Clarence Reynolds, TelecomTV (00:05):
Artificial intelligence is increasingly becoming a critical transformative force across network ecosystems with telecom at the forefront of innovation. Andrew Cower, general manager of software networking at IBM, joins us to explore the current challenges, opportunities, and strategic approaches to AI adoption in telecom. Andrew, thank you for joining and following the great telco debate. It seems there is a long way to go before AI is seriously impactful to telcos. What are the obstacles to adoption?

Andrew Coward, IBM (00:37):
Well, hi Clarence. So I think AI is being adopted by telcos and approaching from the edges, if you like. So the use cases today tend to be some of the things that AI is actually really good at. So for example, telling you why your bill might be higher this month or some of the basic customer blocking and tackling kind of things that you might want to do. Where I think it's struggling is in the ultimate challenge, if you like, of networking, which is, how do I answer really complex questions? Or rather, the answers are complex, maybe not the questions complex. For example, why is Microsoft teams running slow in Chicago today across my network? And if you think about the intricacies of that question and where you would need to look as a human to go find that answer and how long it might take to get to a resolution, then that's exactly the kind of challenges that AI is facing, and that's why it's going to take a little longer to get there to get those answers and AI really contributing. But the prize of doing that's very substantial, if we can radically reduce the amount downtime that networks see, and also very importantly reduce the time it takes to resolve issues, then that's a massive win for the industry and worth the effort to get the prize.

Clarence Reynolds, TelecomTV (02:05):
There was a lot of discussion on our panels about agents or an agentic approach to ai. How could these be deployed in a telco environment to overcome some of the challenges you highlighted?

Andrew Coward, IBM (02:17):
Right. Well, agents are quite the buzzword at the moment, aren't they? When you think about what an agent needs to be, it needs to be able to understand a problem. It also needs to be able to resolve a problem. So it has this kind of independence that we haven't necessarily seen from ai. The way we're thinking about it in networks is that agents have to be responsible for understanding and managing individual elements or areas of the network. And this is around specialization. So we believe that a model that looks after radio might be a very different model or agent that looks after the backhaul, that looks after the IP network that looks after connections into the cloud, for example. And so we're starting to see the need for what I would call a hierarchy of agents and each with their own level of expertise. So my early example of how do we figure out why teams is running slow in Chicago would look at an agent that's looking at the radio network and ask the question, is there issues in that environment?

(03:25):
Is it the back hall and so on? It's a bit like the game clue or as I think it's called in the UK cludo, where you're trying to find out was it the professor in the library with the candlestick and a thing? And each of the agents will be able to provide answers and also provide remediation steps when the problem is being identified and or indeed get to the resolution before anybody's even noticed. So we're starting to see different technology needs at the type of LLMs that are being used, the type of technologies that are needed at each step in that. And one of the interesting things is that LLMs are not necessarily the best way of resolving these things. Traditional statistical analysis is important in this, but also new technologies that are kind of LLM based, but deal with things like time series data.

(04:20):
So it's not well understood. LLMs do not understand time series data very well because they don't really understand time in the way that humans might. And so as an example, IBM created a new LLM called Tiny Time Mixer, which we pushed into open source back in March. It's now been downloaded over half a million times, and it's designed to look at and understand time series data. So in that context you might apply that to say, well, I've got the problem that happened between this window of time, but also how do I understand what happened over time and what changes may have occurred or what problems may have cropped up through a drift, if you like, in the configuration or management of the network. And so these agents then, as we think about them for the network, may be quite different at different places. Another example is when you want to actually provide the what's the resolution?

(05:21):
What do I have to do? What do I have to type, where do I have to go that may well use a traditional LLM to kind of look at the manual for a device and then tell you instructions for how you might want to make that change or when you'd go execute those changes for you. So there is not one master LLM to rule 'em all in this. There are lots of different things that we think will play out and we'll need to coordinate ultimately in there. And that's kind of why the problem is so challenging. So what we're seeing is individual vendors actually provide AI capabilities into their products, but they might just be focused in one domain or their product domain or even on part of their infrastructure. And so there's going to be a real coordination effort required to bring all this technology together.

