Brand Voice at Scale: Teaching AI to Sound Like Your Business

There's a version of AI-generated content that sounds like AI-generated content. Smooth, generic, slightly corporate. It uses words like "leverage" and "solutions." It could have been written by anyone or, more accurately, by no one in particular.

Then there's a version that sounds like your business. Same information, completely different voice. The difference isn't the model — it's whether the model knows who it's supposed to be.

Brand voice profiles are how Workforce Wave makes AI-generated content — from outbound messages to post-call summaries to knowledge base updates — sound like the specific business it's representing, not like AI in general.

What a Brand Voice Profile Contains

A brand voice profile is a structured document that captures the distinctive characteristics of how a business communicates. It's stored per-agent and applied whenever the agent generates text content.

{
  "agent_id": "agt_xyz789",
  "brand_voice": {
    "tone_descriptors": [
      "warm and approachable",
      "clinically confident without being cold",
      "uses plain language over medical jargon",
      "reassuring, especially with anxious patients"
    ],
    "vocabulary_preferences": {
      "preferred": ["appointment", "team", "care", "we'd love to help"],
      "avoid": ["slot", "resource", "utilize", "leverage", "solutions"],
      "industry_specific": ["cleaning", "checkup", "X-rays", "Dr. Chen"]
    },
    "sentence_structure": {
      "avg_sentence_length": "short-to-medium",
      "uses_contractions": true,
      "question_style": "direct and friendly, not formal",
      "list_preference": "prose over bullet points in conversation"
    },
    "sample_phrases": [
      "We're happy to get you in — let me check what's available.",
      "Dr. Chen will take great care of you.",
      "No worries at all — that's a common question."
    ]
  }
}

The profile is applied as a style layer on top of any content the agent generates. When Workforce Wave writes a post-call summary, when the KB sync proposes an update, when the agent drafts an outbound appointment reminder — the brand voice profile shapes how that content sounds.

The Workforce Wave Brand-Voice-Extract Workflow

For new partners who want to establish a brand voice profile from the start, Workforce Wave includes a brand-voice-extract workflow. The input is 3–10 sample content pieces from the business — website copy, past emails, marketing materials, social posts. The output is a fully structured brand voice profile ready to apply.

The workflow is accessible via the admin UI or directly through the API:

POST /v2/agents/{agent_id}/brand-voice/extract
Content-Type: application/json

{
  "samples": [
    {
      "type": "website_copy",
      "content": "At Sunshine Family Dental, we believe a healthy smile should be stress-free. Whether you're coming in for a routine cleaning or something more involved, Dr. Chen and our team will make sure you feel at ease every step of the way."
    },
    {
      "type": "email",
      "content": "Hi Sarah — just a reminder that your appointment is this Thursday at 10am. If anything comes up, give us a call at (843) 555-0182 and we'll find another time that works. See you soon!"
    },
    {
      "type": "social_post",
      "content": "Happy to welcome the Garcia family to our practice this week! It's always special when parents bring their kids in for their first visit. 🦷"
    }
  ],
  "vertical": "dental"
}

# Returns an operationId — extraction runs async, typically 15-20 seconds

The extraction LLM analyzes the samples for tone, vocabulary patterns, sentence structure, formality level, use of contractions, and characteristic phrases. It also runs vertical-aware analysis — a dental practice's use of the word "cleaning" versus "prophylaxis" is a meaningful signal about how clinical or accessible the practice wants to sound.

The output is a brand_voice object that can be previewed, edited, and applied to the agent — or used as the default template for all agents a partner provisions within a given vertical.

Partners who provide more samples get better extractions. Three samples establish the basics. Eight to ten samples give Workforce Wave enough signal to capture subtleties — the way a specific practice acknowledges anxiety versus how a different practice in the same vertical does it, for example.

How Feedback Loops Improve Future Generation

The brand voice profile isn't static. Every time a human edits AI-generated content — correcting a phrase, adjusting tone, changing how something is worded — that edit is a signal.

The brandvoicefeedback table captures these signals:

-- Written whenever a human edits AI-generated content
INSERT INTO brand_voice_feedback (
  agent_id,
  content_type,        -- 'post_call_summary' | 'outbound_message' | 'kb_update'
  original_content,    -- what the AI generated
  edited_content,      -- what the human changed it to
  edit_type,           -- 'vocabulary' | 'tone' | 'structure' | 'factual'
  edited_by,           -- user_id of the human who made the edit
  created_at
);

On a weekly cycle, Workforce Wave's brand voice optimizer reviews accumulated feedback for each agent and proposes updates to the brand_voice profile:

• If an editor has consistently replaced "slot" with "appointment" across 12 post-call summaries, "slot" moves to vocabulary_preferences.avoid
• If edited messages are consistently shorter than generated ones, sentencestructure.avgsentence_length adjusts
• If a specific phrase appears repeatedly in edits ("We're happy to get you in"), it gets added to sample_phrases

The proposed updates are surfaced in the admin UI as a diff — old value versus proposed value — and require human approval before they apply. Partners can accept all, reject all, or accept selectively.

The loop is: AI generates → human edits → Workforce Wave learns from the diff → AI generates better → fewer edits needed.

Brand Voice at the Partner Level

Most partners don't just manage one agent — they manage dozens or hundreds. Brand voice profiles can be set at three levels:

Agent level — specific to one deployed agent, highest specificity Client level — applies to all agents deployed for a single client (useful for multi-location businesses) Partner template level — the default applied to new agents provisioned under the partner's slug, in the absence of more specific overrides

This means a partner who has invested in a strong brand voice template for their dental vertical automatically gets that quality applied to every new dental practice they onboard. The investment in one good template compounds across the whole fleet.

Partners who treat brand voice as a feature — rather than a default they leave untouched — consistently produce content that their clients recognize as sounding right. That recognition is the difference between an AI that clients see as a tool and an AI that clients see as genuinely representing their practice.

The technology is the same either way. The voice is what makes it theirs.

---

This concludes The White-Label Playbook series. For a technical deep dive on the provisioning pipeline, see How Workforce Wave Builds a Voice Agent from a URL in 90 Seconds.