Telecom AI now has an ROI. But can we trust the numbers?

Telecom operators are finally putting hard numbers against their AI investments, with some targeting billions of euros in additional revenues and savings. But calculating what AI is actually worth is far from straightforward. Is AI really delivering the returns being promised, and how should operators measure its true value?

Telecoms operators have counted AI use cases, users and productivity improvements. But as experimentation gives way to widespread deployment, telcos are beginning to ask: what is all this spend actually worth?

Globe Telecom recently put a number on its efforts, reporting ₱1.2 billion in measurable benefits (approximately £14.5 million), comprising ₱825 million in incremental revenue and ₱368 million in cost avoidance, for a 2.8x return on its AI investment, according to its own calculations.

Meanwhile, Deutsche Telekom (DT) is putting even bigger numbers on the table. At its recent Investor Day, DT said it expects AI-related revenue outside the US to reach around €800 million by 2030, alongside approximately €2.5 billion in indirect cost savings from AI and automation compared with 2023.

Globe and DT aren't the only operators feeling the benefits. McKinsey found that 47% of surveyed telecom executives were already experiencing some impact from AI initiatives by late 2025. However, AI accounted for less than 2% of total revenues at more than half of the surveyed operators that had tested AI use cases.

Returns also take time to materialise. Deloitte found that most organisations reported taking between two and four years to achieve satisfactory ROI from a typical AI use case, with only 6% reporting payback within a year.

This suggests that finding an AI application that works and successfully industrialising it across an organisation are two very different challenges.

Telcos may nevertheless be particularly well placed to generate substantial returns. Operators work with enormous volumes of customer and operational data, complex networks and established digital processes. A few minutes saved on one customer interaction, service activation or network incident may be insignificant; multiplied across hundreds of thousands or millions of customers, the impact can be considerable.

Perhaps Globe’s reported return and DT’s targets are therefore evidence not of uniquely valuable AI applications, but of what happens when relatively small improvements are applied at enormous scale.

Telecom AI is graduating from proofs of concept to proofs of value. But agreeing how that value should be calculated may become just as important as demonstrating it.

Once operators start attaching figures to AI, we need to understand what those figures actually represent.

EY argues that businesses risk significantly underestimating the cost of agentic AI by concentrating on token usage alone. It estimates that the full enterprise cost could be roughly three times the token bill once subscriptions, infrastructure, governance, organisational change, expected failures and regulatory costs are included.

Those costs will also depend on choosing the right model for each task, rather than defaulting to the largest general-purpose model.

Calculating the benefits is equally difficult. Interim TM Forum research among 78 operators found that while 62% could easily connect autonomous network improvements to operational benefits, only 37% could link those benefits to financial outcomes.

Reducing the duration of a network outage is obviously beneficial, for example, but calculating the financial return requires understanding what that outage would otherwise have cost in lost revenue and other consequences.

The same problem applies elsewhere. If an AI assistant saves an employee two hours per week, does that productivity gain translate into two hours of economic value? If customer service AI handles more conversations at lower cost but customer satisfaction suffers, has it created value?

And is avoided expenditure equivalent to money actually removed from the cost base?

These aren't arguments against AI investment, but they demonstrate why an operational improvement doesn't automatically translate into a financial return.

There is not yet universal agreement that AI should even be judged by conventional ROI measures. KPMG found that 65% of UK business leaders surveyed would continue investing in AI regardless of their ability to measure a tangible return, reflecting the growing view that AI is a long-term strategic capability rather than an investment requiring immediate payback.

There is some logic to this. Businesses cannot always calculate the immediate financial value of capabilities they may need to remain competitive.

But strategic importance is not a substitute for investment discipline. If AI is considered essential regardless of measurable returns, almost any expenditure can be justified on the grounds of avoiding future competitive disadvantage. Difficulty measuring value is not evidence that value exists.

There's also an opportunity cost. Every pound spent on AI is a pound that cannot be spent elsewhere. For a telecom operator, that might mean less investment in network coverage, capacity, customer experience or existing BSS/OSS applications.

BCG's 2026 AI Radar illustrates the gulf between expectations and actual returns. While 82% of CEOs said they are more optimistic about AI's potential ROI than a year earlier, just 6% of companies are seeing meaningful value through increased revenues and reduced costs.

Companies generating the greatest returns tend to concentrate investment on a small number of high-impact opportunities and redesign entire workflows around AI, rather than simply automating individual tasks.

Time saved only creates financial value if that capacity is put to productive use, increases output or translates into genuinely lower costs. Employees can become more productive without their employers necessarily capturing a corresponding financial return.

None of this diminishes the significance of the numbers coming from Globe and Deutsche Telekom. Rather, they indicate that AI is maturing from something operators demonstrate into something they increasingly expect to justify financially.

But the number of employees given access to AI or the number of proofs of concept completed tells us little about whether the technology has improved the underlying business.

Revenue, costs, customer satisfaction, network availability and time-to-market may be harder to attribute to a particular AI initiative but, ultimately, are the measures that matter most.

For operators, the next stage will be to identify where AI genuinely changes business outcomes and scale those changes across the organisation. Measuring the results credibly will require a clear baseline and a full accounting of implementation costs. It will also mean distinguishing forecasts and avoided expenditure from revenue or savings actually realised.

The winners of the AI-native telco era may ultimately be distinguished not by how much AI they have, but by how convincingly they can answer one question: what is it actually worth?

Find out how Cerillion’s AI-powered BSS/OSS suite embeds AI into real telecom operations – accelerating time-to-market, automating workflows and improving customer experience.