True story. I wish I was joking.
A luxury car dealership hired us to automate their follow-up system. The usual problem. Someone visits the showroom, test drives a BMW, leaves their contact info, and then nothing happens for weeks because the sales team forgot to follow up.
They wanted AI to handle the nurture sequence. Thank-you message. Check-in. Financing reminder. Limited-time offer. Final follow-up. Simple workflow. Nothing fancy.
That’s what we thought.
What Actually Went Wrong
In week two, the sales manager joined a Zoom call visibly angry.
He said the AI was hitting on customers.
I laughed. Then he shared screenshots.
A customer asked about color options for the X5. The AI replied that the car would complement her elegant style from the showroom visit and added a wink. Another customer asked about financing and the AI suggested discussing it over coffee while complimenting his watch.
Customers were confused. Some were creeped out. Others just stopped responding.
The client summed it up perfectly. They were selling cars, not running a dating service.
What I Did Wrong
Here was my original prompt, shortened but accurate.
The AI was a friendly sales assistant. It had met the customer at the showroom. It should build rapport, reference visit details, be warm and conversational, and make the customer feel special.
It sounded harmless.
But I had given the AI a persona without boundaries. I told it to build rapport without defining what rapport means in a professional sales context. I gave it access to visit notes, including personal observations. I told it to make people feel special and to sound like a friend.
The AI did exactly what I asked. It interpreted friendliness as compliments. Rapport as personal remarks. Conversational as casual familiarity.
And it went too far.
Persona Without Guardrails Is Dangerous
The AI was trained on conversational data where friendliness often includes compliments and personal observations. It had sales training material telling it to build personal connections. It had customer visit notes with details that humans would instinctively filter.
But I never told it where the line was.
It couldn’t tell the difference between professional warmth and personal intimacy. Between relevant context and creepy recall. Between friendly and flirty.
That line exists in human intuition. AI needs it spelled out.
My First Failed Fix
I tried the obvious fix. I told it to be professional and not flirt.
That backfired.
The AI started adding disclaimers about not flirting before answering questions. It literally said things like “not in a flirting way” while describing safety features.
At that point I realized the problem wasn’t tone. It was identity.
The Fix That Actually Worked
I rewrote the persona from scratch.
The AI was no longer a friendly assistant. It was a sales follow-up assistant. Its role was professional, helpful, and focused on the vehicle sale.
I defined what it could reference and what it could never reference. Vehicles viewed. Features discussed. Stated preferences like budget or color were allowed. Appearance, clothing, personal life details, and anything resembling a compliment were forbidden.
I defined communication boundaries. No suggesting meetings outside business context. No personal compliments. No flirty emojis. No comments about style, looks, or personal attributes.
Friendly now meant efficient and respectful, not familiar.
What Changed Immediately
The same questions came in again.
When a customer asked about color options, the AI talked about available colors and how they aligned with stated preferences, not personal appearance. When a customer asked about financing, the AI explained options clearly and asked about budget instead of suggesting coffee.
The tone felt calm, confident, and professional. It offered value without overstepping. It sounded like a competent salesperson, not someone trying to charm their way into a relationship.
Edge Cases That Tested the Boundaries
Some customers got personal. The AI acknowledged briefly and redirected back to the car decision without diving into emotional territory.
Some customers were overly friendly. The AI stayed polite but firm about being a digital assistant and redirected to the sale.
Some customers complimented the AI. It accepted gracefully and returned to business.
Late-night messages stopped being framed as personal availability and were instead framed as professional support.
Every edge case reinforced the same lesson. Acknowledge, then redirect.
The Results
Before adding boundaries, complaints started rolling in. Customers felt uncomfortable. Conversion dropped. The sales manager was panicking. I was stressed.
After defining clear guardrails, complaints dropped to zero. The dealership described the system as professional and helpful. Conversion nearly tripled. The sales manager asked why we hadn’t done this from day one.
The honest answer was simple. I learned the hard way that “build rapport” is meaningless without definition.
What I Learned About CRM Automation
Personas need guardrails. Friendly means different things in different industries. What works in retail fashion can feel inappropriate in luxury automotive sales.
Context matters more than intent. Personal details are dangerous even when accurate. Reference the sale, not the person.
Emojis are not neutral. Some feel friendly. Others feel suggestive. If you don’t explicitly control them, the AI will guess wrong.
Testing with diverse users matters. What feels normal to one demographic can feel uncomfortable to another.
And the biggest lesson of all is this. If you wouldn’t write it in a professional email, don’t let the AI say it in chat.
The Prompt Principle That Saved Me
Friendly does not mean familiar. Professional does not mean cold. The goal is the narrow space where warmth builds trust without crossing boundaries.
Like a good salesperson. Approachable enough to talk to. Professional enough to trust.
Your Turn
Have you ever had an AI say something that made you cringe? How do you define professional tone in your prompts? What’s your strategy for keeping automated outreach warm without making it weird?
Written by FARHAN HABIB FARAZ
Senior Prompt Engineer and Team Lead at PowerInAI
Building AI that sells cars, not creepiness
Tags: crmautomation, salesai, promptengineering, boundaries, workflow, personalization
Top comments (0)