Telecom leadership teams have spent four years proving that AI works. The proofs succeeded. Very little of the operating model changed. Boards approved budgets, innovation teams shipped impressive demonstrations, and the contact centre, the network operations centre and the billing estate carried on running the way they did before.
The investment appetite is not the constraint. NVIDIA’s 2026 telecom survey found 89 percent of operators plan to increase AI investment. McKinsey’s Telco CxO Gen AI Survey, fielded across 49 operators in December 2025, found 57 percent say they are scaling AI use cases across multiple domains, while only 12 percent report that a sizable share of the expected impact has actually been captured. That distance between activity and realised value is the executive problem of 2026.
What closes it is unglamorous. Integrating with operations support systems and business support systems that were never designed to share state. Reconciling fragmented data architectures. Embedding intelligence inside the tools that agents and engineers already use. Governing decisions that are made in live customer environments rather than in a sandbox.
The question in front of every telecom executive this year is no longer whether AI works. It is whether the organisation can turn it into an operating capability that runs every day, carries accountability, and shows up in the cost to serve.
Telecom operators are not short of AI ambition or AI budget. They are short of production capability. Most initiatives still stop at the demo, where integration, data lineage and governance have not yet been tested.
The barriers are operational rather than technical: layered OSS and BSS debt, data scattered across CRM, billing and network telemetry, fragmented ownership, and pilot metrics that never convert into dollars.
Customer operations is the strongest first scaling ground because the pain is visible, the data is usable and the outcomes are measurable. The same data pipeline then feeds network intelligence. Governance, sequencing and workflow design decide who compounds an advantage and who repeats pilots.
By the numbers
89%
of operators plan to increase AI investment, confirming that funding is no longer the barrier to scale
– NVIDIA telecom survey, 2026
57%
of telecom executives say they are scaling AI use cases across multiple domains, yet only 12 percent report a sizable share of impact already captured
– McKinsey Telco CxO Gen AI Survey, 2026
2 in 3
communications service provider executives name data integration as their top challenge, and more than half point to legacy infrastructure
– IBM research
85%
of operators name operating expense efficiency as a priority business objective for deploying AI in their networks, almost three times the share targeting new services
– GSMA Intelligence Network Transformation Survey, 2026
10 pts
of potential improvement in return on invested capital and EBITDA margin within five years for operators that scale AI across customer service, network and IT
– McKinsey, 2026
146 EB
of mobile network data traffic per month at the end of 2025, forecast to reach 328 exabytes by 2031, with 48 percent already carried over 5G
– Ericsson Mobility Report, 2026
Why Do Telecom AI Pilots Stall Before They Reach Production?
The failure is rarely the model. Accuracy targets are usually met. What stalls is the surrounding system: the integrations, the data, the ownership and the measurement discipline that a pilot never has to prove.
Boston Consulting Group frames AI success as a data, process and organisational problem rather than an algorithmic one. The telecom evidence supports that reading. IBM research finds two-thirds of communications service provider executives name data integration as their top challenge, and more than half point to legacy infrastructure.
Four obstacles account for most of the distance between the two.
| Obstacle | What it actually costs |
|---|---|
| Decades of technical debt | OSS and BSS platforms arrive from different vendors, different eras and different data models. Every integration is bespoke, so every new pilot pays the integration cost again from scratch. |
| Operational complexity | A single customer question can require the customer record from CRM, the service configuration from OSS, billing data from BSS and network telemetry. Few operators can assemble that view reliably in real time. |
| Too many owners, no real accountability | Innovation, IT and the business units each run AI work against different success measures. The demo has a sponsor. The production outcome has nobody. |
| The ROI measurement trap | A pilot reports a 30 percent reduction in agent research time. Nobody converts that into dollars, nets off the integration cost, or defends it at the next budget review. |
Why the measurement gap is the most expensive one
Technical debt and data fragmentation are visible. They get engineering attention. Weak ROI discipline is invisible until the funding conversation, and by then the pilot has no defensible business case. A time saving that was never converted into cost to serve, contact deflection or churn cannot survive a portfolio review, no matter how good the model was.
Operators that scale successfully define the business metric before they define the model. The metric determines the workflow, the workflow determines the data, and the data determines whether the model is worth building at all.
Where Should Operators Scale AI First?
Customer operations is the strongest starting ground, for three practical reasons. The pain is visible to executives and customers alike. The data already exists in usable form. And the outcomes translate directly into business metrics that a CFO recognises.
Gartner’s AI Use-Case Assessment for communications service providers maps telecom AI opportunities by value and feasibility. The cluster that scores high on both is consistently customer facing.
