Chatbot testing that catches made-up answers before customers do
A support chatbot can sound confident and still be wrong. We test what it says, how it says it, and when it should hand over to a person.
The whole story, short
The problem
The bot invents details, such as refund terms or features the company does not offer.
What we did
- Build a golden conversation set
- Test multi-turn scenarios
- Add LLM-as-judge scoring
- Check tone and policy
- Verify fallback and handoff
The outcome
Golden-set scoring with human review
What goes wrong
- The bot invents details, such as refund terms or features the company does not offer.
- Conversations break down after a few turns, when customers change topic or correct themselves.
- Tone drifts between formal and casual, and off-brand wording slips through.
- Too many conversations escalate to a human, which erases the expected support savings.
Why the usual fix falls short
Scripted tests expect one exact response, so a correct rephrasing fails while a confident wrong answer can pass. Manual review of a few conversations a week is too small to show a trend.
How we tested it
Build a golden conversation set
Use real, anonymized conversations and agree the correct outcome for each one with the support team.
Test multi-turn scenarios
Write scenarios where customers change topic, correct themselves or refer back to earlier details.
Add LLM-as-judge scoring
Score responses for correctness and faithfulness, with QA engineers reviewing a sample to catch judge errors.
Check tone and policy
Turn the brand guide and policies into explicit checks for off-brand wording and promises the company cannot honor.
Verify fallback and handoff
Confirm the bot admits uncertainty and passes full conversation context to a human agent.
Questions buyers ask
How do you test an LLM chatbot?
Start with a golden set of real conversations and agreed outcomes. Run them repeatedly, because LLM output varies between runs. Score answers for accuracy and policy compliance, test multi-turn context, and confirm that the bot hands off cleanly to a person when it cannot help.
What metrics matter in chatbot testing?
The core metrics are answer accuracy, hallucination rate, resolution rate, and escalation rate. Intent coverage shows how much of real usage your tests reach. Together they reveal whether the bot is correct and useful, and whether customers are actually getting their problems solved.
Can chatbot testing be automated?
Much of it can. Golden datasets, regression runs, and LLM-based scoring all run automatically and can sit in a release pipeline. People still need to review samples, judge tone, and investigate unusual failures, because an automated judge can also be wrong about what counts as a good answer.