Short answer. You wrap the stack instead of replacing it. The reliable path to production is a governed data and orchestration layer over the existing OSS/BSS, one AI use case scoped to a single task, and the legacy system kept as the record of truth. Ableneo runs this pattern in CEE telecom: at T-Mobile Czech Republic, document validation now runs 14 times faster, from 7 hours to 30 minutes, at over 90% accuracy, in production on the existing stack.
For most telecom operators, the obstacle is integration, not the AI itself. GSMA Intelligence found that around 60% of operators already run at least one generative AI use case in production, yet only about one in five book cost savings above 10%. The value stalls between the model and the twenty-year-old billing, CRM, and provisioning systems it has to touch. Wrapping the stack fixes that: the legacy OSS/BSS stays the record of truth while AI reads from it and writes back to it through a controlled layer.
Three patterns carry most of the production work in telecom today:
The pattern that fails is the one that tries to fix the core and add AI in the same program. Nokia’s operator survey found 98% of operators say they must modernize BSS for new 5G services. Treating that rebuild as a precondition for any AI use case pushes the first measurable result out by years.
Telecom is a regulated environment, and AI on the stack inherits those obligations. Since 2 August 2026, Article 50 of the EU AI Act requires that any AI system interacting directly with people, including customer-service chatbots and voice agents, tell the person clearly that they are talking to an AI at the first interaction. The Digital Omnibus deferred the Annex III high-risk deadlines to 2 December 2027, but it left Article 50 transparency on its original date. An operator with no high-risk AI still carries these obligations the moment it runs a customer-facing agent.
Two more rules bite on a legacy stack. Automated decisions that significantly affect a customer, for example an automated contract approval or credit check, fall under GDPR Article 22 and need a route to human review. Network and information security duties under NIS2 apply to the systems the AI touches. This is why the wrapping layer matters beyond speed: it is where you place the disclosure, the audit trail, and the human review control these rules require, without re-engineering systems that were never built for them.
You reach production faster by wrapping the legacy OSS/BSS in a governed data layer than by replacing it, and the core stays the system of record.
The blocker is rarely the model; it is data access and data quality. A twenty-year-old OSS/BSS holds customer, billing, and network data across systems that were never designed to be queried in real time, often with inconsistent identifiers and no clean history. Delivery timelines depend far more on that integration and historical data quality than on model accuracy. Before any AI reaches production, the operator needs a reliable way to read legacy data and a place to run, monitor, and roll back the model. Skip that foundation and the pilot works in a demo and dies at go-live.
The second blocker is organizational. Spending skews heavily to technology and away from the people who have to trust the output. Regulated firms share the same imbalance: roughly 72% of AI budget goes to technology and 28% to change management, so the model ships and the frontline keeps working the old way. A real production plan names both the integration path and the adoption path, and staffs both.
Put a data and orchestration layer between the AI and the core. Change Data Capture mirrors legacy data into a fast, governed layer the AI can query, while the old system stays authoritative and unchanged. Event-driven integration lets the AI subscribe to what happens in the stack instead of polling brittle interfaces. Where the vendor exposes open APIs, use them; where it does not, the layer adapts around them. This avoids the trap of point-to-point connections that multiply with every new use case. Operators that take this route move use cases to production in months rather than the years a full replacement demands, and every write back to the core passes through a controlled, logged path.
Add AI around it, then let the AI tell you what to modernize. When agents repeatedly stall or escalate on a particular process, that friction shows exactly where the legacy system holds the business back, and that is where deeper modernization earns its cost. This sequences spend against evidence instead of a five-year rip-and-replace plan built on assumptions. The core replacement, when it is genuinely needed, becomes a series of targeted moves with a measured payback, not a precondition that blocks every AI result until it is finished. The 13% annual growth in the OSS/BSS modernization market is real, and it pays back fastest when the AI layer has already shown where the money is.
Set the production bar before the pilot starts, not after. A use case is ready when it clears three things: a measured result on real data against the current process, not a demo; a governed path to read and write legacy data with monitoring and rollback; and the control the regulation requires, such as Article 50 disclosure and a human review step for decisions that affect customers. Ableneo’s delivery record shows the gap this closes. About 4 of 5 projects reach production, because the production conditions are designed in from the first sprint rather than discovered at the end.
Ableneo builds AI into the systems telecom operators already run, not beside them. The work is grounded in production: 22 AI agents shipped to production across the 2025 portfolio, a document-validation deployment live at T-Mobile Czech Republic at 14 times the previous speed and over 90% accuracy, and a competitive tender won against Big Four firms at Slovak Telekom. That record is the difference between advice and a system that carries real traffic. See Ableneo’s production track record in CEE telecom for how the wrapping pattern turns a legacy stack into a platform AI can run on.
Key takeaways
Planning AI in a regulated business? Ableneo takes systems from classification to governed production.