Agentic AI at MWC26: A short primer on what to look for

If you haven’t already got the memo, we should all expect to be flooded with discussions and buzz about AI and what it means for the telecom industry at MWC this year. We can expect to see AI infused into nearly every conversation, from the RAN and Edge through the transport layer, to the core and let’s not forget on-device AI. 

The GenAI craze has barely crossed the hype cycle crest and settled into more rational discussions, but it is already at risk of being upstaged by Agentic AI. This is certainly true for the telecom industry as well and the locus of conversations has rapidly shifted from deploying GenAI to improve front-office and customer facing functions to discussions around cognitive core networks with dozens to agents and sub-agents scuttling through the network, “autonomously” executing on high-level business intent. There was plenty of signaling at MWC25 that agentic AI would become a major trend, and we expect this year to build on that momentum.  

Agentic AI and a future of Autonomous Networks

A common refrain for telecom networks is “complexity”. Throw in a spoonful of legacy and a pinch of regulation for seasoning, and you have a potent recipe for a dearth of innovation that has caused telcos to slip, slowly but surely, behind their much faster and rapidly iterating cousins in the tech world. Recent years have seen much discussion about a transition from telco to “techco” but there has been precious little by way of evidence that would move the needle. In fact, most operators have not crossed Level 2 in the oft-cited TM Forum framework on Autonomous Networks, forget about Level 4 which is the baseline for mostly autonomous operations in the network until reaching Level 5.

So how do mobile operators get to this promised Land of agentic AI and autonomous network operations? On top of the expected slate of announcements and massive hype, we also expect to see discussions on the following issues regarding agentic AI at MWC26. 

• The need for telco-specific models – one of the pressing questions for operators evaluating the deployment of large language models (LLMs) as the foundation for GenAI applications has been whether to rely on one or more of the plethora of LLMs available today from the likes of OpenAI and Anthropic, or to build models that are much more narrowly based on telco domains with the requisite knowledge to avoid hallucinations and induce better results from superior reasoning. Multiple companies from Netcracker and Amdocs to Totogi and NetoAI will be present and sharing their visions for this roadmap.
• Building an AI-ready data foundation – this sounds so obvious, but it remains amongst the biggest challenges for operators, who have been saddled with years of legacy equipment and software, compounded by poor short-term choices. The result has been heavily siloed data that is very difficult to extract, let alone scan and analyze for outcomes. Many operators have invested heavily in datalakes but while many data streams have been established, much works remains to make the data easily accessible, readable, and consistent. Operators can’t wish away the legacy but they can abstract away from it to offer the base layer of AI-ready network data. Who are the partners and solutions that will help them build this layer? One of the major “themes” at MWC this year will be Intelligent Infra, with a whole session sponsored by Snowflake worth watching.  
• The right tools for the right objective – do operators need agentic AI to replace everything that they have been doing so far? Many operators already have sophisticated tools and ML models for things like churn management. In some cases, operators increasingly use APIs in a deterministic way to enable third party applications and use cases. Does it make sense to replace these with agentic AI which is a new, non-deterministic “semantic” interaction that is arguably more complex but also a lot more expensive? Or do operators focus on making their APIs more compatible with agents through a new semantic layer? Operators will need to grapple with these questions as they strategize on the value of agentic AI and the way forward.  
• Total cost and affordability – agentic AI holds transformational promise for operators, but this comes with a very hard tradeoff in terms of cost. That is, what is the total cost in terms of tokens and compute to entrust the entire network to “agents” to execute truly autonomous network operations? Most operators, if presented with the current highly “centralized” options for compute, would balk at the costs. However, these costs could be more manageable not only by deploying smaller telco domain-specific models, but also by rearchitecting the network to enable inferencing closer to the edge of the network thereby mitigating egress costs to compute resources in the cloud.
• What happens when things go wrong? For better or worse, operators today have established processes and protocols in place to troubleshoot network issues and keep their networks secure. Even APIs have the requisite degree of observability. But what happens with agentic AI? Operators will need new governance models, as well as tried and tested observability and audit capabilities, to enable them to entrust operations to agentic AI. This is an emerging space but will command increasing attention. 

For all the promise, there is plenty of work to do to realize the potential rewards from a shift to agentic AI for mobile operators. Ultimately, agentic AI will need to be deeply embedded into telco operations and integrated with telco-specific models, which are well versed in telco terminology and taxonomy so that they can bridge both new and legacy systems. Operators would do well to “start small and scale smart” after addressing the challenges discussed above. We are at the end of the beginning when it comes to the AI story for operators, and MWC26 will play host to many of the critical discussions which will propel this transformational technology and its adoption within the telecoms industry.