Insights from Grace Mulinski, Brand Content Designer & AI Trainer at Envive AI

For many brands, the biggest hesitation when considering integrating AI into their customer experience is protecting their brand voice: Will this serve as an extension of who we are? 

I’m referring to the brands where you see their name and immediately have a visual or a feeling in mind—where every touchpoint, from social media to product descriptions, has been built to make that recognition automatic. They’ve worked hard to make that recognition automatic, and that’s exactly why they’re the most at risk when deciding to include AI in their customer experience. For them, an out-of-the-box solution isn’t enough. The only real answer is to design the AI agent experience around the brand, from the ground up. 

This concept is something I’ve built a process around when designing conversational AI agent experiences. But how do I approach this process authentically? 

Phase One: Building the Brand Voice

First step: research. This phase is about learning the brand’s anthropology: understanding not just what a brand says, but how it says it. 

My starting points are always the same: the website, their social media presence, and any existing copy that best represents their voice. I'm looking for patterns. What words do they reach for repeatedly? How long are their sentences? Do they write to their customer directly, or do they narrate? Is the tone warm, edgy, playful, authoritative? Do they use emojis—and if so, which ones, and how often? 

For one women's fashion brand I worked with, this research revealed a voice that was bold, bordering on sultry, and deeply intentional: to embolden young women to embrace their bodies. Their customers weren’t just buying clothes; they were buying into a community that made them feel empowered. Every piece of copy, from product descriptions to Instagram captions, was written to make them feel like the most confident person in the room. That insight became the foundation of their agent’s persona.

‍

The output of this phase is the Brand Voice: a set of documents and other assets that captures the brand's personality, tone guidelines, key messaging pillars, language style, and specific phrases to both embrace and avoid. It defines who the AI is, and is a core component that determines all aspects of the agent’s behavior.

‍

Getting this layer right is what separates an AI agent experience that feels native to a brand from one that feels tacked on. 

Phase Two: Testing 

After presenting the woman’s fashion brand with their cultivated sales agent, their internal testing came back with unexpected feedback: the agent was too branded. Bordering on cringey. The voice I had carefully researched and built into the prompt was being over-indexed on. It was reaching for the brand's most recognizable phrases so repetitively that it came across as inauthentic. A human copywriter uses those phrases sparingly, instinctively, but the AI didn't have that instinct. 

The fix was deliberate suppression—I designed explicit guardrails that were added to the sales agent. This produced a result that felt more seamless, less performative. Feedback like this is invaluable. It's how we achieve an agent that feels like an extension of the brand.

‍

A significant part of our testing involves creating test cases that we run through LLM-as-a-judge evaluations; in this process, a separate agent scores our sales agent’s responses against defined criteria. These test cases are designed by industry and product category; this is integral because a brand that sells outdoor backpacks and one that sells luxury handbags will both use the word "bag," but their customers ask entirely different questions and expect entirely different answers. 

What an LLM-as-a-judge does well is evaluate accuracy: is the agent returning the right product, the correct sizing information, the appropriate collection? What it can't reliably judge is whether the agent sounds like the brand. For that, we rely on Human-in-the-Loop evaluation. Personality, tonality, and the feeling a response conveys require a human to read it and determine whether it's right. 

Getting It Right 

Designing an AI experience that truly reflects a brand is not a shortcut. It is a process that requires research, precision, and a willingness to iterate when the first version isn't quite right. That iteration looks different every time. A suppressed phrase here, a restructured prompt there, a round of test cases that reveals if the agent is answering the right question the wrong way. The work is granular, and that granularity is crucial. 

What I've learned through this process is that the gap between an AI agent experience that feels native to a brand and one that feels like a generic chatbot is a matter of care. Care in the research, care in the prompting, care in knowing when to suppress and when to amplify. 

Brands that hesitate before adopting AI are right to do so. The decision deserves thoughtfulness. But that hesitation doesn't have to be a barrier. With the right process behind it, AI experiences can do what great brand voices have always done: make a customer feel like they're talking to someone who genuinely gets them.

‍