Catch the video at top, listen to the audio edition and read our transcript below, or watch this and future episodes on YouTube.
This podcast is written and hosted by Diana Goovaerts. It is edited by Diana Goovaerts and Matt Rickman. Liz Coyne is our executive producer. Special thanks to guests Mark Austin and Louis Powell.
To learn more about the topics in this episode, check out:
Diana Goovaerts: AI is moving from experimentation into operational workflows, but not as a simple plug-and-play layer. Today, we're unpacking what it takes to make AI useful, trustworthy and economical inside telecom networks.
I'm Diana Goovaerts, and this is "The Five Nine."
Telecom networks have been using automation and machine learning for years, but generative AI and agentic systems promise something more ambitious: Models that can understand telco standards, reason through network problems, surface context, recommend fixes and eventually take action. The catch? Networks are critical infrastructure.
"Pretty good" isn't good enough. So we asked Mark Austin, a VP in AT&T's data office, and GSMA's Director of AI Technologies, Lewis Powell, what it actually takes to put AI to work in telco networks - not in a demo, not in a slide deck, but at production scale
Louis Powell, GSMA: You can't just sort of take a demo and then scale because of the multidimensionality of evaluation you need to go through in order to understand what it means to take in production.
Mark Austin, AT&T: When you get into closed loop things, though, you have to crawl, walk, run, just like we've been doing for years with self-optimizing networks. So you would start in a small cluster and you would kind of apply it, get confident there, human in the loop, for sure, many times, and then you get very confident and you can kind of grow in size and scale.
Diana Goovaerts: For all the hype around AI, operators have a very specific problem. General purpose models don't always understand telecom. Ask a model what a cell is, and if it thinks biology before base stations, you have a problem. That sounds almost funny until you imagine that kind of confusion multiplied across millions of inferences inside a network operations workflow.
That's the gap that the Open Telco family of models, or OTEL, is trying to close. GSMA and a group of operators, vendors, and other industry players have been building and benchmarking models trained on the language of telecom - standards, network documentation, research, performance data, and all the other messy domain-specific knowledge that general models often miss.
Louis Powell, GSMA: The frontier models, the majority of open source models are trained on the internet, and a lot of the data that defines telecoms, the standards, the technical documentation, the vendors, the network data, isn't readily available on the large pre-training datasets that the frontier models and open source providers use to train their models. So we thought it was very important that we trained models made them open source and available so that these models deeply understood the telecoms domain.
Diana Goovaerts: We asked Austin and Powell about what comes next, how telcos move beyond model demos into real network use cases, what kind of guardrails have to be in place, and why the future of AI in telecom may come down to three things: data, compute and expertise.
The first point both made is that telecom is not just another enterprise vertical. Telco networks often run on specialized standards, configurations, performance indicators and operational language. And historically, big closed models didn't score especially well when tested against that domain knowledge.
Mark Austin, AT&T: We graded all the closed source models against telco, and they didn't do that well. They were hitting, like, if you ask it a question about various telco things that it should know from the standards, it would get it, like, 60 to 70% of the time it'd get the answer.
Diana Goovaerts: That matters because AI in the network is not a trivia contest. It is about troubleshooting, planning, root cause analysis, ticket handling, and, for some future use cases, closed loop action. If the model starts from the wrong context, everything downstream gets riskier.
Louis Powell, GSMA: We've noticed this by benchmarking how models understand things like 3GPP standards, the standards documentations for telecoms, and noticed that even as models were getting more advanced in maths and coding, they weren't getting any better in the telecoms domain. So it sounds simple, but at scale, when you deploy these models and you're inferencing millions of times, it's really important for them to be accurate on these technical details.
Diana Goovaerts: OTEL is essentially an attempt to teach AI to speak telecom. According to Austin, the work pushed telco benchmark performance from roughly 60 to 70% accuracy up to 90%. And both he and Powell said the models have already seen more than 30 million downloads, a surprisingly high figure for an enterprise domain model.
