It’s easy to imagine how artificial intelligence could benefit communications service providers (CSPs), not least by automating complex operations in radio access networks (RAN). Often though, that story gets boiled down to how AI will “drive down network costs and complexity,” and that’s a problem. First, because people working in CSP organizations will hear that message as, “We’re going to replace you with AI.” But the bigger problem—and the reason why telco RF engineers and network management teams aren’t going anywhere—is that it’s just not accurate.
In the coming years, CSP networks will grow more complex, not less. That’s just a side effect of adopting more virtualized, disaggregated, cloud-native infrastructures. Where AI will provide tremendous value is in improving the manageability and maintainability of those infrastructures. Far from replacing CSP teams, AI will empower them to manage an ever-larger and more dynamic RAN more effectively. So that the same people, with the same resources, can manage 30% more sites daily, without increasing their workload. That’s the real power of AI. There are, however, two key ingredients that CSPs must have to capitalize on AI: operational best practices on which to train machine learning models and granular network data. The best practices are already in place in the institutional knowledge built up over decades in CSP network teams. It’s the granular data that’s the issue. That data exists—it’s traversing the network right now—but to use it for AI, telcos need to be able to access it. Today, they can’t.
Who owns your network data? (Hint: not you.)
To understand why granular RAN data is so important for the future of telco AI, we should review how CSPs plan to use AI in the first place. Ultimately, you’re training a model to do something in the network—optimize spectral efficiency, accelerate root cause analysis, identify golden configurations for cell sites in X locations with Y characteristics, and so on. To do that, you once again start with the right operational processes. (There’s little value in an AI that can automatically do things if it does them poorly.) But you also need a huge data set to train on, so the model can analyze how a given use case was performed, and how the rest of the environment responded, across thousands of repetitions and variations in granular detail.
Effective AI engines learn how minute changes impact the health and performance of the network. This is, after all, the whole point: to detect patterns and correlations that would escape human analysts. Today though, CSPs can’t access network telemetry data at the level of detail necessary for algorithmic training. Not because the network doesn’t collect it (it does), but because network vendors choose not to share it.
Vendors share maybe 30% of the data that RAN equipment collects under OSSii licensing agreements. But they keep the rest to themselves, even though 100% of that data reflects a CSP’s own subscribers and operations. If you’re wondering why telcos continue to accept a model where they literally pay a vendor to access their own data—and the vendor doesn’t even give them all of it—you’re not alone.
Envisioning AI-enabled operations
Having even a small slice of telemetry data makes a big difference in helping CSP teams manage modern networks. But imagine what would be possible if they had access to all network telemetry and could collect it in one place. They could:
- Continually optimize operations: There are thousands of tasks involved in engineering and operating a mobile network, from cell site turnup to vendor management to equipment acquisition, and many others. Practically all of them could benefit from the ability to identify best practices and train an AI to execute them. This is basically what CSP engineering and operational teams do today—just more slowly and manually.
- Build language model (LLM) assistants: WhenCSPs collect telemetry information in a data lake, they can use LLMs to assist human engineers. For example, RAN engineers typically start their day by bringing up the top alarms in the network. Often, those same problems cropped up previously on different sites and will pop up again on still other sites in the future. The most difficult issues can float around the network for months. When CSPs have all network data in one place though, RF engineers can enter queries like, “Show me the health of sites with X configurations over the last 90 days,” to isolate root causes more quickly. They can then update golden configurations and have the LLM generate subroutines to automatically implement them across similar sites. (For CSPs using Open RAN architectures, this capability already exists in applications like RANGPT from Aira Technologies.)
- Use network digital twins: Many CSPs are now implementing virtual representations of the network in the lab to understand how changes will affect services before implementing them. These can be especially helpful for difficult RF problems, such as engineering handoffs across local access and transport area (LATA) boundaries. When engineers can access all network data and feed it into a digital twin, they gain the freedom to explore how small changes will affect performance during such handoffs, without risking disruption to live services. These are high-order operations that require years of RF expertise, but the combination of AI and digital twins will help engineers perform them much more quickly.
- Monetize network data: Finally,CSPs have an opportunity to exploit the expert engineering and management know-how they’ve built up to not only optimize their networks, but monetize those insights. Keep in mind, the telemetry data in the network is not random. It’s the result of years of careful curation as CSPs painstakingly improved processes. For example, RF engineering teams can turn up new cell sites much more quickly than they could 20 years ago. That only happens through trial and error across thousands of deployments—making a minor change here when using this frequency, changing this setting there for another—to identify golden configurations. The resulting data sets reflect decades of institutional investment and effort. And they’re enormously valuable to anyone seeking to train AI models for telco use cases.
All these use cases and many others are now possible, but only when CSPs can access granular network data.
It pays to be open
CSP operational teams have wanted more and better network data for years. In fact, that was supposed to be among the biggest benefits of Open RAN. Yes, you would increase complexity as you disaggregated the network, but the payoff would be that you captured much, much more telemetry data to optimize and automate it. Today, some CSPs are pulling back from wide-open multivendor RAN initiatives, instead choosing a single RAN partner to ease integration. There’s nothing wrong with that strategy, provided you hold the line on demanding open access to network telemetry data.
This is, after all, your data. It’s extraordinarily rich and extremely valuable—to network equipment makers, independent software vendors, even other telcos. Someone is going to make a fortune applying it to AI and algorithmic training. Shouldn’t that someone be you?
By: Ken Gawelek, Michael McGroarty, Janco Terblanche
