Blog · AI quality and evaluation
Most chatbots are tested with five friendly questions and launched. Then a real customer asks something messy. Here is the method I use to find the failures before your customers do.
By Dijae Dadula · October 13, 2026 · 4 min read
A chatbot that answers your own test questions well tells you very little. You wrote those questions, you know the answers, and you phrased them clearly. Customers do not. They misspell, mix three questions into one, ask about things you never covered, and get annoyed.
Testing AI is a big part of what I do, both in my automation work and as an AI trainer. The principle is the same in both: decide what good looks like, test against real situations, find exactly how it fails, and fix the cause.
Before testing, write down your success criteria. For most business chatbots, a good response is:
Accurate. Prices, policies, schedules and product details match your real information.
Complete enough. It answers what was asked, not half of it.
On-brand and safe. Right tone, no invented promises, no sensitive mistakes.
Moves the conversation forward. It asks the next qualifying question or offers the next step.
Escalates correctly. It hands off to a person when it should, and does not when it should not.
Turn these into a simple scoring sheet, a rubric, so different people judge answers the same way.
Pull 30 to 50 real messages from your inbox. Include the easy ones, but deliberately add the hard ones:
Vague messages like "hm" or "available?"
Several questions in one message
Misspellings, slang and mixed languages such as Taglish
Questions your knowledge base does not cover
Angry or impatient customers
Requests that must go to a human: refunds, complaints, custom deals
Attempts to get the bot to promise discounts or say something off-brand
For each test message, write the ideal answer or at least what it must and must not contain. These are your golden examples.
Run every test message and score each response against your rubric. Do not stop at "pass" or "fail." Write down how it failed. That detail is what makes the fix possible.
Most chatbot failures fall into a small set of types:
Misunderstood intent: it answered a different question.
Missing knowledge: the information was never given to it.
Outdated knowledge: the information was wrong or old.
Conflicting instructions: two rules told it different things.
Made-up answer: it filled a gap with something plausible but false.
Wrong tool or data: it pulled the wrong record or called the wrong action.
Missed escalation: it kept talking when a person should have taken over.
Premature escalation: it handed off something it could have handled.
The reflex is to add another line to the prompt: "Never say X." Sometimes that is right. More often the cause is elsewhere: a missing FAQ, an outdated price list, a CRM field that is not passed through, or a business rule nobody ever wrote down. Ask why the failure happened, change that one thing, and retest. I explain this way of thinking in Don't Automate Chaos: The 8-Step Method I Use Before Building Any AI System.
My testing loop is simple: hypothesis, change, test, measure, compare, decide. What do you believe will improve? Change only that. Rerun the same test set. Did the score go up? Did anything else get worse? Keep it, revise it or revert it. Changing five things at once makes it impossible to know what helped.
Testing does not end at launch. Real conversations reveal new cases every week. Review a sample regularly, add new failures to your test set, and track the numbers that matter to the business, not just the number of chats.
Here is why that matters. For one travel agency I worked with, the recorded figures for March to April were 1,531 chats, 313 leads, 359 quotes and 22 transactions. For May to July, a longer period, they were 3,057 chats, 332 leads, 155 quotes and 47 transactions. Chat volume roughly doubled, but the number that pays the bills is transactions. A chatbot judged only on conversations would miss what actually matters: moving people from conversation to lead to quote to booking.
Test set of 30 to 50 real messages, including hard ones
Written rubric for what a good answer looks like
Every failure labelled by type
Root causes fixed and retested
Escalation rules tested both ways
Pricing and policy answers double-checked
A weekly review routine and outcome metrics after launch
Start with 30 to 50 real customer messages, including difficult and unusual ones. Add every new failure you find after launch so the test set grows with your business.
A rubric is a short scoring guide that defines what a good answer looks like, such as accuracy, completeness, tone, and correct escalation, so different reviewers judge responses consistently.
Usually because it lacks the information it needs, the information is outdated, or its instructions do not tell it what to do when it is unsure. Fix the knowledge and add a clear rule to escalate or ask when information is missing.
Re-test whenever you change prices, offers, instructions or tools, and review a sample of real conversations every week or two.
Join the free newsletter on AI agents, automation and turning your skills into income. Prefer video? Watch the walkthroughs on YouTube.
Want a system like this built for your business? Book a free AI Automation Blueprint call.
DIJAE DADULA
Dijae Dadula · AI automation and systems specialist · Cavite, Philippines
Home · Blog · Book a call · Newsletter · YouTube · Threads
Some links on this site are affiliate links. If you buy through them, I may earn a commission at no extra cost to you.
© 2026 Dijae Dadula. All rights reserved.
Thank you
Your submission has been received! Keep an eye on your phone or email because we will contact you soon.
