Skip to content
Musab Edriss
← All case studies
AICVM

Conversational and agent intelligence: AI on customer conversations

On-premise AI that reads customer chats for sentiment and buying intent, explains bills in plain language, and automates callbacks and ticket handling — keeping customer data in-house.

Role
Function lead — Data Science, AI & Analytics
Where
stc Bahrain
On-premise — data stays in-houseModelsSentiment, intent, bill driversSignalsSentiment, Bill Explainer, callbacksLive
On-premise LLMsNLPPythonAirflowStreamlit

Some specifics are anonymised to respect commercially sensitive information.

Context

Customer conversations — chats, calls, callback requests, tickets — carry the clearest view of what customers feel and what they want next. Customer care also carries the cost: handling time, disputes, and work queued by hand.

Problem

Sentiment was not tracked systematically across channels. Conversations carried buying signals that were never captured. Customers and agents struggled to see what was driving a bill amount, and missed calls and callback requests were queued and prioritised manually. Sending conversations to an external AI service was not an option for customer data.

What I did

I built this track on our local-model layer: on-premise models that keep sensitive customer data in-house while cutting inference cost. My team delivered five use cases on it:

  • ROVA AI Sentiment — a sentiment analytics platform that surfaces trends and hotspots across chat conversations.
  • Chat Intent Sales — AI models that turn consumer chats into subscriber-level targeting signals across mobile, devices, broadband, roaming, gaming and rewards. Each captured intent is designed to land somewhere useful: analytics and trend visibility, richer personalised recommendations, precise segmentation for up-selling, or an automated next-best action.
  • Bill Explainer — an AI assistant that breaks down a customer’s bill and explains the charges in plain language.
  • Callback Automation and Ticket Distribution — automated workflows that queue, prioritise and trigger callbacks, and distribute care tickets consistently.

Result

Sentiment analytics and the Bill Explainer are live and rolling out to users; callback automation and ticket distribution are live and being extended; chat intent capture is in build. The live tools give earlier detection of dissatisfaction and sharper agent coaching, fewer bill disputes and shorter interactions, and faster, consistently prioritised responses to missed contacts. Chat intent adds a new stream of actionable customer signals for targeting and personalisation.

What I learned

Keeping the models on-premise was a design decision, not a constraint to work around — it is what made it acceptable to point AI at customer conversations at all. And an AI signal is only worth capturing if it lands somewhere: every output here is designed to feed analytics, a model, a campaign or an automated next step.