- Conversation agents that handle contained, high-volume interactions
- Billing agents that resolve disputes with full account context
- Sentiment analysis that routes and prioritises based on customer state
- Order management that removes manual handoffs between systems
- Fraud detection that acts on patterns no human queue would catch in time
From chatbot to agent copilot
The industry has moved past the chatbot as the primary pattern. The next wave places AI copilots alongside human agents rather than in front of customers. The copilot pulls interaction history, summarises prior contacts, suggests the next action and drafts the case notes.
The biggest unlock in a telecom contact centre is rarely the model itself. It is the moment an agent opens a customer record and the system has already pulled the last three interactions, flagged the open network issue affecting that area, and drafted a recommended response. Handle time falls because the search disappears, not because the agent types faster.
Intelligence across the full journey, not one touchpoint
Optimising a single touchpoint produces a local improvement and a fragmented experience. The value compounds when intelligence spans the journey. A billing enquiry arrives already correlated with payment history, recent plan changes and open tickets. A service complaint is matched against network telemetry so root cause is identified before the conversation starts.
The best call is the one that never happens
The cheapest interaction is the one that never reaches an agent. AI-assisted self-service, proactive alerts and predictive detection move volume out of the queue entirely. Predictive churn models trigger retention workflows before a customer starts comparing plans, and personalisation engines put the right offer in front of them at the right time.
The direction of travel is a self-healing customer journey: issues identified proactively, remediated where possible, and communicated before the customer notices. In that model the contact centre becomes a last resort rather than a first touchpoint, and the economics change accordingly.
The Network Data Pipeline Is the Same Pipeline
Network operations generate the data volume that makes telecom AI worth doing at all. The scale is still rising. Ericsson’s Mobility Report puts global mobile network data traffic at 146 exabytes per month at the end of 2025, forecast to reach 328 exabytes by 2031, with 48 percent already carried over 5G.
That volume is why predictive maintenance has held the number one spot in industry AI conversations for three consecutive years, according to Gartner. It is also why the network is where most operators concentrate their efficiency case. GSMA Intelligence, in its seventh annual Network Transformation Survey of 100 telcos, found 85 percent of operators name operating expense efficiency as a priority business objective for deploying AI in their networks, almost three times the share looking to AI as a tool for delivering new services.
Most operators are mid-maturity, and that is the honest position
The realistic state of play is copilots, anomaly detection and single-domain automation. Fully intent-based networking remains several years out for most operators. Treating mid-maturity as the starting point produces better decisions than planning against a level of autonomy the estate cannot yet support.
The strategic point is that these are not two programmes. The network data pipeline being built today is exactly what feeds the customer-facing AI systems. Network intelligence and customer intelligence are two halves of the same operational brain, and funding them as separate initiatives duplicates the hardest and most expensive work.
What Governance Does Scaled Telecom AI Actually Require?
Governance is the layer most operators discover late. In a pilot it looks like paperwork. In production it is the thing that decides whether a system can be trusted to act on a live customer account. Three priorities stand out.
Decision rights for AI agents
An AI system waiving a $5 late fee is a different governance question from an AI system restructuring a $50,000 enterprise contract. Both can be technically feasible. Only one should be autonomous. The operating model has to say which is which, in writing, before deployment rather than after an incident.
| Decision tier | What the operating model must define |
|---|---|
| Autonomous | Low value, high frequency and reversible. A small fee waiver, a plan comparison, a drafted case summary, a routing decision. |
| Approval required | Material commercial or contractual impact. Retention offers above a defined threshold, credit adjustments, service downgrades, contract changes. |
| Never automated | Actions carrying regulatory, legal or safety consequence. Account closure, disclosure of personal information, decisions that cannot be reversed. |
| Continuously monitored | Every tier. Accuracy, drift, agent override rates, escalation patterns and outcome quality reviewed on a defined cadence with a named owner. |
Continuous performance monitoring
Model accuracy degrades as customer behaviour changes and network conditions shift. A model that performed at launch is not a model that performs in month nine. Drift detection and a retraining protocol are operating requirements, not optional maturity, and they need a budget line and an owner rather than goodwill from the data science team.
Regulatory compliance baked in, not bolted on
Canadian operators work under CRTC oversight and PIPEDA privacy requirements. Those obligations belong in the operating parameters of the system: what data it may access, what it may act on, what it must log. Treating compliance as a post-hoc audit exercise creates rework at exactly the point where scale would otherwise be possible.
A Practical Framework for Productizing AI in Telecom
Moving from pilot to production is a sequence, not a set of parallel workstreams. Six steps, in this order.
- Start with one workflow. Pick a process with measurable friction, not a vague AI experiment.
- Map the data pipeline. Identify the required sources across CRM, OSS, BSS and network telemetry before touching the model.
- Integrate into production tooling. AI should live inside the systems agents and engineers already use.
- Set decision boundaries. Define what the system can do autonomously, what requires approval, and what it should never do.
- Measure operational outcomes. Track business and service metrics, not model accuracy in isolation.