Clarence Reynolds, TelecomTV (06:10):
Andrew, you mentioned the network and networking is a much larger business than telecom. How is AI being adopted in other parts of the ecosystem and more broadly with IBM's customers?

Andrew Coward, IBM (06:22):
Yeah, I mean obviously AI has massive reach across pretty much everything we do at a societal level for most enterprises then who use networking. And indeed, most enterprises I'd argue are absolutely reliant on the network. If you are a bank or online website of any description, then you need that. And so managing the kind of two vectors of uptime and time to resolution, reducing time to resolution, those are your top priorities. Now, I think the difference between that and telco, obviously there are millions of users affected, but when a telco network has a problem and the societal impact of that can be very strategic, it's no surprise that telcos end up in front of senate committees where networks go down, but so do banks. And so this criticality of uptime, and if you like, not making the front page of the Wall Street Journal applies across the networking industry.

(07:26):
I think for IBM, what we're realizing across our customer base more broadly than networking is that there's a kind of specialization that's needed in each domain. If you're in car manufacturing looking at AI to reduce downtime on a production line, that's a different problem set than obviously, than the networking one broad models that understand many things may not be appropriate for the conversation we've just had around solving those things. So again, for each of the industries that IBM focused, which is kind of multifaceted in our case, we've got teams that are looking at how do we apply these broad technologies into each of these industries to go after these specific industrial use cases that then bring huge rewards to our customers as a consequence of deployment. I think most of the early use cases are document text base because that's where the LLM value has come from. But increasingly it's about the interplay between complex systems, topology of things, the sequence in which things happen, and that's really the next wave of AI and the applicability of AI into each of these industrial and business spaces that's going to change society the most dramatically as we roll these things out.

Clarence Reynolds, TelecomTV (08:50):
So what is IBM's approach to delivering AI products and solutions to telco operators?

Andrew Coward, IBM (08:57):
So there's a number of things that we're really approaching. First of all, we're embedding AI and have embedded AI in a number of our networking products. Sev one's a great example of that. A company required almost four years ago now, and using AI to predict changes in the network or predict outcomes as a consequence of changes in the network. So for example, if we see the number of low balances that might be across the network reduce, we can predict with accuracy when that's going to impact your network or how that's going to impact your customers more broadly. We have a number of initiatives that are introducing AI across telco infrastructure, looking at time series data, looking at individual domains and bringing that into a broader solution than we expect to talk more about that mobile congress at the end of February. And then more broadly in things like billing, customer care and so on, the let's call the more traditional, if that's even such a thing in LMS today, are absolutely being deployed into the telco space through our consulting practice.

(10:09):
And so I think IBM's very uniquely placed by having this combination of consulting group that's just absolutely capable of helping telcos go through to deployment. And then alongside the technology, and since I joined IBM around four years ago now, we've done five acquisitions in network space, each of them bringing a domain expertise and some level of analytics and AI into that and will continue to drive that. From IBM's perspective, the telco ecosystem needs to stay very open. We see challenges with vendors locking down or seeing an opportunity to lock down the ecosystem. Now we've gone from an environment where networks were and are very open today from the very foundations of things like SNM, standard ways of communicating and getting information out of systems to one where vendors are starting to see an ability to lock their customers up, if you like, with an end to end AI enabled solution. And that's not good for the ecosystem because it means that you're not able to compare AI models and information between vendors at that point. And so we see one of our roles in this environment as AI comes to the fore as really being, keeping the ecosystem for the industry open. And hence everything that we do, everything we build and everything we drive is around that open environment and hence why we're also publishing these models into open source, they can be consumed as widely as possible.

Clarence Reynolds, TelecomTV (11:53):
Well, Andrew, thank you very much for illuminating the complex integration of AI and telecom.

Andrew Coward, IBM (11:59):
You're welcome. It's been great seeing you again.