PAGE
PROGRESS
0%
·10 min read

How to Train AI on Your Brand Voice and Tone

How to Train AI on Your Brand Voice and Tone

make ai sound like you mygom article

How to Train AI on Your Brand Voice and Check It Stays On-Brand

To train AI on your brand voice, start with three things: 20 to 50 pieces of writing that already sound like you, a short written list of your voice rules, and a second model that grades every draft against those rules. Most teams get there with prompts and examples. Fine-tuning comes later, and only when the output stays inconsistent.

The second part is the one most guides skip. A model that writes in your voice is useful. A model that can tell you when a draft has drifted off-brand is what lets you publish without rereading every line.

Below is the process we use, step by step, plus what we learned building an AI tool that writes our own commercial proposals.

Do you need fine-tuning to train AI on brand voice?

Usually not at the start. OpenAI's own guide to optimizing LLM accuracy (opens in new tab) says "Prompt engineering is typically the best place to start", and recommends adding few-shot examples to improve consistency before anything else.

The same guide lists the cases where fine-tuning starts to make sense, and one of them is exactly this topic: "the tone or style of speech is not correct". So the order is:

Each step is cheaper than the next one, and each one produces the material the next one needs.

Step 1: Collect writing that already sounds like you

Gather website copy, sales emails, case studies, social posts and support replies. Pick the pieces your team is proud of, the ones customers answered or shared. A few dozen strong samples teach a model more than a thousand generic lines.

A quick filter: read a sample out loud and ask whether your CEO would say it in a meeting. If the answer is no, leave it out.

Checkpoint: open three random samples from the folder. If any of them could have been written by any company in your industry, keep curating.

Step 2: Write the voice down as rules

Adjectives like "friendly" or "confident" give a model almost nothing to work with. Rules do. Write a one-page voice sheet with:

We keep our own content rules in one document like this. Most of the lines on it started as a correction to a draft that sounded wrong.

Step 3: Give the model examples, with labels

Put 3 to 5 of your best samples straight into the prompt, next to the voice sheet. If you write for several channels, label each sample, so the model learns that an email and a LinkedIn post can share one voice and still sound different:

javascript
{ "text": "Your next shipment leaves the warehouse on Thursday.", "tone": "direct", "channel": "email" }
{ "text": "One missing zero. 15,000 ordered, 1,500 delivered.", "tone": "playful", "channel": "linkedin" }

Checkpoint: generate drafts for prompts from your real daily work: a reply to a client, a product update, a follow-up email. Put each draft next to a human-written piece on the same topic.

How to train a model to recognize brand tone in AI outputs

This is the step that makes the rest scalable. The quickest way to recognize brand tone is a second model with a rubric that grades every draft, with no extra training.

Anthropic's guide to building evaluations (opens in new tab) describes this method, LLM-based grading, as "suitable for complex judgment" and gives a tone example: a model rates each response on a scale from 1 to 5. Three of its recommendations matter most for brand voice:

The same guide notes it is generally best to grade with a different model from the one that wrote the draft. A simple grader prompt looks like this:

plaintext
Grade this draft against our brand voice rules.

<rules>{voice_sheet}</rules>
<draft>{draft}</draft>

List every rule the draft breaks, then give a score from 1 to 5.
5 = could be published as written. 1 = rewrite.

Checkpoint: before you trust the grader, give it 20 drafts your team has already scored by hand. If its scores disagree with your team's, the rubric is too vague. Rewrite the rule it got wrong and test again.

When fine-tuning is worth it

Fine-tune when steps 1 to 3 plus the grader still leave you with uneven drafts, and you have enough consistent examples to train on. By then you have the two things fine-tuning needs: a cleaned, labeled dataset from step 3 and a grader from the previous step to measure whether the trained model is better.

Keep a part of your examples out of training. OpenAI recommends "maintaining a hold-out set to use for evaluation following training to detect overfitting." Run the grader on the hold-out set before and after training and compare the scores.

If you use Hugging Face, the training setup can start this small:

python
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=8,
)

Start with the platform defaults and change one setting at a time, so you can see what each change does to the grader's scores.

What we learned building our own proposal generator

We applied the same idea to the documents we write most: commercial proposals. Our AI Proposal Generator (opens in new tab) learned our structure and tone from proposals our team had already written and won with, and pulls prices from our rates instead of a spreadsheet.

Preparing a proposal went from 3 to 4 hours to 30 to 60 minutes, which is 70 to 85% less time. Every proposal now has the same structure and pricing logic, whoever prepares it. A person still reviews the draft and decides what goes out.

The lesson for brand voice: the tone came from the reference files, real proposals that had already won clients. And brand voice reaches far beyond marketing. Any repeated document your team writes, a proposal, a specification, a reply to a client, has a voice, and an AI system can learn it the same way.

For the content side, we wrote about how we built an AI content generator plugin for PayloadCMS (opens in new tab). And our AI Business Brain (opens in new tab) keeps the company context in one place, so a new draft doesn't start from zero.

Frequently asked questions

Can AI learn my brand voice? Yes, if you give it real examples and written rules. A model with only adjectives like "friendly and professional" will write like everyone else. A model with 3 to 5 of your best pieces and a list of banned phrases gets much closer on the first draft.

How many examples do I need to train AI on brand voice? For prompting, 3 to 5 strong examples per channel is a good start, picked from a library of 20 to 50. For fine-tuning, you need more, all in the same format. Quality matters more than volume: one off-brand sample in the training set teaches the model the wrong thing.

How do I get an AI support agent to match my brand's tone of voice? The same way: voice rules plus real support replies as examples, and a grader that checks replies before they go out. Support has one extra rule worth writing down: when the agent isn't sure, it should hand the conversation to a person.

Does brand voice survive translation? Only if you check it in each language. Write the voice sheet for each language separately, because phrases that sound natural in English often sound translated in another language. Then run the grader on the translated drafts too.

How long does it take? A voice sheet, examples and a grader can be set up and tested within days. Fine-tuning adds data preparation and several rounds of testing, so plan for longer.

How Mygom can help

We build AI systems that write and check repeated documents in your company's own voice: proposals, client replies, reports. If you want to see where that would save your team the most time, get in touch (opens in new tab).

Gabriele J.

Marketing Specialist

Connect on LinkedIn

Let’s work together

Let’s work together

Ready to bring your ideas to life? We're here to help.