How We Trained a Custom GPT for a Mills 50 Agency’s Brand Voice

<i>A small marketing shop in Orlando’s Mills 50 district was drowning in rewrites. We built a custom GPT that finally made junior writers sound like the brand — without the buzzwords.</i>

(Client details are anonymized and some specifics composited at the client’s request.)

A few months back, the owner of a small marketing agency in Orlando’s Mills 50 district called me up. She had six employees — three junior writers, two account managers, and a designer. Her problem was straightforward: every piece of content that went out to clients sounded like a robot wrote it. Press releases, social posts, email newsletters — all that stiff, corporate tone that made her wince. She was spending 15 hours a week rewriting copy herself. The junior writers were solid people, but they’d default to jargon and buzzwords every single time. She needed a way to train them — or train an AI — to match her brand voice.

I’d seen this before. Small agencies usually have a strong voice in the founder’s head, but it never makes it to the rest of the team. Traditional solutions — style guides, templates, endless feedback loops — only get you so far. This owner had tried a shared document of “approved phrases” and weekly editing sessions. Neither stuck. Writers would slip back into press-release mode within days.

So we built something different: a custom GPT that ingested her brand voice, her style guide, and 200 pieces of approved content. The goal was simple — give the junior writers an AI assistant that’d catch hype words before they hit the page.

The Situation: A Brand Voice That Existed Only in One Person’s Head

The agency specialized in helping local service businesses — plumbers, electricians, landscapers — tell their stories. Their brand voice was warm, direct, and a little irreverent. Think “your neighbor who knows a thing or two about pipes,” not “we are a leading provider of plumbing solutions.” But that voice lived entirely in the owner’s head. The style guide was a two-page PDF with vague rules like “be friendly” and “avoid jargon.” It wasn’t helping anyone.

When a junior writer drafted a blog post for a client, the owner would mark it up in red pen: “too formal,” “sounds like a press release,” “this word is banned.” She’d rewrite entire paragraphs. Each piece of content needed two to three rounds of revision. The agency was billing clients for work that cost more to produce than it earned.

The owner tracked her time: 12 hours per week editing and rewriting. That’s 48 hours a month — basically a full workweek — spent fixing tone. She wanted something that could catch tone problems before the draft even landed on her desk.

What They’d Tried Before (And Why It Didn’t Work)

The agency had taken a swing at three different approaches before calling me:

• A style guide update: They rewrote the style guide with more examples. The junior writers read it, nodded along, and went right back to writing corporate-speak. That guide sat in a Google Doc that nobody ever opened.
• Template libraries: They built templates for common content types — client spotlight, service announcement, seasonal tip. The templates helped with structure, but the voice still slipped. Writers filled in the blanks with their own stiff language.
• Weekly editing workshops: The owner spent an hour every Friday going over edits. It helped a little, but the feedback didn’t stick. By Monday, the writers had already forgotten the nuance.

None of these addressed the real issue: the brand voice wasn’t codified in a way that could be applied consistently. It was a feeling, not a system.

The AI Work: Building a Custom Brand-Voice GPT

We decided to build a custom GPT using OpenAI’s GPT-4 with a retrieval-augmented generation (RAG) architecture. The idea was straightforward — give the model a “memory” of the brand’s approved content and style rules so it could generate copy that matched the voice and also critique drafts written by humans.

Step 1: Gathering the training data. The owner gave us access to 200 pieces of content she considered “on voice.” Blog posts, social media captions, email newsletters, website copy — the whole range. We also got the two-page style guide and a list of “banned words” — about 40 terms that just didn’t fit. We cleaned the data, removing anything that could identify the actual business.

Step 2: Building the RAG pipeline. We used a vector database (Pinecone) to store embeddings of each approved piece of content. When a user asked the GPT to write something, it’d first pull back the three most similar examples from the approved set. That gave the model concrete examples to mimic. We also embedded the style guide rules as additional context.

Step 3: Fine-tuning guardrails. We wrote a system prompt that explicitly listed the banned words and told the model to flag any that appeared in a human-written draft. The GPT would output a score (0–100) for “voice fit” and highlight problematic phrases. We tested it against 50 sample drafts from the junior writers and iterated five times to cut down false positives.

Step 4: Human-in-the-loop for tone. We deliberately kept the owner in the loop for final approval. The GPT could generate drafts and flag issues, but the owner reviewed the output before it went to the client. This wasn’t about replacing her — it was about reducing her editing time. The GPT handled the first pass, catching about 80% of the tone problems. The owner only needed to review the remaining 20%.

Step 5: Integration into their workflow. We built a simple web interface using n8n — an automation tool — that connected to the agency’s Slack channel. A writer could type “/brand-voice draft [topic]” and get a draft back in seconds. They could also paste a human-written draft and get feedback. The system logged every request so we could track usage and quality over time.

Where We Deliberately Kept a Human in the Loop

Look, I’m a believer in AI as an assistant, not a replacement. For this project, we kept the owner involved in two key spots:

• Final approval: The GPT never sent content to a client. The owner reviewed every piece. This caught subtle tone issues — a joke that fell flat, a reference that didn’t land — that only a human would catch.
• Edge cases: When a topic was outside the training data (like a new service offering), the GPT’s output wasn’t as reliable. The owner wrote those drafts from scratch, then added them to the approved set to expand the model’s knowledge.

