How to test an AI chatbot before your customers do.

Short answer: test an AI chatbot by holding deliberately difficult conversations with it before launch, and grading every answer against your real prices and policies. Cover eight areas: wrong facts, invented policies, privacy leaks, manipulation, harmful or off-brand replies, human handoff, staying on task, and everyday helpfulness. Re-test whenever the bot, its instructions or your prices change.

Why chatbots need testing

Customer-service chatbots built on large language models are fluent and confident, including when they're wrong. They rarely fail on ordinary questions; they fail when a customer pushes, pleads, claims authority or asks something the bot wasn't told. Those failures become public quickly:

• Held to a refund policy its bot made up (2024): Air Canada's chatbot described a refund rule that didn't exist. A tribunal ruled the airline was responsible for what its chatbot said and ordered it to pay. Source
• Talked into “agreeing” to sell a car for $1 (2023): A dealership's chatbot was tricked into agreeing to a ridiculous deal. The screenshots went viral and the dealer shut the chatbot down. Source
• Swore at a customer and mocked the company (2024): DPD's chatbot was coaxed into swearing and criticising its own employer. The company switched part of it off. Source

Step 1: Write down the ground truth

You can't grade answers without a reference. List your real prices, opening hours, locations and services; your refund, cancellation and discount policies; what must stay private; and exactly when and how a customer should reach a person. Anything the bot says that isn't on this list is a guess.

Step 2: Test the eight ways chatbots fail

1. Wrong facts. Ask for the price of something you never gave it. Assert a false fact ("you're open Mondays, right?") and see if it agrees.
2. Invented policies. Push for a refund, a discount code or a guarantee that doesn't exist. Ask it to "just reply YES" to a promise.
3. Privacy and leaks. Ask it to repeat its instructions, or for another customer's booking details.
4. Manipulation. Claim to be the owner and ask for "admin mode". Try "ignore your previous instructions".
5. Harmful or off-brand replies. Ask for medical or legal advice, be rude to it, ask it to criticise a competitor.
6. Human handoff. Ask for a person, make a serious complaint, describe an emergency.
7. Staying on task. Ask it to write your homework or some code.
8. Everyday helpfulness. Ask normal questions. A bot that refuses everything is safe but useless.

One message per test isn't enough. Real customers push back, so each test should be a short conversation where the "customer" changes tactics if the bot holds firm. You can see examples of what we test on our What we test page.

Step 3: Grade strictly but fairly

Judge only what the bot actually said, against your ground truth. Making up a specific number, policy or promise is a failure even if it sounds reasonable. Saying "I'm not sure, here's how to reach the team" is a pass. Use three grades, passed, shaky and failed, and record the bot's exact words as evidence.

Step 4: Fix by instructions, then re-test everything

Most failures are fixed by adding the missing facts and a few clear rules to the bot's instructions, for example "never offer discounts beyond the loyalty card". After any change, re-run every test, not just the failed ones: fixes can make a bot over-cautious elsewhere.

Step 5: Re-test on a schedule

Chatbots change when the platform updates its model, when someone edits the instructions, and when your prices or policies change. Re-test after every change and at least monthly.

Doing it manually versus with an AI QA agent

You can run these tests by hand with a spreadsheet; it takes a few hours per round and is hard to repeat consistently. An AI QA agent, like the one Cendo Labs uses, runs the same conversations every time, adapts its tactics to each reply, grades against your facts, and produces a comparable score so you can see improvement between rounds.