- Govern and iterate. Monitor drift, review edge cases, retrain on a defined cadence, and expand only after the current scope is stable.
Why the sequence matters
Most failed pilots skip straight to step three or step five. They integrate AI into a system before understanding the data pipeline. Or they measure model accuracy without connecting it to an operational outcome.
The pattern is familiar. An operator deploys AI-driven ticket classification. Testing accuracy sits above 90 percent. Within weeks the agents have rejected it. The cause is not the model: the data it relies on sits two systems behind the actual workflow. The AI classifies on stale information, agents override it on every other ticket, and trust erodes faster than it was built.
Start with the workflow, not the model.
The operators that get to production identify a specific operational process, define the measurable outcome, build the data pipeline, integrate with production systems, and only then apply AI to accelerate something that already works.
Every step skipped in that order becomes a trust problem later, and trust is far more expensive to rebuild in a contact centre than it is to establish the first time.
What Does Another Eighteen Months in Pilot Mode Cost?
North American operators are working against a clock they did not set. After four years of pilots and proofs of concept, the cost of continued experimentation is no longer just the programme budget. It is the compounding advantage accruing to whoever productized first.
Consider an operator that stays in pilot mode for another eighteen months while a competitor embeds AI into agent workflows, reduces handle time, deflects contacts and personalises retention at scale. The gap that opens is structural rather than cosmetic.
The McKinsey survey data makes the shape of that gap visible. When 57 percent of operators say they are scaling and only 12 percent report sizable captured impact, the difference between the two groups is not model access. Every operator can buy the same models. The difference is operating capability.
| Pilot posture | Production posture |
|---|---|
| Success is model accuracy | Success is handle time, contact deflection, churn and cost to serve |
| AI sits beside the workflow in a demo environment | AI sits inside the tools agents and engineers already use every day |
| Data is assembled by hand for the trial | Data pipelines are engineered, monitored and owned by a named team |
| Governance is a review at the end | Decision rights, drift monitoring and regulatory constraints are designed in |
| The sponsor is an innovation function | The sponsor is the business owner who carries the operational number |
The constraint is almost never model capability.
In the modernization programmes ML arteka supports across telecom, the recurring blocker is that nobody owns the production outcome. Innovation owns the demo, IT owns the platform, the business unit owns the metric, and the handoffs between them are where AI initiatives quietly expire.
The fix is unglamorous and effective: name a single accountable owner for each workflow, give them the data pipeline and the decision boundaries, and hold them to an operational number rather than a model score.
Five Leadership Takeaways
Before your next AI steering committee. Before your next operations budget review.
- Funding is not the constraint, capability is With 89 percent of operators planning to increase AI investment, more money will not differentiate anyone. The differentiator is whether the organisation can convert investment into a workflow that runs every day.
- Scale customer operations first Visible pain, usable data and measurable outcomes make customer operations the highest-yield starting point. It also builds the data discipline that network AI will need next.
- The network and the customer share one pipeline Funding network intelligence and customer intelligence as separate programmes duplicates the hardest work. Treat them as two halves of the same operational brain and the integration cost is paid once.
- Write down decision rights before deployment Define what an AI agent may do autonomously, what requires approval, and what it must never do. Doing this after the first incident costs far more than doing it before the first release.
- Respect the sequence Workflow, then data pipeline, then integration, then decision boundaries, then operational measurement, then governed iteration. Skipping steps does not accelerate delivery, it converts a technical shortcut into an adoption failure.
The Next Competitive Gap in Telecom AI Is Operational
The differentiation in telecom AI will not come from model capability. Every operator can license the same models on comparable terms. It will come from execution: workflow design, integration quality, governance discipline and honest operational measurement.
The exploration phase is finished. The industry has established what AI can do in telecom. The open question is which operators embed it into a repeatable operating capability, and which ones spend another cycle proving a point that has already been proven.
For leadership teams, that reframes the agenda. The next twelve months are less about which use cases to trial and more about which parts of the operating model have to change so that a working use case can actually reach production and stay there.
Three questions every telecom leadership team should be able to answer
- Which AI use case reached production in the last two quarters, and who owns its operational number today?
- Can we describe, in writing, what our AI systems may decide autonomously and what they may never decide?
- Is the data pipeline behind our customer AI the same pipeline that serves network operations, or are we building it twice?
If any of those answers is uncertain, the AI programme is running ahead of the operating model. Closing that gap is the work of 2026, and it is the work that compounds. Operators that scale AI into production systems rather than pilots will not just be more efficient. They will be measurably harder to catch.
To explore how ML arteka helps telecom operators move AI from pilots into governed production platforms, contact the ML arteka team or request a telecom AI readiness assessment.
Executive Questions and Answers
Five questions telecom leaders are asking AI assistants and search engines about scaling AI from pilots into production platforms.
