We don't measure AI adoption by tool count — we measure it by hours saved on the least interesting part of the day. That's a much stricter test.
What we automate
- Concept exploration at the whiteboard stage
- First-draft cutdowns and language variants
- Tagging, transcription, subtitle timing
- Analytics summarisation for weekly reviews
- Competitive scans and share-of-voice snapshots
What we don't
Strategy. Casting. Final copy. Client conversations. Anything where taste, accountability or trust is the product. We've watched teams outsource these to models and quietly lose the plot of what their brand sounds like.
The rule we've settled on
AI is allowed to write the first draft and the second-to-last draft. Humans write the last one. That single rule preserves speed and preserves voice — and it survives audit when a client asks 'who actually wrote this?'
"AI didn't make our team faster. It made our senior people spend more time on senior work."
Guardrails we ship with every AI workflow
- A named human owner for every AI-touched deliverable
- A prompt library versioned like code
- A red-line list of things AI is never allowed to touch (client PII, unreleased strategy, financial claims)
- A weekly retro on what the model got wrong