This approach gave the agency speed without sacrificing quality. The owner’s editing time dropped from 12 hours per week to about 4 hours — a 67% reduction. The junior writers felt more confident because they had a tool that caught their mistakes early.

“The GPT caught hype words I didn’t even know were banned. It saved me from embarrassing myself in front of clients.” — Junior writer at the agency

Measured Results: What Actually Happened

We tracked metrics for six weeks after launch. Here’s what the numbers showed:

• Editing time: The owner’s weekly editing time dropped from 12 hours to 4 hours. That’s 8 hours saved per week — enough to take on another client.
• Revision rounds: The average number of revisions per piece fell from 2.5 to 1.2. Writers got it right the first time, more often.
• Voice fit score: We asked the owner to rate each piece on a 1–5 scale for voice consistency. Before the GPT, the average was 2.8. After, it was 4.3.
• Banned word count: The GPT flagged an average of 2.3 banned words per draft in the first week. By week six, that dropped to 0.4 — writers were actually learning to avoid them.
• Client satisfaction: The agency’s clients didn’t notice the change directly, but the owner reported fewer revision requests from clients — down from 3 per month to 1.

One unexpected benefit: the junior writers started using the GPT as a learning tool. They’d write a draft, get feedback, and internalize the rules. Within a month, their unaided writing improved. The GPT was acting like a coach, not just a crutch.

What We’d Do Differently (Honest Caveats)

Honestly, this project wasn’t perfect. A few things were harder than expected:

• Data cleaning took longer than building the model. The 200 approved posts were scattered everywhere — Google Docs, PDFs, email threads. We spent two weeks just standardizing the data. Next time around, I’d ask the client to dump everything into a single folder upfront.
• The GPT struggled with humor. The brand voice relied on a dry, witty tone that didn’t translate well to AI. The GPT would sometimes generate jokes that felt forced or off-key. The owner had to rewrite those sections herself. We tried fine-tuning the model on more humorous examples, but it never fully nailed it.
• Banned words are a moving target. The owner kept adding new words to the list as she saw them in drafts. The system could handle updates, but it meant the GPT’s feedback wasn’t always consistent. We eventually built a simple interface where she could add banned words herself.
• Small sample size. With only six employees and 200 posts, the training data was limited. The GPT performed well for common content types but struggled with niche topics. A larger dataset would’ve helped.

Despite these caveats, the project was a clear win. The agency saved time, improved quality, and gave their junior writers a tool that made them better.

What This Means for Other Small Businesses in Central Florida

Here’s the thing — this approach isn’t just for marketing agencies. Any business that produces content — blogs, emails, social posts — can benefit from a custom GPT trained on their brand voice. The key is starting with a clear idea of what “on voice” looks like, and keeping a human in the loop for the final check.

If you’re a small business owner in Orlando, you’re probably facing the same problem: your team doesn’t write like you do. A custom GPT can help bridge that gap. It won’t replace your judgment, but it’ll catch the mistakes that eat up your time.

We offer an AI Readiness Assessment to help businesses like this one figure out where AI can make the biggest impact. For companies that already know they need a custom tool, we provide Fractional AI Officer services to scope and build solutions like this brand-voice GPT.

The Mills 50 agency is now producing content faster, with less stress. The owner spends her saved time on strategy and client relationships — not rewriting press releases. That’s what matters.

If you’re ready to stop sounding like a robot, contact us. We’ll help you build an AI that sounds like you.

"The GPT caught hype words I didn't even know were banned. It saved me from embarrassing myself in front of clients." — Junior writer at the agency

Frequently asked questions

How long does it take to train a custom GPT on brand voice?

For a small business with 100–200 approved content pieces, the initial build takes about two to three weeks. Most of that time is spent cleaning and organizing the training data. Once the model is live, you can start using it immediately.

Do I need technical expertise to use a custom GPT?

No. We handle the technical setup — embedding the data, building the interface, and integrating it into your workflow. Your team just needs to know how to type a prompt or paste a draft. We provide training and support.

Can the GPT handle multiple brand voices for different clients?

Yes, but it requires separate models or a more complex setup. For an agency with multiple clients, we recommend building a separate GPT for each brand voice, or using a single model with a prompt that specifies which voice to use. We can help you decide the best approach during a consultation.

What if my brand voice changes over time?

The model can be updated by adding new approved content and modifying the banned word list. We recommend a quarterly review to keep the voice consistent. Our system makes it easy to add new examples without retraining the entire model.

How much does a custom brand-voice GPT cost?

Costs vary based on complexity and data volume. For a small agency like the one in this case study, the initial setup runs between $2,000 and $5,000, plus a monthly hosting fee of around $100–$300. We provide a detailed quote after an initial assessment.

Will the GPT replace my writers?

No. The GPT is designed to assist writers, not replace them. It catches tone issues and generates drafts, but human writers still add creativity, strategy, and final judgment. In our case study, the junior writers became more confident and improved their own skills.