The challenge
A US wireless provider wanted to address customer concerns raised on Facebook, Twitter, blogs and company forums. It used Attensity 360 and Radian6 to capture posts.
The goal was to classify each post accurately as positive feedback, actionable negative feedback or non-actionable negative feedback.
What we did
- Built a series of simulation-based keyword algorithms in a statistical software package.
- Tested accuracy against a manually classified set of posts.
The result
The client could accurately identify both direct and indirect mentions and reach customers with concerns more effectively than before.
This helped increase customer loyalty and raised CSAT scores.
How we would build it today
- LLM-based classification with a labelled evaluation set, so accuracy is measured, not assumed.
- Topic clustering that shows emerging issues week by week.
- Actionable posts routed to the care team with a drafted reply for a human to approve.
This describes our current toolkit for the same problem. It is not part of the original engagement.