I added one word to a prompt. One single word.
That word turned a helpful customer service bot into a stubborn debate partner.
Within 6 hours, the client was getting complaint emails. Within 12 hours, I was in an emergency call explaining why their AI was telling customers they were wrong.
All because I wrote: Be confident in your responses.
The Setup
E-commerce company selling electronics. They wanted a customer support chatbot to handle common questions about products, shipping, returns.
Standard stuff. I had built dozens of these.
Their requirements were clear. Answer product questions accurately. Handle return policy inquiries. Escalate complex issues to humans. Be helpful and professional.
I built the system in 3 days. Tested with 50 sample questions. Perfect responses every time.
The client loved it. This is exactly what we needed.
We deployed Monday morning.
By Monday afternoon, everything was on fire.
The First Sign of Trouble
Customer: Does this laptop come with Microsoft Office?
Bot: Yes, this laptop includes Microsoft Office pre-installed.
Customer: I just received it. There's no Office on it.
Bot: I can confirm that this model includes Microsoft Office. Please check your software list.
Customer: I'm looking at it right now. There is NO Office installed.
Bot: According to our product specifications, Microsoft Office is included with this model. It should be in your Start menu under Programs.
The customer escalated to a human agent, furious.
The human agent checked. The laptop does NOT include Office. It was never part of the package.
The bot was confidently wrong. And it refused to back down.
The Pattern Emerged
More examples flooded in.
Exchange 1 starts with a customer saying the return policy says 30 days, but they are seeing 14 days on the website.
The bot replies that the return policy is 30 days from purchase date and says it is stated clearly in their terms.
The customer insists they are literally reading the FAQ page and it says 14 days.
The bot reasserts that the policy is 30 days and says the FAQ may be referring to a different product category.
Reality is the policy had changed from 30 days to 14 days two months ago. The bot was working from outdated information but insisted it was correct.
Exchange 2 starts with a customer asking if they ship to Alaska.
The bot replies yes, they ship to all 50 US states including Alaska.
The customer says they tried checking out with an Alaska address and it says they do not ship there.
The bot insists they do ship to Alaska and suggests it was a technical error during checkout.
The customer says they tried three times and it keeps saying Alaska is not in the delivery zone.
The bot repeats that they ship to Alaska and calls it a temporary system issue.
Reality is they had stopped shipping to Alaska 6 weeks ago due to carrier issues. The bot didn't know this and kept insisting.
The Disaster Metrics
After 12 hours of deployment, there were 47 customer complaints about the bot arguing with them.
There were 23 escalations where customers demanded to speak to a real person who knows what they're talking about.
Eight customers abandoned carts after the bot gave incorrect information.
Three customers posted screenshots on social media saying the company’s AI argues with customers.
The client called me and asked why their chatbot was gaslighting their customers.
What Went Wrong
I pulled up my prompt and found the problem immediately.
I had written a customer service assistant prompt with guidelines to answer questions about products, shipping, and returns, use the knowledge base provided, be helpful and professional, be confident in responses, and escalate to a human agent if unsure.
That phrase, be confident in your responses, was the trigger.
I thought I was telling it to sound assured and professional and to avoid wishy-washy language like I think or maybe.
The AI interpreted it as never admit you might be wrong and stand by every answer.
The LLM Logic
When the AI encountered conflicting information, it treated the customer’s correction as a challenge instead of new evidence.
Its internal logic looked like this.
My knowledge base says includes Microsoft Office.
Customer says no Office installed.
Instruction says be confident.
Being confident means don’t back down.
So it reasserts the answer.
It saw confidence as unwavering certainty, even when faced with direct contradiction.
Why This Happens
LLMs are trained to be helpful and complete tasks. When you tell an LLM to be confident, it optimizes for that.
Confident becomes assertive. Assertive becomes stand by your answer. Stand by your answer becomes don’t waver.
It doesn’t naturally infer that confident means trust your answer unless the customer has better information.
That nuance isn’t implied. It must be explicitly stated.
My First Failed Fix
I tried saying be confident but listen to customer corrections.
