Telecom AI Spending Risks Bloating Opex Without Productivity Gains

Telecommunications operators investing heavily in artificial intelligence face a critical operational challenge: scaling AI without simply layering token costs onto already-bloated legacy operating expenses. Without deliberate redesign, the industry risks a cost-creep scenario where AI spending increases total opex without delivering proportional improvements in productivity, customer experience, or revenue growth.

Bain & Company projects that within three to five years, telecom operating cost structures could shift to an "agentic operating model" where traditional expenses account for 70-80% of total opex, with AI agent and token costs comprising the remaining 20-30%. In this model, humans make critical decisions while specialized AI agents execute or augment much of the work across customer care, network operations, software engineering, and enterprise functions.

Why it matters

Unlike digital-native companies, telecom operators must modernize while managing complex legacy business support systems, large outsourced workforces, long-term vendor contracts, and regulatory constraints. The industry's unique operational complexity means AI transformation will be both more valuable and more difficult than in other sectors. Operators that fail to redesign their operating models risk becoming less competitive even as they increase technology spending.

Three predictable cost traps

Bain identifies three common pitfalls that prevent telecom operators from realizing AI's economic potential.

The first trap assumes falling model prices automatically reduce AI costs. While model prices have declined roughly tenfold annually, total token bills often balloon as usage expands unpredictably. Employees discover new applications, power users in network operations and customer care consume tokens at scale, and teams migrate to newer models with more complex reasoning chains that consume more tokens per request. Model inference represents only part of total cost—tool invocations, orchestration, runtime evaluation, observability, storage, and human oversight all contribute to AI workflow economics.

AT&T addressed this by redesigning its AI orchestration so large "super agents" delegate work to smaller, specialized models. By matching model capability to task complexity rather than defaulting to the most powerful option, the company reportedly reduced costs by up to 90% while tripling throughput.

The second trap involves bolting AI onto legacy processes without workflow redesign. This creates a dual operating model where traditional teams continue existing processes while AI performs isolated tasks around the edges. Human-led operations don't evolve, SaaS licenses keep growing, and AI becomes another line item rather than a transformation driver.

Vivo demonstrates the alternative approach with its AI-driven, closed-loop network operations process. Rather than optimizing isolated activities like fault detection, the company redesigned the complete detect-to-resolve workflow, enabling AI to detect anomalies, diagnose causes, execute corrections, validate outcomes, and escalate only when necessary—resulting in significantly faster incident resolution with substantially less manual intervention.

The third trap mistakes low-risk demonstrations for transformation. Internal chatbots, call summaries, and code assistants create value but rarely reshape economics. The bigger opportunity lies in using AI to solve customer problems that were previously too slow, manual, or expensive to address—such as autonomous network incident resolution, proactive churn prevention, or AI-enabled self-service kiosks in rural communities where conventional stores would never be viable, as Telia has piloted.

Five immediate actions

Bain recommends telecom leaders take five steps now: measure cost per task on one workflow rather than total token spending; create a dedicated AI compute budget as a strategic resource; select three high-value workflows for end-to-end redesign; eliminate wasteful AI usage such as agents that loop without converging or duplicate guardrails; and establish governance with named owners, measurable outcomes, and financial accountability for every AI agent before proliferation occurs.

The analysis and recommendations were detailed in a Bain & Company brief published in September 2026 by partners Hannes Schneider, Alex Bhak, Danielle Stekelenburg, Alex Martynov, and Fabio Caiazzo.