Brittany Maslo
A support lead opens the weekend queue and finds the same complaint three times, each from a different customer: the chatbot sounds like a chatbot. One person called it polite but weirdly stiff. Another said it felt like talking to a brochure. The third asked, flatly, whether a human was available.
The underlying model is fine. It answers accurately, follows the prompt, stays on topic. What it fails to do is sound like the brand around it, and customers pick up on that within two exchanges. That’s the decision product teams are wrestling with right now: ship the generic voice that comes in the box, or put in the work to train a model that sounds like the company it works for.
Decide Whether Voice Is a Vendor Setting or a Product Decision
The first choice sits earlier than most teams realize. Voice gets filed as a configuration task — a sentence or two in the system prompt, a dropdown labeled “tone,” maybe a few style rules copied out of the brand book. That framing hands the most audible part of your product to whichever vendor you happened to pick.
A Harvard Business Review piece on AI voice argues that this is a behavioral variable the business should own, rather than an infrastructure detail. Own it, and the budget conversation shifts — owning the voice can mean anything from a tighter prompt to a private LLM trained on your own material. File it as a vendor setting and you get what the vendor ships to everyone else.
Decide How Far to Go Down the Customization Stack
There are roughly four moves available, and each one costs more and controls more than the last. The right answer depends on how distinctive your voice really is and how much of your business runs through the model.
- Prompting. A detailed system prompt, a few examples, maybe a style guide pasted in. Cheap, fast, and the ceiling is low. Good enough for internal tools and early pilots.
- Retrieval. Point the model at your own content — help docs, past tickets, approved copy — so answers are grounded in your material. This does more for accuracy than voice, but it helps.
- Fine-tuning. Train the model on paired examples of inputs and the outputs you’d actually ship. Voice starts to feel owned instead of borrowed.
- A private model. Host the weights yourself, train on proprietary data the vendor never sees, and keep the behavior inside your perimeter. Teams going this route usually bring in a partner rather than standing it up from scratch.
Most teams land between the second and third rung. The mistake is skipping to the fourth because it sounds serious, or staying on the first because it’s easy.
What Does “Your Voice” Actually Mean in Training Data?
Fine-tuning a model on your voice means building a dataset that shows the model, over and over, what a good answer from your brand looks like. OpenAI’s supervised fine-tuning documentation describes the shape of the work: prompts paired with the responses you want, enough of them to shift the model’s defaults toward your style.
The pipeline is the easy part; the hard part is agreeing on what the voice is. Pull a hundred recent support replies and you’ll find three different writers, two different reading levels, and at least one tone nobody wants to repeat. Curating the dataset is where the brand decisions get made.
If marketing, support, and product can’t agree on what a good reply reads like, the model can’t either. Pick the reviewers before you pick the training set.
Decide Where the Human Review Lives
A trained voice is still a work in progress. Someone has to read what the model is saying in production, flag the drift, and feed corrections back into the next training run. Put that loop on the roadmap before launch, not after a customer screenshots something awkward.
The teams that get this right treat voice the way they treat any other product surface: shipped in versions, measured against examples, improved on a schedule. The teams that don’t end up with a chatbot that sounds like everyone else’s, and a quarterly meeting about why engagement is soft.
Pick the Smallest Version That Still Sounds Like You
The temptation on a project like this is to over-build — bigger model, more data, custom everything. Ship the smallest version that passes a blind read against your own writing samples. If a reviewer can’t tell your human-written reply from the model’s reply, you’re done for now. Scale up when the product needs it, not before.
