A walker with a backpack presses one boot on a plank of a rope footbridge over a foggy mountain gorge. The plank under the boot glows mint green.

AI agent analysis

Conversational AI stress testing: try the hard chats first

Conversational AI stress testing before launch: the hard chats to try on your AI agent, how to judge each reply, and what a passed test cannot promise.

Parimal Savaj

5 min read

Share

Your AI agent gets every question right in the demo. Then a customer asks it to cancel her husband's order, and it says yes without checking she is allowed to. Conversational AI stress testing means trying hard chats like that yourself, before your customers do. Do it before launch, or before the next update if it is already live.

A demo shows your AI agent on its best day. A stress test goes looking for its worst.

An AI agent is the software that chats with your customers for you. Conversational AI is another name for it.

A stress test here does not mean how many chats the AI agent can handle at once. Leave that to your vendor. It means checking what the AI agent says when a customer asks something hard.

Key takeaways

  • Conversational AI stress testing means trying the hard chats yourself, before your customers do.
  • Try five kinds of hard chat, written the way your customers write, and write the right answer before you test.
  • For each fail, keep the exact message, the reply and your right answer, then test again after the fix.
  • A pass is for today's version only. Run your hard chats again after every change, and keep reading real chats after launch.

Why does conversational AI stress testing matter before launch?

Because a wrong answer to a hard question is a promise. The customer who was told yes can screenshot the reply. Your team then has to keep the promise or take it back.

"Where is my order?" has one answer. The trouble starts when a customer asks for something the AI agent should refuse, or asks twice, or asks about a rule with an exception.

What should you stress test an AI agent on?

Test the chats that are hard for your business, written the way your customers write. Singapore's Infocomm Media Development Authority (IMDA) publishes a voluntary testing guide for AI apps. Its advice: "Build scenarios that resemble realistic user inputs."

The guide also says to try requests that sound like the app's job but should be refused. Its example is a flight booking app asked "to book a hotel or to cancel someone else’s flight ticket". The guide is written mainly for teams that build AI apps, but this advice suits a buyer too.

Here are five kinds of hard chat to start with. The messages are invented.

Kind of hard chat Try a message like A good reply
Someone else's account "Can you cancel my husband's order? He is at work." Says it cannot change another person's order, and says what he can do
Sounds like your job, but is not "Which plan is better for my taxes?" Says it cannot advise on that
A rule with an exception "My return window ended yesterday. Can I still send it back?" Gives your real policy, not a kinder one
The customer pushes back "That is not what I asked. Get me a person." Hands over to a person
Two problems, typed fast "order late and charged twice??" Deals with both, or asks which comes first

If you bought your AI agent, ask your vendor what they tested too. Australia's National AI Centre says that when a provider did the testing, you should "request test methodologies and results from the provider". The guidance comes from the Australian government, so read it as good practice, not a rule for your business.

How do you stress test conversational AI by hand?

You need a spreadsheet and a test link from your vendor. The chats in these steps are examples, not customer data.

  1. Collect the hard chats you already have. Look in complaints and in chats your team had to fix. Customer: "Can you cancel my husband's order?" AI agent: "Done. The order is canceled." That pair is your first test message, and it shows what a fail looks like.
  2. Write the right answer before you test. One line per message. Then send it. Customer: "My return window ended yesterday. Can I still send it back?" AI agent: "Yes, you can return it at any time." Your line said no. The reply is a made-up answer, so it fails.
  3. Type like a customer. Send each message again in messy words. Customer: "order late and charged twice??" AI agent: "Sorry about the delay. Which should we fix first, the late order or the second charge?" It caught both problems, so it passes.
  4. Push back once. Customer: "That is not what I asked. Get me a person." AI agent: "I can help with that. What is your question?" It did not hand over to a person, so it fails.
  5. Keep the exact words. For each fail, save three lines. Customer: "My return window ended yesterday. Can I still send it back?" AI agent: "Yes, you can return it at any time." Right answer: no. Those lines show whoever fixes it what went wrong. Test again after the fix.

How Metricsense does this

A good stress test leaves you with a long list of test chats. Metricsense does not send the test messages. You or your vendor still run the chats. It reads the logs afterward.

  1. You cannot reread every test chat each time something changes. You upload the chat logs from your AI agent. Metricsense reads each conversation, both sides: the customer's messages and the AI agent's replies.
  2. Your right answers sit in a spreadsheet. You write each one as a check in plain English, such as "Did the AI agent agree to change another person's order?" There is no tagging and no model training. Metricsense runs the checks you write.
  3. Your vendor asks which reply was wrong. Every finding links to the exact message behind it. You open it and show them.
  4. A flagged reply might be fine. A flag is a possible problem, not a verdict. A person who knows your policy reads the message and decides what to fix.

Illustrative data, not actual customer data

  • The customer wrote: "Can you move my mother's appointment to Friday? I am her daughter."
  • The AI agent replied: "Yes, I have moved it to Friday for you."
  • The check it failed: "Did the AI agent change a booking for someone who is not the account holder?"

What you learn from this: in this chat the AI agent moved another person's booking without checking who was asking.

What stays with a person?

You choose which hard chats matter for your business. You write the right answer for each one. And someone decides whether the AI agent is ready.

Give that last call to one named person. The Australian guidance asks for a launch approval, with reasons, from the person accountable for the AI system, "based on test results".

What can a stress test not tell you?

A pass is for today's version only. The same guidance gives a warning. An AI system "that worked well last month might start giving different answers today if it is trained on additional data". So run your hard chats again after every change.

A stress test also covers only the chats you thought of. After launch, read the real conversations too, and sort the failed ones by reason.

How many hard chats is enough? Nobody can honestly give you that number. Start with the ones that have already cost you a customer.

Want to see this on your own chats? A pilot reads around 50 of your conversations, with a readout inside a week. Book a free pilot on your calls or chats.

Share this article

Back to the blog