A support ticket came in last week for our saddle brand, Iberosattel. The subject line was simply: “Question.” The body read: “Does the saddle come in brown?”
It’s a simple, common query. Answering it takes a human agent about 90 seconds: look up the customer, open the ticket, type a polite response listing the available colors, and close the ticket. While not a difficult task, it’s a recurring one. Multiplied by dozens of similar questions every day across our brands, this process can become a bottleneck. The real cost isn’t those 90 seconds; it’s the context switching that pulls my team away from complex, high-value customer issues.
My immediate thought, as always, was: can we systematize this? That question marked the start of a new experiment at JvG Labs: using AI to handle the initial customer response. Our goal isn’t to replace the support team, but to build a digital triage assistant that’s fast, accurate, and, most importantly, empathetic. The true challenge isn’t the technology—it’s calibrating the machine to sound human.
The Modern Customer Service Dilemma: Speed vs. The Human Touch
Before building anything, I looked at the current landscape. Today’s customers expect a sense of immediacy. In fact, research from Salesforce shows that 83% of customers expect to engage with someone immediately upon contacting a company. This demand for instant acknowledgment creates a huge strain on human-only support teams.
At the same time, speed without quality is disastrous. A study by Zendesk found that the single most important factor for a good customer service experience is being able to resolve an issue quickly. Yet, Microsoft’s research paints an even starker picture: 62% of consumers say they have stopped doing business with a brand due to poor customer service experiences.
This is the core tension: customers demand the speed of a machine but crave the understanding of a person. My experiment is focused on navigating this paradox: How can an automated system provide immediate acknowledgment without sounding like a soulless robot, and how does it know when its capabilities are exceeded?
My First-Response System: The Initial Blueprint
The objective was never to have an AI solve a complex fitting issue for a Westernsattel. Instead, the goal is more modest and, I believe, more achievable:
- Acknowledge Instantly: Let the customer know their message has been received and is being processed, 24/7.
- Triage & Categorize: Understand the basic intent. Is it a sales question, a support request, or a shipping inquiry?
- Provide Initial Information: If it’s a simple, FAQ-style question (like „Do you have brown saddles?“), answer it directly.
- Escalate Intelligently: Hand off the conversation to a human expert at the first sign of complexity, frustration, or nuance.
This logic boils down to a simple decision tree. This first contact is critical because it sets the tone for the entire interaction. A bad first impression can frame the experience negatively, even if a human agent later provides a perfect solution.
Calibrating Tone: The Art of the Prompt
The heart of this system is prompt engineering. It’s less about code and more about language, psychology, and defining clear operational boundaries for the AI. A generic prompt like „You are a customer service agent“ produces generic, corporate-sounding responses—the equivalent of giving a new employee a title but no training.
Instead, I focused on building a detailed persona and instruction set. This involves defining variables that make the response feel personal and contextual.
My prompt framework includes directives like:
- Tone: „Your tone is helpful, calm, and professional, like a knowledgeable and friendly colleague. Avoid overly enthusiastic or apologetic language. Be direct but warm.“
- Context: „You are assisting a customer of a premium saddle company. Our customers are passionate horse owners who value quality and expertise.“
- Action: „Your primary goal is to understand the customer’s need and provide immediate, accurate information if possible. If the query involves a custom order, a fitting problem, or any expression of dissatisfaction, your only goal is to confirm you understand and that you are routing them to a specialized human agent immediately.“
My experience with automation in marketing (https://www.patrick-thoma.com/blog/automation-in-marketing) was useful here. Just as in marketing, personalization and context are everything. We aren’t just inserting a name; we’re tailoring the entire interaction based on the user’s implicit and explicit signals.
The Most Critical Component: The Escalation Protocol
An automated system is only as good as its safety net. The single most important part of this project was designing the escalation triggers—the rules that tell the AI to stop and hand the conversation over to a person. An AI that tries to solve a problem it doesn’t understand will only create more frustration.
I mapped out the non-negotiable triggers for immediate human escalation. These triggers include:
- Sentiment Analysis: Keywords like „disappointed,“ „frustrated,“ „angry,“ „unacceptable.“
- High-Stakes Keywords: Terms like „cancel,“ „refund,“ „legal,“ „complaint.“
- Query Complexity: If the initial message contains more than two distinct questions.
- Product Specificity: Any mention of a specific order number or a past support ticket ID.
- Ambiguity: If the AI’s confidence score in understanding the customer’s intent is below 95%.
This protocol is the system’s humility, a built-in recognition of its own limits. By escalating effectively, the AI handles the high-volume, low-complexity queries, freeing up our human experts to apply their skills where they matter most: in conversations requiring deep product knowledge, empathy, and creative problem-solving.
Frequently Asked Questions (FAQ)
Will AI replace my human support team?
Based on my current experiment, no. The goal is augmentation, not replacement. The AI acts as a „digital assistant“ that handles repetitive first-contact tasks, allowing human agents to focus on the complex, relationship-building conversations that drive customer loyalty. It’s about elevating the role of the human agent, not eliminating it.
How do you measure the success of an automated first-response system?
I’m tracking several key metrics:
- First Response Time (FRT): This should drop dramatically, ideally to under one minute.
- Human Agent Load: A reduction in the number of simple, Tier-1 tickets that reach human agents.
- Escalation Rate: A measure of how many conversations the AI correctly passes to a human. This shouldn’t be too low; a high escalation rate early on means the safety net is working.
- Customer Satisfaction (CSAT): The ultimate measure. We’ll be monitoring CSAT scores to ensure this system improves the customer experience, not detracts from it.
Isn’t an AI response cold and impersonal?
It can be, if not designed correctly. The entire focus of my experiment is on calibrating the tone, language, and rules to avoid this. By providing deep context about our brand, customers, and desired voice—and by programming it to escalate any emotionally charged situation—we aim for a response that feels efficient and helpful, not robotic.
What tools are needed to build a system like this?
The core components are a customer support platform (like Zendesk, Gorgias, or Intercom), an AI model (accessible via an API like OpenAI’s), and often a „connector“ tool (like Zapier or Make) to link them together if a native integration doesn’t exist. The real work, however, is not in the tools but in the strategic design of the prompts and escalation logic.
The Journey of Systematizing Trust
This experiment is still in its early stages, but it’s a fascinating look into the future of customer interaction. The challenge is not technological but deeply human: encoding empathy, humility, and judgment into a system. It’s a perfect example of what I find most compelling about designing scalable business systems (https://www.patrick-thoma.com/blog/designing-scalable-business-systems)—creating structures that put human expertise to better use.
The goal is to build a system that our customers don’t just tolerate, but appreciate—one that respects their time and gets them to the right answer, whether from an AI or a person, as seamlessly as possible. I’ll be monitoring the data closely and will report back on how the system behaves in the wild.
