Dear Reader,
Ninety per cent of telecom operators say AI is delivering positive ROI. Eighty-nine per cent plan to increase AI spending this year. Both figures come from Nvidia’s 2026 State of AI in Telecommunications survey of more than a thousand industry professionals — and neither is the number that matters.
The number that matters: roughly half of those respondents said network automation is the top AI use case driving return. Not customer chatbots. Not churn models. Not the forty or fifty use cases on most roadmaps. The network.
The industry’s problem is not a shortage of AI use cases. It is a surplus of them with no logic for sequencing. The conversation I hear most often from telco leadership: “We have fifty use cases identified. We don’t know which one to start with.” That is not a technology problem. It is a problem of strategy, priorities, and data.
Why the paralysis happens#
Telecom vendors sell point solutions. An AIOps vendor sells network automation. A CRM platform sells churn prediction. An NLP vendor sells customer service automation. Each arrives with a business case for its own product. None of them tells you which use case generates the data that makes the next use case work.
The pattern repeats: pilots in silos. A churn model trained on billing data, disconnected from network performance data. A customer service AI that logs sentiment but doesn’t feed it back to the retention engine. Each initiative evaluated on its own ROI, none evaluated on what it unlocks downstream.
The Heavy Reading/Omdia 2025 AIOps survey of 84 global network operators confirmed the structural problem: 52% still operate with siloed data systems. Three years into the AI era, the data integration problem is not solved — it is the problem.
The Prioritisation Matrix#
Before selecting a use case, evaluate it across three dimensions.
Data availability. Does the data exist, is it clean, and is it accessible across systems? A churn model requiring unified network, billing, and customer service data will fail in most telco environments today. An AIOps model running on structured network telemetry has a clear path to production.
Regulatory exposure. AI managing critical network infrastructure may fall under EU AI Act Annex III, Section 2 — safety components in critical digital infrastructure. The August 2026 compliance deadline applies. Churn prediction is not in Annex III, but GDPR Article 22 applies wherever automated scoring affects how customers are treated without a documented human decision in the loop. Both create obligations that most telco AI programmes have not yet mapped.
Revenue impact. Network OPEX reduction from AIOps: 25-40% in documented deployments. Churn reduction at scale: at 15-30% annual churn across most operators, a meaningful improvement in predictive accuracy has direct P&L impact. Customer service automation: measurable in handle time and resolution rates, but the Nvidia data suggests the larger operational returns come from internal process automation — fraud detection, billing anomaly management, technician scheduling — not front-facing bots.
Use these three dimensions to rank, not just list. Most roadmaps contain lists. A portfolio contains a sequence.
| Use Case | Data Availability | Regulatory Risk | Revenue Impact | Sequence |
|---|---|---|---|---|
| AIOps / network ops | High — structured telemetry exists | EU AI Act Annex III, Section 2 | OPEX -25–40% | 1st |
| Internal automation (fraud, billing, scheduling) | High — structured internal data | Low | Fraud loss, handle time | 1st / 2nd |
| Customer service AI | Medium — requires CRM + NLP | GDPR Art. 22 if decisions affect service | Moderate | 3rd |
| Churn prediction | Medium-low — requires unified network + CRM + billing | GDPR Art. 22 (differential treatment) | High P&L lever | 4th |
Start with the network#
Network automation is where the clearest ROI sits — and where the data foundation for everything else is built.
AIOps use cases include predictive maintenance, traffic optimisation, fault detection, and radio access network optimisation for 5G coverage and energy efficiency. These are not exploratory: 47% of operators in the Omdia survey report assurance operations that are already autonomous for specific use cases. AT&T’s Geo Modeler simulates geographic and environmental variables before infrastructure deployment — AI as a capital allocation tool, tested before concrete is poured.
AI managing network infrastructure is not the same as AI recommending a product offer. When the model gets it wrong at 3am, service degrades for hundreds of thousands of subscribers. The industry has already internalised this: 58% of operators use digital twins — parallel simulations of the live network — to validate AI decisions before live deployment. That is the right architecture. Human oversight embedded in the engineering process, not bolted on as compliance theatre.
Bain’s February 2026 report found that fully autonomous networks remain aspirational for most operators. The productive path is targeted automation in service assurance, network planning, and operations support, layered progressively. Each deployment builds the telemetry that powers what comes next.
Customer service: not the starting point#
Customer-facing AI gets the coverage. It also delivers lower returns than most leadership teams expect.
Nvidia’s data is instructive: internal operational improvements — billing reconciliation, fraud pattern analysis, ticket routing, workforce scheduling — are outperforming customer service chatbots on ROI metrics. AT&T’s autonomous agents for fraud reduction and customer wait-time management work because they operate on structured, high-volume internal data. The mechanism matters more than the interface.
The governance question in customer service AI is different from network ops. Customer interactions generate sentiment signals — frustration indicators, service quality complaints, escalation patterns. That data matters — but only if it feeds a churn model that then triggers differential treatment — different offers, different service priority, different retention effort — without a clear human decision documented in the process, Article 22 exposure follows. The human consultant placing the retention call is not automatically the oversight mechanism — not unless there is documented authority and a genuine ability to override the model’s recommendation.
