Building an AI content system at scale
Moving a 16-person team beyond ad-hoc AI usage to a structured system with scenario-specific prompts, dedicated tools and a mandatory human review layer.
Moving a 16-person team beyond ad-hoc AI usage to a structured system: scenario-specific prompts, dedicated tools and a mandatory human review layer.
Metrics and outcomes
The challenge
AI tools had become a daily part of the team's workflow, ranging from polishing drafts and generating variations to keeping up with concurrent launches and Figma reviews. Without a structured system behind it, AI copy fell into the same failure mode: error-free but flat, inconsistent and disconnected from brand identity.
A single style guide didn't hold up. Output became inconsistent whenever it lacked specific context, and that compounded as the team and product surface grew. The user-facing risk was real: unreviewed AI copy reaching production could fail exactly where it mattered.
The problem to solve: build a system that let a 16-person team use AI at scale without losing brand consistency or shipping unreviewed copy.
Gathering insights
Moved beyond ad-hoc prompting
I established early that isolated, one-off prompts weren't going to hold up across a 16-person team. Without something more structured, like reusable agentic models or skills, consistency would break down the moment more than one person touched the same kind of task.
Audited what already existed
I then gathered the artefacts already in place, like the editorial style guide, UI guidelines and terminology repository, and used them as the reference the system would be built against. Most of the gaps weren't missing documentation; they were documentation nobody had connected to how AI was being used day to day.
Understood where it was actually breaking
I sat down with product managers to find out where inconsistency showed up in practice and where turnaround speed mattered most to the business, rather than assuming which scenarios needed fixing first.
Set baseline metrics before building anything
Before changing the workflow, I established what "better" would actually mean, which came down to lead time, consistency, ease of use, a content-quality baseline and a way to evaluate AI-generated A/B variations. All of it tied back to solving real problems and creating business value.
Ideation
Built an on-brand, structured AI system
Scenario-specific reference sets cover tone of voice, error messages and onboarding flows separately, each with worked examples rather than abstract principles: tone pillars like human, insightful, confident and witty anchored to concrete scenarios.
Built dedicated prompts and tools
Gems for Google Gemini and GPTs for ChatGPT were built for specific tasks: push notifications, translation, UX writing, long-form content, localisation, paid ads, banner copy and more. No general-purpose prompts asked to do everything.
Built a mandatory human review layer
Every AI-assisted string was checked against the live Figma flow and user journey before production. Copy that reads well in isolation often fails in context, and the review layer caught what the prompt couldn't. AI accelerated variation generation for A/B tests, with human judgement kept as the deciding layer on what shipped.
Rolled it out with the team, not at them
I trained the team on the new system directly rather than issuing it as a mandate, working through real examples from their own output so the reference sets and tools reflected how the team already thought about tone. That way the new system never read as a decision imposed from above.
Results
- 70%+ increase in content production velocity while staying on-brand
- Adopted across close to 40 domains, each with its own variations
- Near-instant turnaround for time-sensitive communications
