One of the hot topics in telecom these days is the role AI could play in managing opex. I blogged about this a bit recently, and got some LinkedIn comments. The operative presumption is that telecom could cut jobs through the use of AI, and Light Reading said that “The combined workforce of the 20 operators tracked by Light Reading shrank by another 52,000 jobs or 4% last year, and AI is only just getting started.” Truth is that job-cutting at telcos is hardly new, and AI hasn’t played any role in it so far. What could AI do that’s not already being, or been, done?
I’ve dug back to the foundation of Andover Intel and my conversations with 88 operators to get as good a handle on their views as possible. Remember that I’m not surveying or questioning them, but analyzing the things they say spontaneously as they react to industry events and my own views.
Let’s make an assumption, one I think is correct, to open this discussion. Opex for telcos is mostly human costs, which means that reducing opex means reducing jobs. In order to do that, you have to take stuff that humans do and either do them another non-human-intensive way (automate them) or you have to change your process so that stuff isn’t needed at all.
The economic cost of technology has always been compared to the economic benefits if could bring, and in many cases this comparison would contrast human-driven solution cost to IT cost. When I got into computer programming as a young man, it was common to relate the power of a computer to the number of experts it would take to do the same job; “It would take a hundred mathematicians a hundred years to do this”. Obviously, that many mathematician-years would be costly, so computers’ cost is likely justified. The notion that AI could save telcos money is based largely on this paradigm, which we could call the “human displacement theory of opex reduction”. We do stuff with AI instead of workers. But this paradigm has two weaknesses.
First, you have to actually be able to displace the workers. Eliminating work is not the same thing as eliminating workers. You can make a worker more productive with AI, but unless that improvement can be translated into either job cuts or reductions in incremental hiring, you’ve not saved anything. Is the target worker’s need truly displaced, or is the worker still needed often enough that they have to be retained? Is there anything valuable they could do with the time AI saves them? Can you acquire a worker at a lower wage if AI augments their effort? You need to address these questions.
Second, you have to compare the benefits of automating the work with the potential of eliminating the need for it. Can the process itself be re-framed to make it unnecessary to do the thing the worker does? Is there a more cost-efficient way of building infrastructure, a better way of handling a business process, that will eliminate the need to do certain things? If there is, then doing that will save both human cost of doing the job and the cost of doing it with AI. You then have to see what the cost of the refactoring would be, and whether savings justifies it.
The reason this is important is that for decades, telcos have been eliminating jobs by refactoring. I remember having to make any interexchange calls through the human operator. I remember when if you got broadband, you had a tech come out and do the installation. Today, increasingly, you’re shipped something and given instructions yourself, and you can’t get a human operator to assist you on a call no matter what you do. The great majority of jobs eliminated by telcos were eliminated by refactoring, and that’s still true.
The extent to which refactoring works depends on two things. The first is simple; does the task actually require hands-on action? The more human work is explicit in a task, the less amenable the task is to our refactoring options. The second is more complicated; can the human work be refactored or reassigned? If current practices require that a trench be dug, a cable run, and a connection made in a remote concentrator, then you either need to do that yourself or convince the customer to do it. If you can pre-deploy these assets, or convince a real estate developer to do that, you can reassign the work. If you can’t do either, and want AI to displace the human initiative, you better be shopping for a robot that can trench without hitting pipes, trees, flower beds, pets, kids, and so forth.
This is one reason why operators tend to focus on mobile or FWA. The greatest concentration of operations cost is in the access portion of a network. With wireline, access is physical and so physically touching things is essential. In mobile, access is almost virtual. You can beam-form RF to impact where it’s available, you don’t have to dig a new hole (unless you’re out of range of current towers). RF access makes a lot of the access opex tasks hands-off, which means that they are more easily refactored into something AI can impact.
Another factor to be considered is the green versus brown field problem. The most expensive element in a wireline network is the wires, or fiber, and it’s not something you can change or move easily. Thus, much of the wireline upgrades have been happening either in areas of locally high “demand density”, meaning higher income areas, or in new developments. Mobile networks, in contrast, get refreshed when the 3GPP releases a new generation of technology, and operators routinely budget for the upgrade. During this upgrade, opex-saving or opex-automating technologies could be introduced without requiring the changes meet their own business case.
All this explains operator views I hear. In the main, they believe that while AI could likely provide some benefits in netops, netops in the sense of the management of the network and its interior elements, isn’t a huge component of opex overall. Many of them believe that even here, they’d be smart to spend money on simply building in more capacity and route density to make the network more self-healing and resilient, which is a form of refactoring to eliminate the need for a task rather than shifting how the task is handled.
It’s hard to say, based on the input I get from operators, whether there really is any likelihood that AI will play a major role in opex reduction, meaning whether it will impact jobs. The fact is that most of the low apples in that space have already been picked, and my own models suggest that to go much further you’d likely have to seriously refactor infrastructure. Would the benefit justify the cost of that? I don’t know, but I do think that it’s very likely that you could change network technology to eliminate the need for some types of jobs more easily and cheaply than you could automate the jobs. AI, then, may not figure strongly in the telco opex future at all.
