Over the next three to five years, the operating cost structure used by telecom operators is likely to undergo one of the most significant shifts in decades. As AI agents become embedded across customer care, network operations, software engineering, and enterprise functions, tokens will account for a growing share of operating expenditures.

One increasingly plausible scenario is an agentic operating model in which traditional operating expenses make up approximately 70% to 80% of the total, while the remaining 20% to 30% consists of AI agent and token costs. Humans will continue to make critical decisions, but much of the execution will be performed—or augmented—by specialized AI agents working alongside them.

This transition will be both more valuable and more difficult in the telecom industry than in most other industries. Unlike digital-native companies, telcos must modernize while operating complex legacy operations and business support systems; managing large, outsourced workforces; navigating long-term vendor relationships; and complying with regulatory and labor constraints. AI is also beginning to shift the traditional balance between capex and opex, as software stacks evolve toward common systems of record with intelligent agent layers on top.

Telco chief technology and information officers, chief financial officers, network operations executives, customer service leaders, and other executives face an immediate challenge: How can they scale up AI without simply adding token costs to an already-heavy legacy operating cost base? The risk is a dangerous cost-creep scenario: higher operating expenses without proportional gains in productivity, customer experience, or growth (see Figure 1).

Figure 1

Avoiding that outcome requires redesigning how work is performed while building new mechanisms to govern AI as an operational resource. Many organizations fall into predictable traps.

Cost trap #1: Cheaper models, bigger bills

The first misconception is that falling model prices automatically mean lower AI costs. Although model prices have declined roughly tenfold each year, the effective cost per task often stays flat and the total token bill balloons unpredictably as usage rises (see Figure 2). Employees discover new uses, power users consume tokens at scale (especially in network operations and customer care, given high-volume workflows), tasks become more complex, and teams shift to newer models—whose more complex reasoning chains consume more tokens per request—instead of picking the most cost-efficient model for the task.

Figure 2

Most organizations still manage AI spending like cloud infrastructure costs, tracking total API invoices rather than asking a more important question: What did this business outcome actually cost? For AI, the meaningful unit of economics isn’t cost per token. It’s cost per resolved customer issue, network incident, proposal generated, or software release.

Model inference represents only part of the total cost. Tool invocations, orchestration, runtime evaluation, observability, storage, and human oversight all shape the economics of an AI workflow. Optimizing model choice alone leaves substantial savings unrealized.

The companies managing token economics effectively treat it like any other strategic production resource, rather than as discretionary experimentation. That means creating a dedicated, protected AI compute budget with clear governance and taking a portfolio approach that matches the right model to the right task. For example, AT&T examined how it could more efficiently process billions of tokens per day. The company redesigned its AI orchestration so that large “super agents” delegated work to smaller, specialized models. Rather than reducing AI usage, it matched model capability to task complexity, which by AT&T's account reduced costs by up to 90% while tripling throughput.

Cost trap #2: Bolting AI onto a legacy process landscape

Many organizations deploy AI without redesigning how work actually happens. The result is a dual operating model. Traditional teams continue executing existing processes while AI performs isolated tasks around the edges. Human-led operations don’t evolve, SaaS licenses continue growing, and AI simply becomes another line item on the income statement. Instead of transforming operations, organizations automate fragments of legacy complexity.

The better approach is to redesign workflows from end to end. Rather than asking, “Where can we insert AI?” leaders should ask, “Which processes would benefit the most from AI, and to what degree should each be automated?” This often means eliminating hand-offs, simplifying approvals, reducing manual monitoring, and replacing dashboard-centric work with autonomous agent interactions.

Vivo’s implementation of an AI-driven, closed-loop network operations process demonstrates this shift. Rather than optimizing isolated activities such as fault detection, the company redesigned the complete detect-to-resolve workflow, letting AI detect anomalies, diagnose causes, execute corrective actions, validate outcomes, and escalate only when necessary. This resulted in significantly faster incident resolution with substantially less manual intervention.

Organizational redesign extends beyond workflows. In parallel, it requires thoughtful workforce planning and execution as the required skills and talent mix evolve, as well as renegotiating SaaS contracts as agents handle more tasks for workers.

Cost trap #3: Mistaking demos for transformation

Telcos often focus on low-risk AI demonstrations that only marginally improve efficiency. This sets the ambition too low and misses the P&L upside. Internal chatbots, auto-generated customer call summaries, code assistants, and productivity copilots create value but rarely reshape a telco’s economics.

The biggest opportunity lies in using AI to solve customer problems that were previously too slow, too manual, or too expensive to address. Instead of asking how AI can reduce the cost of existing work, leaders should consider which customer problems become economically viable only because AI exists. That shift changes the conversation completely.

Examples include autonomous network incident resolution, proactive churn prevention, personalized B2B proposal generation, or fully self-service retail experiences in locations that could never support traditional stores. Telia piloted AI-enabled self-service kiosks in rural communities, serving customers where conventional stores would never have been viable.

Leading organizations recognize that AI transformation is a journey, and low-risk individual experimentation is only the first step.

Five actions every telco leader can take right now

Although the transformation will take years, leaders can make these changes immediately:

1. Measure cost per task on one workflow. Instead of focusing on total token spending, assess the true cost of resolving a customer issue, triaging a network incident, or producing a sales proposal today. What would it cost with AI?
2. Create a dedicated AI compute budget within the organization. Treat AI as a strategic operating resource rather than hiding it within existing IT budgets. If business unit managers are forced to scrounge for money mid-quarter, adoption in critical domains will stall.
3. Pick three high-value workflows for end-to-end redesign. Focus first on high-cost, high-friction areas, such as customer care, network operations, or software delivery. Start with real work, not demos.
4. Eliminate wasteful AI usage. Some token spending adds little value. Agents can loop without converging, resend the same context on every call, and run duplicate guardrails over the same output. Cap how long an agent can run before it hands off to a person, strip repeated context out of prompts, consolidate overlapping checks, and retire the uses that aren’t delivering returns. Internal incentives that reward indiscriminate token usage can also contribute to the problem. AI token leaderboards are a blunt instrument that can help increase employee AI adoption but also drive up costs. Consider eliminating or fine-tuning them.
5. Establish governance before proliferation. Implementing agentic AI is different than a typical technology transformation because of its autonomous decision making and actions that have real business consequences. “Set it and forget it” won’t work here; someone must be on the hook to regularly monitor model drift and token spending. Leading organizations will assign every AI agent a named owner, measurable outcomes, and clear financial accountability.

The true destination

The goal isn’t simply a 70-to-30 cost ratio. It’s an operating model that’s faster, simpler, and higher performing because it’s AI-native. Deploying the most AI agents or consuming the most tokens won’t determine success. Those that simply layer AI on top of legacy operating models will see costs rise without fundamentally changing their competitive position. The real divide will be between telcos that deliberately reshape their operating model and those trapped trying to optimize yesterday's processes with tomorrow's technology.