The challenge
The publisher was losing digital subscribers at a pace the product team could not stop. Their CRM was template-based: every subscriber got the same 12 variants across the lifecycle, and after a few weeks the trick was over. Engagement flattened, unsubscribes climbed, and the business could not personalise at scale without compromising the editorial voice that made the publication distinct.
The solution
We built an LLM messaging runtime that sits inside their existing Braze stack and replaces the template body with a per-recipient message generated live against a strict brand-voice evaluator. The evaluator was co-designed with the paper's editors — it scores generated copy on tone, accuracy, vocabulary, and a long list of brand red flags. No message is allowed to send unless it scores above threshold. Every outbound message is logged to an auditable store that the editorial desk can inspect at any time, and every generation cites the article(s) it references.
Why templates fail for news
News publishers are voice-first businesses. Readers subscribe because they trust a specific editorial voice. The moment a template reveals itself — the second identical email — the magic is broken. Personalisation is existential for this category, and template-based CRM cannot deliver it.
The editorial-safe generator
The generator is a small, carefully controlled runtime: it has access to the reader’s behaviour (articles read, sections, recency), the publication’s catalogue, and a strict brand-voice evaluator authored by the editors themselves. Every generated message runs through the evaluator before it is allowed to send. Off-brand copy does not ship.
The brand voice evaluator
Built over four weeks of workshops with the editors. It scores generated copy on:
- Tone and register
- Vocabulary (allowed, discouraged, forbidden)
- Factual anchoring (citations required to articles)
- Off-brand red flags (e.g. forbidden idioms, hype words, false urgency)
Messages below threshold are not sent. The editorial desk reviews failures weekly and the evaluator improves.
Governance
Every message is logged with its prompt, its reader context, its citations, the evaluator score, and its delivery receipt. The editor-in-chief can inspect any send at any time. When the AI Act requires us to show explainability, we can.
What the head of product said
“Our editors were nervous. They agreed to the project because the brand-voice evaluator was theirs. Now they are asking us to use it elsewhere in the product.”
Services deployed
- Custom AI Agents (messaging runtime, brand-voice evaluator, audit log)
- Knowledge Engineering (publisher catalogue grounding, citation enforcement)
- AI Governance