AT&T and GSMA warn Telco AI Models Drift

Boost Mobile data scientist Priyank Jain, GSMA director of AI technologies Louis Powell and AT&T data office vice president Mark Austin flagged a specific and seldom discussed risk in telco AI deployments this week: the gap between the data a model was trained on and the data it actually encounters once it is live, a failure mode known as train serve skew. Speaking to Fierce Network on 25 September 2026, Jain explained that the danger is not a crash or a failed job. Nothing alerts an operator that anything is wrong. Jain said the model "just gets quietly worse."

Two Different Ways Models Drift

The article identifies 2 separate causes. General purpose frontier models arrive with no exposure to telecom specific data formats or vendor taxonomies, since none of that appears in their internet scale training data. Telecom specific models face a subtler problem: production data, such as settlement records, arrives late or out of order in ways the training set never captured, and datasets in this domain routinely carry over 100 columns of custom, operator specific parameters. Powell added that frontier models have no inherent understanding of telecommunications and can produce confident, wrong answers, a real risk once a model is embedded in an enterprise workflow. Austin's prescription is procedural rather than technical: match training data to production sources as closely as possible, run a new model silently alongside the live system before cutting over, and keep testing it after deployment rather than treating the launch as the finish line.

Why This Matters Before Autonomy

This is a narrower and more mundane problem than the industry's current preoccupation with agentic AI and autonomous networks, and that is exactly why it matters. The NGMN report on agentic AI concluded that even coordination within a single operator's own domains, let alone accountability between an enterprise and its network provider, requires common information models and audit trails that do not yet exist. Train serve skew shows the problem starts a level below that. A supervised classifier with a known, static task can still fail silently once its inputs drift from what it was trained on. An autonomous agent negotiating across an operational boundary carries the same drift risk plus the coordination and accountability gap NGMN identified. Operators evaluating agentic AI pilots should ask a more basic question first: whether they would know if the model already in production today has quietly stopped working.

Quiet drifts like this example are worse than explicit failures. The quality of the output degrades silently, which degrades the quality of service. Explicit quality control for data quality before and after transformation, and before and after agentic intervention are necessary to enable autonomous networks.

Orange Business Moves Agentic AI Onto Open-Weight Models

Orange Business is moving its most sensitive AI workloads, including its agentic systems, onto open-weight models running on infrastructure it controls, and it says sovereignty rather than cost is the reason. The account comes from a Fierce Network interview with Miguel Alvarez, chief data and AI officer of Orange Business, published on 14 September by Mitch Wagner. Alvarez describes 3 changes in the operator's AI strategy over the past year: a shift of sensitive work from public cloud models to open-weight models it hosts itself, some of them Chinese; a move from chat windows beside applications to agents inside the systems of record; and a 3-month mandate given to a mixed team of AI and operations process experts to automate the front third of the incident management chain. That last piece is the one with a number attached. Qualification, triage and routing of a ticket used to take a technician 30 to 40 minutes working down a checklist of connectivity and configuration tests. Alvarez says an agent now does the same work in about 3 minutes, with 95% of it automated, and that the Level 1 support team "may not survive in its current form".

Sovereignty is the differentiator

The first change is the one Orange Business leads with. Alvarez says "the importance of being able to do AI in a trusted environment has increased a lot", that open-weight models now run work which needed frontier models 4 to 6 months ago, and that the models his teams use most for coding are Alibaba's Qwen and MiniMax, with Moonshot's Kimi under test. The sovereign option is also a product: Orange Business's Live Intelligence platform, built for 100,000 Orange Group employees and sold to 150 enterprise customers, serves Gemini, ChatGPT and Claude alongside Mistral and the open-weight models Orange hosts itself, so a customer can route ordinary work to a hyperscaler model and keep sensitive work inside a French enclave. Fierce reports that the optionality became a practical priority in June, when the US government ordered Anthropic to suspend foreign access to its 2 most advanced models. In July I argued that sovereign AI capacity is the most immediate of the credible operator revenue lines, and that edge and distributed compute become fundable when sovereignty is the demand driver rather than the garnish. Orange Business is the first operator I have seen describe the same logic from the inside, as a buyer of models rather than a seller of GPUs. It did not move to open-weight models because they were cheaper. It moved because a customer, or a regulator, or a foreign government, can now switch a frontier model off.

The second change is the one that matters for the business case. When Bain warned last week that agentic AI could raise telco opex by 30%, the mechanism was that a process augmented rather than replaced removes nothing: the human team keeps running the workflow while agents perform isolated tasks at its edges, and the legacy 70% to 80% of the cost base stays. Alvarez describes exactly that trap in his own first phase. Summarisation, ticket correlation and root cause analysis were the early work, and "you still have the same roles doing more or less the same work in more or less the same way". The 3-month mandate was the correction. It targeted a whole segment of the chain rather than a task inside it, it was staffed by process experts as well as AI experts, and it ends with a change to the organisation: Level 1 teams that supervise agents and handle customer communication instead of running the checklist. Similar exercises are running at Orange Group level across HR, finance, legal, software development and B2B sales, aimed, in Alvarez's words, at whole job lines rather than individual use cases. That is the difference between a demonstration and a P&L entry, and it is the first operator account that describes the retirement of the legacy process as the goal rather than as a consequence to be managed later.

The 2 changes are connected. The reason sensitive work moved onto controlled infrastructure is that agentic systems, "whose autonomy introduces risks requiring greater scrutiny", are now among the workloads. Fierce lists the season's incidents: an autonomous agent system breaking into part of Hugging Face's production infrastructure in July, a swarm of agents using a dormant wiki as a coordination board, a model reaching third-party systems during a security evaluation. Alvarez's response is operational rather than philosophical. Trusting vendors to watch on the operator's behalf is no longer sufficient; vulnerability assessments are more frequent; and the operator needs the ability to quarantine sensitive services and cut their connectivity. That is the containment layer I described when I argued that the agent runtime is not the agent model, and it is the enforcement half of the governance question the industry has mostly discussed as policy. An operator that runs the model on its own hardware can quarantine it. An operator that calls a frontier model through an API can only stop calling. Orange Business's stack, on Alvarez's description, is built on LangChain and LangGraph, OpenTelemetry for observability, Model Context Protocol for interconnection and open-source components for the LLM gateway and MCP registry. Every one of those is a runtime component. None of them is a governance model, and the interview does not describe one.

The main challenge for AI remains trust. Delivering trustable operation, results in a controlled, auditable, traceable manner is paramount. As a result, operating on models that are under your control, deciding what dataset should remain on premise, in jurisdiction or in public cloud and what governance model agentic relies on is a crucial set of decisions for operators.