TL;DR
- Telco AI splits into network operations (RAN, transport, core), customer experience (care, churn, retention), and revenue assurance / fraud.
- TM Forum Open Digital Architecture and 3GPP SA5 set the standards for AI-driven closed-loop network management.
- GSMA NESAS provides the security baseline for inference deployed alongside network functions.
- Data residency by licence jurisdiction is the dominant deployment constraint — per-country inference and key custody are routine.
- Generative-AI-augmented care has been the highest-ROI 2024-26 deployment; RAN intelligence is the largest by data volume.
Overview#
Telcos sit on some of the largest and most under-exploited datasets in any industry — per-cell per-second KPIs across millions of cells, billions of call detail records, decades of customer-care transcripts. The combination of foundation-model time-series, graph models for fraud, and LLMs for care has finally made this data tractable at production scale.
The industry standard frames are TM Forum's Open Digital Architecture (ODA) for service operations, 3GPP SA5 for assurance, and GSMA's security baselines. ETSI's Zero-touch Service Management (ZSM) supplies the closed-loop automation reference. The work is unglamorous — stitching OSS, BSS, charging, and care into one signal — but it is what makes the model outputs operational.
Common workloads#
- Network optimisation and RAN intelligence — self-optimising network for coverage, capacity, mobility, and interference, informed by drive-test and per-cell KPIs.
- 5G slice and SLA assurance — slice lifecycle automation, SLA monitoring, intent-driven slice creation for enterprise verticals on 5G SA.
- Customer-support copilots — care-agent assist and customer-facing chat across mobile, broadband, and IoT plans.
- Churn prediction and retention — joint models over network QoE, care contacts, billing, and lifecycle signals, with automated retention offers.
- Revenue assurance and fraud — IRSF, SIM-box, subscription fraud, bypass detection; LLM enrichment for case files.
- Field-engineer copilots — multilingual SOP search, swap-out procedures, live RF stats on ruggedised tablets.
- Network cybersecurity — anomaly detection for control-plane and signalling attacks; lateral-movement detection.
Regulatory and compliance landscape#
Ofcom is the UK regulator for telecoms and online services; BEREC coordinates EU-level regulation across national regulatory authorities. TRAI (India), MCMC (Malaysia), ACMA (Australia), and FCC (US) play equivalent roles in their jurisdictions. GDPR and ePrivacy treat traffic and location data under sector-specific lawful basis; lawful-intercept boundaries must be respected by design.
On the standards side, TM Forum's ODA and Open APIs define the interaction patterns for intent-based service management. 3GPP TS 28.105 covers AI/ML management for 5G networks. ETSI ZSM (GS ZSM 001 onwards) defines the closed-loop reference architecture. GSMA NESAS (Network Equipment Security Assurance Scheme) provides the security baseline for vendors and operators.
Where AI is shipping today#
Care copilots are the highest-volume production deployment, with several tier-1 operators reporting 20-30% deflection from human queues. Churn prediction is mature; the 2025-26 cohort adds network-experience covariates that detect at-risk customers four to six weeks earlier than billing-and-care-only models.
RAN intelligence (foundation-model time-series over per-cell KPIs) has moved from research into operational pilots at multiple operators; closed-loop optimisation under TM Forum intents remains the next frontier. Revenue-assurance AI is widely deployed and is one of the few areas where ROI is hard-dollar and immediate.
Pitfalls#
- OSS/BSS data fragmentation defeats most AI pilots — inventory, fault, performance, charging, and CRM each live in their own silos with their own data models. AI without a data-stitching investment underperforms classical analytics.
- Generative-AI hallucination in care: a customer-facing agent that invents a tariff or refund policy is a regulatory and reputational hit. Grounding in the tariff library is non-negotiable.
- Cross-operator data pooling is generally illegal under sector-specific data-protection rules — per-operator models are mandatory.
- Closed-loop automation that bypasses human approval on RAN configuration is the fastest way to take a region offline. Operator-in-the-loop for risky changes is the working pattern.
Yobitel stack mapping#
Yobitel runs AI across network operations, customer experience, and revenue assurance for tier-1 telcos and challenger MNOs. Sovereign deployments are routine: per-licence data localisation across IN, EU, UK, MEA, SEA with in-country inference and key custody where required.
- Yobibyte — fine-tuning on tariff library, customer-service transcripts, and 3GPP specs; model registry for assurance.
- Omniscient Compute — capacity routing under residency constraints.
- Agentic RAG over product catalogue, tariff library, network status, and customer record.
- Tokenised inference for per-business-unit cost attribution across consumer, enterprise, and wholesale.