The same core sports journeys exist in three brand versions, each with its own voice. Checking them for consistency meant comparing hundreds of touchpoints side by side. AI could speed that up, but it doesn’t know a brand’s voice unless you teach it, and a confident wrong suggestion is worse than none.
Designed the audit framework first. Five core sports journeys (football, horse racing, tennis, basketball and golf), each assessed across four dimensions: UX copy and microcopy, navigation and IA labels, error and empty states, and onboarding and tooltips, with 40 touchpoints per sport.
Built a custom Claude skill grounded in real brand guidance. It reviewed each brand’s version of a journey against that brand’s actual tone of voice documentation and our shared content conventions. When Betfair’s guide shifted towards an “opinionated sports fan” voice, I updated the benchmarks to match.
Set rules for where brand voice gives way to clarity. On high-stakes actions, especially anything involving money, clarity beats personality. “Place bet” over a personality-led CTA, even on Paddy Power.
Kept a human at the wheel. The AI sped up the review and flagged issues. I checked its findings, investigated the causes and decided what to do.
Built the output for action. A structured spreadsheet with ratings, consistency flags, priority and status tracking, a summary dashboard, and per-brand quick-reference rules. Easy to amend and share with wider team members and leadership.
The most impactful finding was the inconsistent conventions across the three brands, I was able to trace them to two root causes:
Changes that hadn’t been applied across the board: updates made in one place but not carried through everywhere
Content added directly at build, without content design seeing it
That diagnosis turned a list of fixes into a process problem we could solve, by giving content design a clear way into the work before build.
This was all presented internally as a demonstration of how content design can use AI responsibly, something I'm keen to always keep front of mind.
AI is good at volume; reading hundreds of screens and spotting patterns. It’s not good at knowing when Paddy Power’s cheekiness helps and when it gets in the way, or at asking why something went wrong. The value came from designing the system around that: real brand guidance in, clear rules for judgement calls, and a person making the decisions.
It’s an approach I’ve been developing for a while. At PokerStars I built a similar system using ChatGPT. This has been the beginning of how I expect to embed AI into processes, I have more plans to develop usage of it further.