Customer service AI is best deployed second in the sequence: after the network data foundation is built, it becomes a data generation mechanism as much as a cost reduction mechanism.
Churn prediction: the data-hungry endgame#
Churn prediction has the most seductive business case in telecom AI. Annual churn rates of 15-30% across most operators, with prepaid markets higher. AI models demonstrating accuracy of 88-97% in research settings. Targeted retention intervention costs a fraction of acquisition costs.
The operational reality is harder than the pitch. Most churn models run on billing and CRM data. The models that achieve the upper end of accuracy integrate network performance data — subscribers experiencing persistent degradation at their location are measurably more likely to churn than those who are not. That network performance data lives in the AIOps infrastructure. Without the first use case operational, the third one underperforms.
Deloitte’s TMT Predictions 2026 adds a signal most churn models are missing. In developed markets, mobile users may value operator reward schemes as much as — or more than — network performance improvements by end of decade. Deloitte’s framing is blunt: gifts beat gigabits. With no new device categories expected through 2030, loyalty programme engagement data is becoming a primary retention signal — and most churn models do not incorporate it.
Most telco churn models operate as marketing tools: a score is produced, a retention consultant calls, a discount is offered. Whether the automated score constitutes a decision that significantly affects the customer — and whether meaningful human review is documented in the loop — is rarely examined. That is not a compliance exercise. It is the difference between a defensible process and a liability.
The data flywheel#
The case for this sequence — network ops first, internal process second, churn prediction third — is not about the individual ROI of each use case. It is about what each one generates for the next.
AIOps creates structured, reliable network telemetry across millions of events daily. That data, fed into a churn model, provides the signal that billing data alone cannot. Customer service AI creates structured sentiment and interaction data. That data, also fed into the churn model, adds a second predictive layer. Loyalty programme engagement data completes a third layer.
The flywheel only turns if the use cases are connected — if the output of each feeds the next. Every vendor sells their own component. Assembling them into a working system is not their problem.
The Briefing#
Accenture acquires autonomous network AI platform
On 24 February 2026, Accenture acquired Avanseus, a cloud-native AI platform for prediction, anomaly detection, and optimisation in complex network operations. The technology is designed for integration with hyperscaler agentic AI platforms and will serve as a foundation for Accenture’s autonomous network services. This is the vendor consolidation pattern in motion: point solutions are being absorbed into managed services. Telcos buying standalone AIOps platforms today may find those capabilities inside a managed services contract within 24 months.
Network automation, not chatbots, is driving telco AI returns
Nvidia’s 2026 State of AI in Telecommunications (1,000+ professionals): 90% report positive ROI, 89% are increasing AI spend, but roughly half identify network automation — not customer service — as the top returns driver. Internal process automation is outperforming front-facing AI on measurable returns. The use case that gets the budget in a roadmap presentation and the use case that pays back are not always the same.
74% planning AI agents in network ops — 52% still operating in silos
The Heavy Reading/Omdia 2025 AIOps survey of 84 global operators: 74% plan to deploy AI agents across network operations within two years; only 47% have reached operational autonomy for any specific use case today. The primary barrier is not model quality — it is data architecture. Operators planning agent deployment without resolving the silo problem are building on the wrong foundation.
Deloitte 2026: rewards may matter more than signal quality for churn
Deloitte’s TMT Predictions 2026: in developed markets, mobile users may rank operator reward schemes above network performance improvements by end of decade. With no transformative new devices expected through 2030, non-network benefits are becoming a primary retention lever. Churn prediction models that do not incorporate loyalty programme engagement data have incomplete signal — and most do not.
Questions for Your Leadership Team#
How many of your roadmap use cases can reach production with data you already have, clean? The answer is smaller than the list suggests. Start there.
Which of your use cases generates data that improves a different use case? If you cannot draw that dependency graph, you have a list of pilots, not a portfolio.
For your network management AI: does it fall under EU AI Act Annex III, Section 2? Critical infrastructure AI has a compliance deadline in August 2026. Most telco AI programmes have not yet documented a classification.
For your churn model: where is the documented human decision in the process? A churn score that automatically triggers retention action — without a recorded decision point and the ability to override — is an Article 22 exposure. The volume of the list the agent works from is not the answer.
Are your AI initiatives sharing data, or sharing a slide deck? The difference between a portfolio and a list of projects is whether the outputs of one feed the inputs of the next.
The Portfolio Window#
Telecom AI spending is accelerating. The operators that build a data flywheel — network telemetry informing churn prediction, customer sentiment closing the loop — will see returns multiply as each use case feeds the next. The operators that execute independent pilots will continue generating impressive individual dashboards and negligible system-level impact.
The vendor market will not solve this. Every vendor in your roadmap presentation is selling the use case they own. The sequencing question — which use case builds the data foundation that makes the next use case work — is the question nobody in the room has been paid to answer.
Until next issue,
Krzysztof
Sources: Nvidia 2026 State of AI in Telecommunications · PYMNTS Feb 2026 · Heavy Reading/Omdia 2025 AIOps Survey via Radcom · RCR Wireless, December 2025 · Accenture/Avanseus acquisition, Feb 24 2026 · Deloitte TMT Predictions 2026 · EU AI Act Annex III · Bain, Accelerating Autonomous Networks, Feb 2026