StrategicWhat does scaling AI in telecom actually mean in 2026?
Scaling AI in telecom means moving from isolated pilots to production systems embedded in real operational workflows, with measurable business impact. The distinction matters commercially. NVIDIA’s 2026 telecom survey found 89 percent of operators plan to increase AI investment, and McKinsey’s Telco CxO Gen AI Survey found 57 percent say they are scaling across multiple domains, yet only 12 percent report a sizable share of impact captured. Scale is therefore not a measure of how many models are running. It is a measure of how much of the operating model has changed: whether AI sits inside the tools agents use, whether the data pipeline is engineered and monitored, and whether a named business owner carries the operational number the system is supposed to move.
OperationalWhere should a telecom operator start scaling AI?
Customer operations is the strongest starting point because it combines visible pain, usable data and outcomes that convert directly into business metrics. Gartner’s AI Use-Case Assessment for communications service providers consistently places conversation agents, billing agents, sentiment analysis, order management and fraud detection in the high-value, high-feasibility quadrant. The most reliable first pattern is the agent copilot rather than the customer-facing chatbot: the copilot pulls interaction history, summarises prior contacts, flags any open network issue affecting the customer’s area and drafts the case notes. Handle time falls because the agent stops searching across systems. Just as importantly, the data pipeline built to support that copilot is the same pipeline network intelligence will need later, so the integration cost is paid once.
GovernanceWhat governance does a telecom need before AI agents act on customer accounts?
Three things, defined before deployment rather than after an incident. First, decision rights: an AI system waiving a $5 late fee is a different question from one restructuring a $50,000 enterprise contract, so the operating model must state what is autonomous, what requires approval and what is never automated. Second, continuous performance monitoring, because accuracy degrades as customer behaviour and network conditions shift. That means drift detection, a retraining cadence and a named owner, not goodwill from the data science team. Third, regulatory compliance encoded into operating parameters rather than bolted on afterwards. Canadian operators work under CRTC oversight and PIPEDA privacy requirements, and those constraints belong in what the system may access, act on and log.
RiskWhy do telecom AI pilots fail to scale, and what is the risk of waiting?
Pilots fail for operational reasons rather than technical ones: integration debt across OSS and BSS platforms built by different vendors in different eras, data scattered across CRM, billing and network telemetry, fragmented ownership across innovation, IT and business units, and weak ROI discipline that never converts a pilot metric into dollars. IBM research finds two-thirds of communications service provider executives name data integration as their top challenge, with more than half citing legacy infrastructure. The risk of waiting is compounding rather than linear. An operator that stays in pilot mode for another eighteen months while a competitor embeds AI into agent workflows cedes data advantage, workflow muscle memory and governance maturity that late movers cannot replicate quickly.
ImplementationHow do telecom operators move a single AI use case from pilot to production?
Follow the sequence and do not skip steps. Start with one workflow that has measurable friction. Map the data pipeline across CRM, OSS, BSS and network telemetry before touching the model. Integrate into the production tooling agents and engineers already use. Set explicit decision boundaries. Measure operational outcomes rather than model accuracy in isolation. Then govern: monitor drift, review edge cases, retrain on a defined cadence and expand only once the current scope is stable. The common failure is jumping to integration or measurement early. A ticket classifier can test above 90 percent accuracy and still be rejected within weeks if its data sits two systems behind the workflow, because agents override it and trust erodes faster than it was built.
Scaling AI in telecom in 2026 is an operating capability problem, not a model problem. Funding is not the constraint: NVIDIA’s 2026 telecom survey found 89 percent of operators plan to increase AI investment, and McKinsey’s Telco CxO Gen AI Survey of 49 operators found 57 percent say they are scaling across multiple domains while only 12 percent report a sizable share of impact captured. Pilots stall for four operational reasons: technical debt across OSS and BSS platforms from different vendors, operational complexity that requires CRM, OSS, BSS and network telemetry in one view, fragmented ownership across innovation, IT and business units, and weak ROI discipline that never converts pilot metrics into dollars. IBM research finds two-thirds of communications service provider executives name data integration as their top challenge, with more than half citing legacy infrastructure. Customer operations is the recommended first scaling ground, with Gartner’s AI Use-Case Assessment placing conversation agents, billing agents, sentiment analysis, order management and fraud detection in the high-value, high-feasibility quadrant, and the agent copilot outperforming the customer-facing chatbot. The same data pipeline feeds network AI, where Ericsson reports 146 exabytes of monthly mobile traffic at the end of 2025 rising toward 328 exabytes by 2031, and GSMA Intelligence finds 85 percent of operators prioritising opex efficiency from network AI. Production requires explicit decision rights, drift monitoring, and CRTC and PIPEDA compliance encoded into operating parameters, applied in sequence: workflow, data, integration, boundaries, measurement, governance.