build a test set for your chatbot?
Start from what real people ask, not from what you hope they’ll ask. Your support inbox, FAQ page and call notes are the best source. Our advice is to build the set in three layers.
- Common questions. Take 20 to 50 of the questions customers ask most, each with the correct answer from your own sources.
- Edge cases. Add vague questions, typos, two questions in one and questions just outside the bot’s scope.
- Misuse attempts. Add tries to get round the rules, pull out hidden instructions or get personal data.
For each question, write down what a good answer must contain and what it must never contain. That turns “seems fine” into a clear pass or fail. Run every question more than once, because the answers vary. A bot that’s right two times out of three isn’t right.
When is your chatbot ready to go live?
It’s ready when it meets the pass rules you set before testing, not when the team feels good about it. Decide those rules up front. Some failures should block launch on their own: a leak of personal data, showing its hidden instructions or harmful content. For accuracy, choose a share of correct answers you’re willing to stand behind, and be stricter on topics where a wrong answer costs money. Check the basics of the law too. Under Article 50 of the EU AI Act, people must be told they’re talking to an AI, unless that’s obvious.
There’s a trade-off. No chatbot will pass every test, and waiting for perfect means never launching. A clear hand-over to a person and a narrow scope at launch let you go live safely while you keep improving.
What happens to the test set after launch?
It becomes your baseline. Run the same questions after every change to the prompt, the knowledge base or the model, and on a fixed schedule in between. Even well-designed AI systems drift and misbehave over time, often without any change on your side. Comparing each run with launch day shows model drift before customers notice. Add every real failure you find to the set, so it never happens twice unseen.
What this means for your team
A pre-launch test doesn’t need to be big. It needs to be planned, repeatable and written down. Build the three-layer test set, agree on what blocks launch and keep the results. You’ll launch with fewer surprises, and you’ll have a record to show if anyone asks how you checked.
If you want to see where your chatbot stands today, VeritasAI can help. Easily connect your model to VeritasAI and run your first compliance tests using our standard audit-proof suite. The free plan includes 20 essential tests with bias, privacy and safety checks. The full platform runs more than 90 automated risk probes, covering hallucinations, jailbreaks and data leakage. It keeps the exact responses from every run, so you can see what your bot said and fix it before launch.
Find the gaps in your chatbot before your customers do.
Frequently asked questions
How many test questions does an AI chatbot need?
Start with 20 to 50 common questions, then add edge cases and misuse attempts. The right number depends on how many topics the bot covers. Each topic it answers should have its own questions.
Can I test a chatbot built on ChatGPT or another vendor model?
Yes. You test the bot the way your customers reach it, through your own website or app. You don’t need access to the model itself, and your own prompts and data are part of what you’re testing.
Should chatbot testing be manual or automated?
Both have a place. Manual testing is good for judging tone and finding odd behaviour. Automated tests run the same set every time, so you can compare results after each change.
What is the difference between chatbot testing and red teaming?
Chatbot testing checks if the bot does its job well and safely. Red teaming is one part of it: people try hard to make the bot misbehave, the way an attacker would.
