First-level customer support at NOS runs at industrial volume. Every call holds information about why customers get in touch, what confuses them and what might make them leave, but none of it was written down anywhere.

Every insight locked in audio

Calls were not transcribed, summarised or classified. No structured data came out of any call, so none of it could be analysed at scale.

Teams with no way in

Retention and segmentation teams could not reach call-level information at all. The data that could inform their models existed, but only as recordings.

Labels that overlap in meaning

The call taxonomy has categories that sit close together, such as "I don't understand my invoice" and "I disagree with the amount invoiced". With that much semantic overlap, classification accuracy was hard to even define.

The system is a generative AI pipeline on Google Cloud that processes NOS’s first-level customer support calls. A lightweight hosted model from the Gemini Flash Lite family handles both transcription and summarisation, and a hierarchical classifier assigns each call to a topic.

Choosing the model was a cost decision as much as a quality one. Local models on GPU instances and off-the-shelf cloud services were both explored before settling on a lightweight hosted model, which gave the best balance of cost and performance at this volume.

Some features were removed on purpose. Speaker diarization, separating the agent’s voice from the customer’s, was evaluated and dropped: it added complexity without a meaningful result for this use case.

Classification needed a definition of accuracy before it could be measured. Because the labels overlap in meaning, classification rules were redefined through experiments with human annotators, creating a consistency baseline the model could be judged against.

The pipeline is modular by design. Each stage produces an output that can be consumed on its own through the data warehouse, so retention, segmentation and sentiment analysis teams integrate transcripts, summaries or classifications directly, without running the full system.