National telecom operators face a challenge: the data needed to launch AI projects is fragmented, inconsistent, contains gaps, and lacks unified reference data. Attempting to train AI models on such data yields poor results, making it impossible to implement intelligent systems for improving subscriber service efficiency or network optimization.
The cause: architectural chaos and lack of Data Governance
This problem arises from a historically formed OSS/BSS ecosystem, which for large telecom operators can comprise 15–25 systems of various generations. Each system (CRM, billing, network management systems, subscriber support systems) was created to solve its specific narrow task, often without considering the need for a unified customer profile or common reference data. As a result, the same customer may have multiple records with different addresses, contact details, or even names, and service and tariff data are stored in disparate billing systems. This leads to the Customer 360 concept (a single, complete view of the customer) not working, and the time-to-market for new tariff plans being limited by the need for manual data reconciliation between legacy systems.
Solution options: from MDM to Data Mesh
Solving this problem requires a systematic approach to data management. One option is implementing MDM (Master Data Management) with a consolidate model, where the MDM hub (central master data repository) only matches and aggregates data from various sources, while the data itself remains in the source systems. This approach is relatively quick to implement but does not guarantee integrity during updates, as sources can continue to create duplicates. A more reliable but complex option is MDM with a registry model, which creates a centralized master data model with an API to synchronize changes back to the sources. This ensures a higher level of consistency but requires significant effort for integration and process restructuring. For large telecom operators with decentralized teams and high business domain autonomy, the Data Mesh approach is promising, viewing data as a product where each domain is responsible for the quality and availability of its data, eliminating a single point of failure and scaling data management.
A common mistake
Often, companies, especially in the telecom sector, rush into launching AI projects, focusing solely on technologies and the accuracy of AI models. They spend significant resources developing complex algorithms without defining clear business success metrics at the outset. This results in projects that may show high model accuracy but deliver no real business value, as it’s unclear how this accuracy translates into reduced operational costs, increased customer satisfaction, or profit growth.
The correct path is to define the business metric first: this could be the percentage of time saved by contact center operators, the percentage reduction in errors during request processing, or the specific ROI from AI implementation. Model accuracy is an important engineering metric, but it is secondary. Business value must be the priority, and only then do investments in AI become justified.
Using the example of a typical national telecom operator with millions of subscribers, providing both B2C and B2B services, we see how the AI Act (which came into effect in the EU in 2024 and already impacts companies working with European partners) creates new challenges for VoIP and contact centers. Requirements for transparency, explainability, and non-discrimination of AI systems mean that an operator cannot simply implement a voice bot to automate calls without ensuring full transparency of its operation and the ability to audit its decisions. This also applies to LCR-routing (Least Cost Routing) systems, which may use AI to optimize traffic. Without proper data management and clear data governance policies, meeting these requirements becomes impossible.
Technologies for AI in communications
For successful integration of AI into telecom infrastructure and compliance with AI Act requirements, it is important to use proven technological solutions:
- Apache Kafka — a distributed event streaming platform used for asynchronous delivery of data change events between various OSS/BSS systems, such as between CRM and billing (topics customer.created, account.updated). Without Kafka, synchronization would occur through point-to-point calls, which with 15+ systems turn into a web of dependencies. Any change in one system causes a cascading failure in three others, and the integration time for a new system increases from weeks to months.
- API Gateway — a single entry point for all API calls, providing centralized access management, routing, and monitoring. In our scenario, the API Gateway is used to unify access to customer data from various OSS/BSS systems, providing a single interface for AI services. Without it, AI model developers would have to integrate with dozens of different APIs, significantly complicating development, maintenance, and security.
- Kubernetes — an open-source system for automating the deployment, scaling, and management of containerized applications. Kubernetes ensures the scaling of integration microservices and AI models during peak loads, for example, on days of mass mailings. Without it, peak hours would lead to either API timeouts or excessive infrastructure allocation “for reserve,” which is expensive and inefficient.
In practice, the DooxSwitch Platform (a VoIP platform for telecom operators from DooxSwitch), which includes softswitch, multi-tenant PBX, SIP-routing, billing, and WebRTC, is used for managing VoIP traffic and integrating AI models for voice data analysis. This allows, for example, detecting fraud or automatically determining subscriber intent in real-time, ensuring transparency and auditability of AI systems in voice channels in accordance with AI Act requirements.
Risks
Despite the benefits, there are risks that can harm projects involving AI integration in telecom communications. The biggest one is the lack of a clearly defined master data owner at the start of the project. If there is no single responsible party defining data quality standards and update policies, the MDM initiative is doomed to failure, and AI models will continue to operate on poor-quality data. Another risk is underestimating the complexity of integrating legacy systems. Even with the use of Kafka and API Gateway, synchronizing data from systems that are 20+ years old can turn out to be significantly more expensive and time-consuming than expected if a deep audit of their architecture is not conducted.
Consistent implementation of data governance and MDM practices, preceding the launch of AI projects, transforms data from an “unmanaged asset” into a ready resource for AI models and management analytics. This eliminates “garbage-in/garbage-out” surprises and allows telecom operators not only to comply with AI Act requirements but also to derive real business value from intelligent systems. For example, Softengi develops custom AI models for detecting traffic anomalies and potential fraud in VoIP networks, while Softline and Data Management IG teams have experience in implementing comprehensive solutions for building a stable omnichannel architecture.