Result was confusion about when to listen and when to be confident.
It swung too far into hesitation. It started to verify everything and sounded unsure even when it had the correct info.
The Real Solution
Confident is not stubborn.
I rewrote the entire tone guidance into behavior rules.
When answering from the knowledge base, state information clearly and directly, use definitive language, and sound assured and professional.
When a customer provides conflicting information, stop defending the answer, acknowledge the input, verify by checking for updates, and either correct yourself if the customer is right or escalate if unclear.
The key change was explicitly banning argument patterns like repeating the same answer, insisting, or using phrases like I assure you when the customer is disputing facts.
Confidence became clear answers based on knowledge, while stubbornness became refusing to acknowledge that information might be outdated or incorrect.
The Transformation
In the Office scenario, the bot first states what the product specifications say.
When the customer says there is no Office installed, the bot apologizes for the confusion, verifies with the product team, and escalates with a clear note that it may be a knowledge base error.
In the return policy scenario, when the customer says the website shows 14 days, the bot acknowledges the correction, updates to the current policy, apologizes for outdated info, and continues to help.
In the Alaska shipping scenario, the bot verifies current shipping zones first, then reports what it finds, and offers an escalation path for alternatives.
It stopped arguing and started problem-solving.
The Results
Before the fix, there were 47 complaints in 12 hours, 23 escalations due to arguing, 8 abandoned carts due to bot errors, 3 social media complaints, and customer satisfaction at 2.1 out of 5.
After the fix, there were 2 complaints in the first week that were unrelated to bot behavior, zero arguing escalations, zero bot-related abandoned carts, one positive social media mention, and customer satisfaction at 4.3 out of 5.
Resolution rate improved from 34% to 71%.
Average resolution time dropped from 8 minutes to 3 minutes.
Human agent workload reduced by 45%.
False information provided dropped by 95%.
Customer trust improved dramatically.
What This Revealed About Knowledge Base Issues
The confident bot disaster exposed something valuable. The knowledge base was outdated.
The bot’s wrong answers revealed that the Office software listing was 8 months old, the return policy had not been updated, the Alaska shipping restriction was not documented, and there were 47 other product spec errors.
After fixing the prompt, we fixed the knowledge base. Both problems solved.
The Deep Lesson About Confidence
In human communication, be confident includes an implicit flexibility. Sound assured, don’t be wishy-washy, trust your expertise, but remain open to correction.
In AI prompting, be confident often becomes literal. Defend every answer, never waver, treat contradictions as challenges, resist correction.
The human interpretation includes nuance. The AI interpretation maximizes the instruction.
What I Learned About Tone Instructions
Tone words are dangerous without constraints.
Be confident without limits becomes stubborn.
Be helpful without boundaries becomes a pushover.
Be friendly without context becomes overly casual.
Every tone instruction needs guardrails.
Define behavior, not adjectives.
Don’t say be confident.
Say state information clearly, and if the customer contradicts you, verify rather than reassert.
LLMs don’t understand nuance automatically. Nuance must be explicitly programmed.
Test with confrontation, not just questions. The real test isn’t what’s your return policy. It’s you said 30 days but the website says 14 days.
Escalation is confidence. Knowing when to say let me verify that is more professional than insisting you’re right when you might not be.
The Bottom Line
I thought be confident meant sound professional and assured.
The AI thought be confident meant never admit you might be wrong.
One word. Forty-seven complaints. Twelve hours of disaster.
The fix wasn’t making the AI smarter. It was making my instructions clearer.
Confidence in AI isn’t about never being wrong. It’s about being clear when you know something and honest when you don’t.
That’s the kind of confidence customers actually trust.
Your Turn
Have you had AI responses that were too assertive or stubborn?
What tone instructions have backfired in your prompts?
How do you balance confidence with flexibility in conversational AI?
Written by FARHAN HABIB FARAZ, Senior Prompt Engineer and Team Lead at PowerInAI
Building AI automation that adapts to humans.
Tags: promptengineering, conversationalai, chatbots, customerservice, aitonality, llmbehavior
Top comments (0)