Build a Brand Voice Model for Marketing Agency Clients

Two pieces of content can say the same thing and sound completely different. "We help teams move fast." vs "Built for velocity-obsessed teams who ship before they sleep." Same information, different voice.

Brand voice is what separates memorable brands from forgettable ones. Generic AI produces the first version. A fine-tuned model trained on the brand's approved content produces the second. This is one of the highest-margin deliverables an AI agency can build: $5,000-8,000 setup, $300-500/month retainer, 10-15 hours of work.

What Makes Voice Different From Prompt Engineering

A brand voice system prompt tells GPT-4 to "be conversational, use short sentences, speak like a founder." This instruction is processed new every call. The model is not calibrated to the brand's voice — it is trying to approximate instructions it has never learned from.

A fine-tuned brand voice model has seen hundreds of approved content pieces from this brand. It has learned their specific vocabulary (the words they use), their rhythm (sentence length patterns), their energy level (punchy vs measured), their referencing style (technical? pop culture? industry-specific?). These are not things you can capture in a system prompt. They are pattern-matched from the data.

The practical difference: A brand voice model generates copy that the client's marketing team immediately recognizes as "their" brand. No heavy editing. The system-prompted model generates copy that vaguely sounds like the direction but needs significant rework to feel authentically on-brand.

Data Collection

Source: Every piece of brand content that was approved and published:

• Website copy (homepage, landing pages, product pages, about us)
• Email sequences and newsletters
• Social media posts (approved/published, not drafted and rejected)
• Ad copy (winning variants)
• Long-form content (blog posts, case studies) written in brand voice
• Sales collateral

What to avoid:

• Rejected drafts (teaches the model what not to do, which can contaminate outputs)
• Content written by agencies or contractors in a style that was later revised
• Older content that predates a recent brand voice evolution

Size target: 200-600 pieces of content for initial training. If the brand has less than 200 pieces, supplement with annotated examples (the agency writes 50-100 examples in the brand's voice, reviewed by the brand).

Structuring the Dataset

Brand voice training requires examples of different content types. Use type tagging so the model learns that short social posts have a different structure than email subject lines, which differ from website hero copy.

{"messages": [
  {"role": "system", "content": "You are the content writer for [Brand], a B2B SaaS company building workflow automation for operations teams. Voice: direct, confident, slightly irreverent. No corporate speak. Short sentences preferred. Use 'you' not 'your team.' Speak like a thoughtful founder, not a marketing department."},
  {"role": "user", "content": "Write a LinkedIn post about our new integration with Slack. Content type: LinkedIn post (150-200 words). Topic: New Slack integration that shows real-time workflow status."},
  {"role": "assistant", "content": "Your workflow broke at 2pm on Friday.\n\nYou found out Monday morning.\n\nThat's the problem we just fixed.\n\nToday we're shipping real-time workflow status directly in Slack. Every trigger, every step, every failure — visible in the channels your team already lives in.\n\nNo more tab-switching to check what ran. No more 'did that workflow actually send?' Slack messages to engineering. No more Sunday night anxiety about Monday's data pipeline.\n\nWhen something breaks, you see it instantly. When something works, you know.\n\nOne channel. All your workflows. Real time.\n\nIt's live for everyone today. If you've connected Slack already, the integration activates automatically. New to Slack? Settings → Integrations → Slack. 30 seconds.\n\nShip less worry. [link]"}
]}

Include examples across all major content types: social, email subject, email body, headline, landing page, ad copy. The model needs to know the format conventions for each.

Training the Model

Base model recommendation: Mistral 7B Instruct for brand voice. It produces more natural, flexible prose than Llama models for creative writing tasks.

With 400-600 examples across content types, training takes 30-60 minutes. The fine-tuned model should:

1. Maintain the brand's vocabulary choices
2. Match their sentence rhythm and length patterns
3. Apply the appropriate register for each content type
4. Avoid vocabulary they never use (if the brand never says "leverage," the model should not say it)

Evaluating Brand Voice Quality

Standard accuracy metrics do not apply here. You need human evaluation.

Evaluation method — Blind Comparison Test:

1. Generate 20 content pieces with the fine-tuned model
2. Generate the same 20 pieces with GPT-4 + a well-crafted system prompt
3. Present pairs (without labels) to the brand's content lead
4. Ask: "Which sounds more like us?"

Target: Fine-tuned model wins 70%+ of comparisons.

Secondary metric — Edit Rate: Have the brand's team write 20 content pieces using the model. Track what percentage they publish without edits vs with light edits vs with major edits.

Target: 50%+ publish without significant edits.

Pricing and Delivery

Initial build: $5,000-8,000

• Includes: Data audit, dataset curation, model training, evaluation session with client, delivery of Ertas project + Ollama endpoint

Retainer: $300-500/month per brand

• Includes: Monthly usage, quarterly retraining with new content, voice consistency monitoring, model updates

What makes this premium pricing defensible:

• Generic AI tools are subscription commodities at $20-100/month
• A brand voice model is proprietary to this brand — it cannot be purchased off the shelf
• The agency delivers something no one else sells: this brand's voice, trained, deployed, and maintained

---

Further Reading

• Marketing Agency AI Opportunity — The full marketing vertical overview
• Content Agency AI Cost Reduction — Replacing content production API costs
• Fine-Tuned Copywriting Model for Agency Clients — Ad copy and conversion copy models
• AI Agency Retainer Model — Building recurring revenue from model maintenance