My main takeaway from Network X: AI is moving faster than the telecoms industry’s ability to govern, cost and operationalise it.
A paradox at the heart of telecoms
Networks have become essential infrastructure, yet the industry is being urged to adopt AI before it can consistently explain the cost, value, operational risk or resilience implications.
Telecoms is becoming increasingly complex, arguably unwieldy, while society depends on digital infrastructure more than ever. We have moved from 3G being a novelty, when fixed phones were still the default, to mobile and fixed networks forming part of the critical fabric of society. That makes the current AI enthusiasm both understandable and worth questioning.
AI is a probabilistic technology. That is not a criticism; it is a fundamental characteristic. It is exceptionally capable of identifying patterns, correlating large datasets and democratising access to data analysis. But probability is not the same as certainty, and networks still need to behave deterministically when it matters.
What I heard on cost, value and trust
The appetite to invest in AI is clear. The commercial discipline behind it is, in many cases, less so. A few observations from the event:
- Cost and value are hard to pin down. Across the headline sessions, few executives could clearly quantify either the full cost of AI adoption or the hard financial value it will generate in network operations. One anecdote suggested Uber exhausted its annual AI budget in four months.
- Some operators are setting measurable targets. Deutsche Telekom Group has set an ambition to deliver €2.5 billion in savings by 2030, alongside potential growth upside. That is the kind of measurable outcome the industry needs more of.
- Trust is conditional. When asked directly whether they trusted AI, opinions became notably more cautious, and rightly so.
Three approaches to adoption
The operators on stage were taking noticeably different routes.
- Melinda Szabó, T-Mobile Czech Republic and Slovakia, spoke about prioritising “big bet” opportunities. The important question is whether conviction is being matched by robust business cases.
- Alejandro Plater, A1 Austria, described an internal AI sandbox, giving employees a safe environment to experiment.
- Mariusz Gatzá, Orange Slovakia, described a more controlled model: a multidisciplinary team focused on delivering practical use cases each quarter, while other initiatives were reined in.
What agentic AI means for the network
The network implications may be even more significant than the business questions. Agentic AI is not simply another application category: it will both consume and generate workloads. As Omdia noted, only around 10% of tokens are currently associated with human activity. That challenges the traditional assumption of downlink-heavy traffic and forces a rethink of network architecture, capacity planning and traffic flows.
Chinese operators are already considering architectures organised around classes of data centre. Meanwhile, valuable use cases are emerging:
- Improving customer engagement
- Triaging tickets
- Optimising RAN performance
- Recommending remedial actions
- Correlating operational data to identify root causes faster
Resilience: an unmanaged dependency?
There is a major resilience question here. Benjamin Vigouroux of Orange Business described having to revise a presentation overnight following US restrictions affecting access to an LLM model in Europe. Imagine that model sitting inside a live operator workflow. Is that operational resilience, or an unmanaged dependency waiting to become an incident?
Knowing where probabilistic AI belongs
The most telling observation may be that autonomous-network initiatives often struggle to get beyond proof of concept. The reason, in my view, is not a lack of enthusiasm. It is a lack of clarity about where probabilistic AI is appropriate, and where deterministic technology remains essential. If a network must be configured precisely as prescribed, assured and repeatable automation matters more than AI-generated confidence.
The unanswered question is whether the compute and data required for effective AI can come together quickly enough. If latency becomes a critical constraint, operators may need entirely new distributed architectures, perhaps closer to a CDN model but designed for AI workloads. That could require substantial investment.
The real question
The question is not whether operators should use AI. They should.
The question is which AI use cases create defensible value, what resilience is required, who funds the required infrastructure, and who ultimately pays.
Listen to the Digitel Talk wrap-up
We have also recorded a wrap-up episode on Digitel Talk, covering our reflections from Network X. Please go and listen, and let us know what you took away from the event.
By Ian Ginn