The bigger point, though, is strategic. GSMA doesn't want telecom's AI future to depend on only a handful of outside intelligence providers. Powell framed open models as a way to preserve choice for operators, whether they want hyperscaler partnerships, sovereign on-prem deployments or a hybrid approach
Louis Powell, GSMA: We want to look at the health of industry in the market, and we're very concerned about acute supply chains or where we see shortages in choice and options for our operators. So you look on the hardware side, there's very few vendors of RAN, not many in each geography, and we're very concerned that we have the same situation on this sort of intelligence frontier, where there's only four or five big players who own intelligence in the telecoms industry.
So what we want to do as an trade association is ensure there is a diversity of supply so that our operators have choices in which direction they go.
Diana Goovaerts: So what does putting AI to work actually look like in practice? Today, Austin says much of the value is still in decision support, helping planners, designers and troubleshooters ask better questions, identify likely causes and move faster through complex workflows.
Mark Austin, AT&T: You have to have a lot of expertise to actually deploy and run those things. So, it's your standard capacity dimensioning that you have to do, and then you have to have resiliency, so you have to have failover. So you have to think about you're in the driver's seat to do all those things, and it's not easy. So you have to get really good at that.
Diana Goovaerts: It's worth restating. When it comes to AI, operators have to crawl, walk, run. Start with human-in-the-loop recommendations. Test in a small cluster. Build confidence. Expand carefully. Only then do you start talking about closed-loop actions.
Louis Powell, GSMA: We want accurate models. We want efficient models. We don't want to use the large models for every single use case, so we want to make sure they're efficient, but also we want to make sure they're safe.
So evaluating the responses, they might be accurate, but are they compliant to the guardrails put in place around a network? Are they accessing the right tools to make the right decisions, or are they looking for permissions that could violate, uh, internal standards?
Diana Goovaerts: Trust is the big unlock here. Before an AI system can recommend a fix or someday trigger one, it has to be tested against known truth sets, operator data, holdout sets, guardrails and failure modes.
Mark Austin, AT&T: It starts with evals, right? So any time you're building any kind of system, you have to have a test set, a truth set, how you're gonna evaluate it, you have a training set and you have a holdout set. So you put it on the training set, then you test it on the holdout set, did it actually get it right, and you try to learn.
Diana Goovaerts: Powell added another layer. The threshold for trust depends on the use case. If the benefit is marginal and the downside of a bad answer is high, the business case starts to fall apart. But if the AI delivers a transformational lift in productivity, savings or revenue, and the guardrails are strong enough, the risk may be worth exploring
Louis Powell, GSMA: It's really a sort of, it's a case-by-case basis, and some of the dimensions I would look at is what is the uptick? Lots of times you deploy these models, but the results aren't transformational. So if the results aren't transformational, you know, they're marginal, they're 5%, 10%, is it worth the cost and risk of dealing with these non-deterministic models?
I would say, well, probably not worth the effort. If you do see areas where you have a significant uptick in productivity, in cost saving, in revenue generation, then maybe there's a, a risk they're worth exploring, and then ultimately, it de- depends on the sort of outcome of a bad answer. So with customer service, you probably want very high standard because reputations can be lost very quickly. Having a sense of what is the benefit, what is the cost of a negativity dictates what your threshold is for sort of accuracy.
Diana Goovaerts: That's where telecom differs from many enterprise AI deployments. A bad chatbot answer can be embarrassing. A bad network decision can be operationally and financially consequential. Powell pointed to runtime safety, tool permissions, sovereignty and explainability as key parts of the trust equation.
One of the most interesting pieces from AT&T's approach is its AI smart router. Austin described it as a kind of command and control layer for token economics, deciding where a request should go, which model should handle it, what latency and accuracy targets apply, and how to fail over if capacity or performance becomes a problem
Mark Austin, AT&T: My boss likes to call it money ball for tokens. We almost call it token ball. So we play token ball, and you try to figure out where do you route those things. You pick the most economical place to generally route the model. And everything you have to score, so everything has an eval, and everything will have a service level objective. So you'll have a certain latency that, that hits this thing, you wanna change decision.
Diana Goovaerts: That may sound tactical, but it's actually strategic. Operators don't necessarily need the biggest, most expensive model for every request. Austin said AT&T has found that for some workloads, only a relatively small share of requests require a premium model. The rest can be handled by more economical models if the router is smart enough to know the difference.
