Context
The client was a fast-growing D2C beauty brand with a carefully cultivated voice and a 14-person content team spread across product copy, email, and editorial. Off-the-shelf AI tools had quietly become a problem: writers were using them, brand reviewers were rejecting the output, and nobody was admitting how much time the back-and-forth was costing. The Head of Brand needed a tool that respected the voice rather than fighting it.
The Problem
- Generic AI outputs were breaking the carefully-built brand voice in subtle, repeated ways.
- Brand reviews were catching 60% of drafts for tone or messaging violations.
- Writers were spending roughly 30% of their week on ‘AI cleanup’ — work that did not appear on any time sheet.
- There was no institutional memory — every new writer was starting from scratch with the brand voice.
- The brand team did not trust AI in principle and needed proof, not promises.
Our Approach
We built voice into retrieval, not just into prompts. Every generation pulls relevant brand-approved assets into the context window, scores the output against the brand pillars, and surfaces the source assets that informed the draft. Writers can see why the system made the choices it did and can disagree, which trains the next round.
What We Engineered
- Brand-voice retrieval over 600+ approved past assets and the official style guide.
- Real-time tone scoring against defined brand pillars, delivered inline as the writer drafts.
- Structured templates for product copy, email, social posts, and long-form editorial.
- Inline citation of the brand assets that informed each generation — every claim and turn of phrase is traceable.
- A weekly ‘voice drift’ report that flags trends the brand team should know about.
Architecture and Workflow
- Interface: Chrome extension for in-context drafting plus a standalone editor for long-form work.
- Models: GPT-4o for generation; a custom voice classifier for tone scoring.
- Inputs: 600+ brand-approved assets, the style guide, banned-words list, current product catalogue.
- Stack: pgvector for retrieval, FastAPI backend, Next.js editor with inline citation rendering.
- Outputs: Drafts annotated with tone score and source citations.
- Time to launch: Easy-to-Medium — four weeks from kickoff.
Business Impact
- Writer output per week increased by 6.8x.
- Brand-review rejection rate fell from 60% to 9%.
- Time-to-publish dropped from three days to roughly six hours.
- New writer ramp time fell from four weeks to five days.
- The brand team began using the system as a training resource for new hires.
Why It Worked
The difference between a generic AI writer and one that sounds like the brand is retrieval done well. We engineered that difference into the foundation, then added the visible signals — citations and tone scoring — that earned the brand team's trust.
Risks Mitigated
- Voice erosion — enforced at generation time and monitored weekly.
- Hallucinated product claims — every claim is cited to a source asset.
- Adoption resistance — the tool earned trust by showing its reasoning.
Client Voice
“It is the first AI tool our brand team has actually asked us to give to more people, instead of asking us to turn off.”
— Head of Brand | D2C Beauty