AI copy helped, but the offer mattered more
What I learned about AI while personalizing offers for 100M+ subscribers.
At Indosat we tested offer copy written with ChatGPT for each customer type against our usual copy. The ChatGPT versions did about 5% better. That's useful, and I'd do it again.
But it was a small part of the result.
What mattered more was which offer each person saw, and when. We used next-best-offer logic on the Tech Mahindra billing platform, with RFM and churn scores we built with data science. If no model matched, we showed a default offer. Over time this raised ARPU by about 5%.
Timing mattered too. The best moment to offer more data is right after someone checks how much they have left.
We also had to follow rules. Indonesian government contact rules limited how often we could message each person, and consent rules decided who we could message at all.
Now I build my own products with Claude Code, and I see the same thing. AI makes the work faster. But I still have to decide what to build, who it's for, and when to show it.