Agentic AI gets closer to proper autonomy, but its success depends on the combination of large-scale centralized intelligence and highly-tuned small-scale models in distributed edge environments.
Dynamic reasoning – current agentic AI models lack proper dynamic reasoning and rely on hard-coded logic and domain-specific scripts.
Hybrid architecture – future telecom AI will blend large foundational models and small domain models, making use of cloud-to-edge processing.
Structural stuff – organizational and cultural challenges must be addressed to scale AI effectively across global telecom operations.
Note: This article is continued from a previous entry, available here, and is taken from a longer editorial report, which is free to download – and available here, or by clicking on the image at the bottom. An attendant webinar on the same topic is available to watch on-demand here.
But this shift (see report / last entry) is still in its early stages, even if proposed agentic models, strung out across core-to-edge infrastructure, put the industry much closer to some kind of pervasive AI in telco networks – on paper, anyway. Robert Curran at Appledore Research remarks: “We are seeing lots of progress – from driving automation to driving autonomy, attached to this agentic idea. There is still a role for humans but it is an observatory role, not a doing role. But we are very, very, early with agentic AI.”
Meanwhile, there are other issues to contend with. The AI industry has a problem with ‘reasoning’, for example, as set out by ‘chain-of-thought’ models, which generate intermediate reasoning steps to apply some kind of multi-hop logic, akin to human big-picture planning, in order to solve complex ‘real-world’ puzzles and riddles. Except, as it is, this new neural dynamism in large language models is mostly man-handled through these thought processes.
In other words, current AI workflows don’t truly ‘understand’ context; they follow a script. Fatih Nar at Red Hat says: “It will evolve to the point that it says: ‘Okay, this a telco problem and a 3GPP service architecture, and these are logs from Ericsson and radios from Mavenir, and so I should refer to this document to decode the log, and correlate with the right radio performance data – and maybe even link to revenue impact. That is dynamic thinking, and it is coming our way.”
It is coming our way fast; it’s just not coming yet. There are more turns before we top the hill to reveal some sort of agentic horizon. “Right now, we have to hard-code this chain-of-thought,” says Nar. “The reasoning is still pretty static.” In the meantime, the shift from pattern matching to dynamic reasoning also requires domain models and distributed architectures, and one precipitates the other.
The move to tune smaller models for task-specific enterprise workloads, as referenced in discussion of generative AI (see report / previous posts), will accelerate with the distribution of agents into enterprise functions, and development of localised chain-of-thought reasoning. Nar has a funny example about how a telco (Verizon, his former employer, in this instance) might be required to train an AI agent differently at operations centres in Texas and New Jersey.
“It’s cultural, right? It’s Latin and Italian. And culture reflects in the work of data professionals, too – their way of working, dealing with data, doing things. And so it has to reflect in the chain-of-thought in these venues. So in Salt Lake, your AI agent wears a cowboy hat; in Jersey, it is basically a New Yorker, and a Yankees fan. What I’m saying is that even though your AI model is trained out-of-the-box, it undergoes a post-training process in every industry.”
The post-training puts a cowboy hat or a baseball cap on the AI model, then. Nar has a more grounded example: “Take American Airlines. Every minute a plane idles at the gate, it pays out to ground services. The goal is to keep its planes in the air – that’s the business model. And optimizing that whole workflow – shortest route, fullest cabin, least fuel, best timing, highest fares – is something an AI model can do far better than a human.”
Every domain is different, and each demands different AI training tactics – is the point. This goes for domain functions, as well – meaning different operational silos and even different regional businesses within big corporations. The challenge to deploy and train AI is organisational, as much as it is cultural. Petri Hautakangas at TUPL says: “There are issues to scale [a solution] into different markets, all with different processes [to triage and solve technical issues].”
He is speaking from experience, supplying his firm’s network engineering and energy savings tools into different operating companies. He says: “You need a centre of excellence within the customer so there is an internal mandate to put the best practices everywhere. Otherwise the models will be slightly different everywhere. But these are interesting challenges, and really everyone understands the right direction; it is just how quickly you get there.”
We will return to the point about organisational structures to support AI. But the message is that the future is not giant models, but hybrid ones, which combine large systems for language processing with smaller ones – trainable, stackable, controllable, efficient – for complex tasks. This hybrid evolution is being shaped by dedicated tools from the big beasts of AI, geared for smaller industrial models. Small models will be integral to agentic AI.
Volker Tegtmeyer, product marketing principal and manager at Red Hat, says: “Nobody thinks one large language model will do everything – although some still think it will do most things. But with agentic AI, we will have lots of smaller language models that are very good at specific tasks – whether license plate or fever recognition, or whatever.”
All of which informs his company’s rationale about an open AI architecture, where customers, in whatever sector, can select the best models for their apps.
To be continued…