Mark Austin, AT&T: Depending on what they're asking, you might not need the premium. In fact, what we found is, depending on what you're doing, it could be 10, 15, 20% at most you might need a premium model. You can get away with the economical model at, you know, otherwise. So that's very useful.
Diana Goovaerts: A model router like AT&T's lets an operator decide when to use open models, closed models, on-prem resources, cloud-hosted capacity, or various overflow options. In other words, AI isn't just one model, it's a portfolio, and the economics only work if that portfolio is actively managed.
For all the excitement around agentic AI, Austin offered a useful reality check: not everything in the network needs to become an agent. Telecom has been using machine learning and automation for years. The question is where agents add something incremental, especially in the workflows that require context gathering, tool use and handoffs across systems.
Mark Austin, AT&T: Although I love agents and we use agents for a lot of things, not everything should be an agent. It's almost like not everything's a hammer. You've got a new tool in the toolbox, and this is a really fancy tool. It could do amazing things.
But many times, you can get away with traditional methods, traditional machine learning. It's not a one-size-fits-all. Not everything's a hammer, not everything's an agent.
Diana Goovaerts: One pattern Austin described is using generative AI or agents to explore a problem quickly and then turning parts of that workflow into more traditional machine learning once the classification or decision path is clearer. In other words, use AI to find the path, then simplify where possible.
Powell's warning fits in here too. Scaling frontier models out of the box can blow up the business case. At production scale, small hallucination rates become operational risk and token usage can tank ROI.
Louis Powell, GSMA: The biggest misconception that operators have of deploying models on the network is the fact they can take an enterprise frontier model and deploy it out of the box at scale. What we see is that these models are unbelievably capable, they can be very convincing and they can be very accurate at small scale deployment.
But when you come to production level, where you're requesting or inferencing these models potentially hundreds of millions of times a day, then A, hallucinations or inaccuracies rack up, and B, your token usage will go through the roof, and that makes the business case for this deployment - you know, the ROI just vanishes.
Diana Goovaerts: If there was one understated theme that kept coming up in these conversations, it was context. Austin said agents are really only useful when they have the right information. That means network configuration, historical trends, device context, local conditions, business priorities and the tools required to reason through a problem
Think about a low throughput issue on the RAN. Is it a one-time event? Does it happen every Tuesday at noon? Is it a configuration issue? A device issue? A congestion pattern? A high-value venue or customer-impacting area? The AI system can't reason well unless those kinds of signals are available, standardized and accessible.
Mark Austin, AT&T: Your agents are useful when you give them context. They can't really solve anything unless you give them context. So a lot of it when you're thinking through these things is like, how am I gonna surface the context so that they can actually reason through those things?
And I don't think we're near as mature as to where we need to be in surfacing context in terms of those things. Everybody has a little bit...they have the same conceptual network elements, but sometimes they're configured differently, so there's not really a good standard to surface the context for agents.
Diana Goovaerts: So where does this leave operators? Well, Powell's view is that building, training and deploying models in agentic systems will become a core IT competency, more like websites, apps and Kubernetes than a niche science project.
Operators can partner, and many will, but they can't afford to have no internal understanding of how these systems work.
Louis Powell, GSMA: Now is the time to learn rather than in three, four, five years when these models and systems are becoming fundamental to how you run your business.
I would take a bet that in the very near future, building, training, and deploying models, agentic systems will be a core competency of pretty much all AI teams and IT teams in general. So outsourcing them to partners and not getting a sense of how that works now at this very early stage is just delaying the cost, because the cost will only get higher over time as you're looking to deploy these at scale.
Mark Austin, AT&T: If you wanna experiment with OTEL, you know, take it, customize it. It's out there for your own trying. You know, add your own adapters to train it on your own kind of internal data and documents. That's encouraged. Try to get some value out of that and think about tools that we got. Maybe we can kind of standardize some tools that everybody will reason up as well.
Diana Goovaerts: That is all the time we have for today. Make sure you like and subscribe wherever you are listening or watching, and stay tuned to fierce-network.com during the week for all the news and the latest insights from experts across the industry. We will catch you next time.