Case Study - On-Brand Content Generation with a Human in the Loop

Overview

Every organisation has a distinct voice, a set of internal perspectives, and audiences it needs to speak to differently. Generic AI writing tools ignore all of that — they can't reach into a company's proprietary knowledge, they don't know the difference between a CEO's LinkedIn post and a product update, and they force someone to rewrite nearly every draft. The market for "AI writing tools" is crowded, but the products in it all share the same gap. They can write — but they can't know. The fix isn't a better model; it's a better system around the model.

We built the first sprint as a proof-of-concept pipeline demonstrating the core loop end-to-end. A tagged knowledge base stores source documents with searchable keywords and feeds a retrieval layer that surfaces the right context for any given generation. A Voice × POV × Audience configurator lets the operator select a Voice (e.g. CEO), a Point of View, and a target Audience (e.g. LinkedIn public); the pipeline composes prompts dynamically against those selections so different combinations produce genuinely different outputs. Every generation is preceded by a retrieval pass over the organisation's own material, so the output reflects the brand's positioning rather than the model's training distribution. A simulated audience agent reads the draft as the target reader would and gives the operator structured feedback before the content ships. A provider layer sits in front of generation so the underlying model is swappable as the frontier moves — without rewriting the pipeline.

LinkedIn posts were the deliberate first surface: short-form, voice-sensitive, audience-explicit — the format where generic AI writing fails most visibly. The knowledge base is treated as the long-lived asset, the model is treated as the swappable component, and the human is treated as the irreplaceable arbiter of what reaches the audience. That ordering matters — it's what keeps Robin from drifting into the same flat generic tone every other content tool produces.

What we did