Re-imagining analytics as an AI enablement layer
Repositioned CVM analytics from reporting into an in-house AI enablement layer — models, decisioning, orchestration and local infrastructure — that serves the telco and its adjacent businesses with no licensing cost.
- Role
- Function lead — Data Science, AI & Analytics
- Where
- stc Bahrain
Some specifics are anonymised to respect commercially sensitive information.
Context
Analytics in CVM had grown up as a reporting and campaign-support function. The business needed something different: analytics as a tool to solve key business problems and open new opportunities — across the core telco and a growing set of adjacent businesses in fintech, insurance, loyalty and travel.
Problem
Scaling AI-led work ran into the same walls every time. There were no GPUs to hand, every stakeholder tool meant a long front-end cycle or a vendor, and models that worked in a notebook struggled to become something that ran reliably every week. Buying a platform would have added licensing cost before a single use case had proven its value.
What I did
I reset the function’s approach around four moves:
- Re-imagine the analytics infrastructure — an architecture built for scaling AI initiatives, at zero licensing cost.
- An AI model factory — problem-specific models delivered lean, focused on incremental conversion rather than model count.
- Agent intelligence and automation — putting insight into front-line interactions, with real-time push notifications.
- Analytics-as-a-Service beyond CVM — lead intelligence, tailored ML models and performance analytics for other teams and business units.
Underneath sits an AI enablement layer with four tiers: models and scoring (propensity, risk, fraud), decisioning logic (next-best-offer and next-best-action, combining rules with LLMs), orchestration (SAS, pipelines, workflow automation) and a local infrastructure engine. The principle is a clear division of labour: the technology function provides infrastructure and platforms; CVM converts them into intelligence and value.
The delivery engine is our own. A small cluster of virtual machines runs orchestration and on-premise LLM inference on the primary node, with worker nodes for model execution, dashboards and failover — all open-source, with automatic recovery instead of manual ops. Airflow schedules the ETL and ML workflows end to end: building model-ready tables, creating features, scoring, publishing, and monitoring with retries and alerts — for inbound sales, telesales and loyalty, API uptime, ML scoring and competition monitoring. Streamlit turns outputs into interactive apps shipped straight to stakeholders, all reachable from a single super app entry point.
Result
The function now runs more than 20 models in parallel on its own compute, with 100% uptime and no licensing cost, and ships stakeholder tools without long front-end cycles or vendor dependency. On-premise models keep sensitive customer data in-house while cutting inference cost, and the lack of GPUs stopped being a blocker. The same stack now carries a portfolio of AI and analytics use cases across commercial, customer care, insurtech, fintech and core CVM.
The programme is measured on business outcomes, not models shipped: lower handling time, less manual complaint processing and faster resolution on the cost side; incremental revenue, campaign conversion and ARPU on the growth side; and first-contact resolution, churn-model precision and omnichannel AI decisioning for customer experience. Next, we are aligning with the Big Data team to move compute onto the enterprise data platform — lower GPU and infrastructure cost, and a path to fully real-time processing.
What I learned
Owning a small, well-automated platform beat waiting for a large one: it let the team prove value use case by use case, and it now gives us a clean path onto the enterprise platform rather than competing with it. And being explicit about who owns what — platforms with technology, intelligence with CVM — removed most of the friction that usually slows analytics teams down.