There’s a lot of buzz around the use of AI to reduce opex. Some is explicitly about “AIops”, some about “autonomous networks”, some about agents, and the obvious question is just how real this is. I’ve heard from over 400 enterprises and 78 operators on the topic, as well as a lot of my friends on the vendor side, and here’s a review of what they’re saying.
The vendors are a good place to start. All my friends, representing 27 different vendors or service companies say that they love an opex-reduction story. They’re sorry to see the number of projects driven by business benefits has declined, for the simple reason that this shrinks the pot they can draw on for increasing their own sales.
On the buyer side, it’s a lot more complicated. Senior management (VP-level) tends to like opex reduction as much as the vendors, because there’s pressure on them to cut costs if their companies can’t achieve target growth rates on sales improvement. But, and it’s a big “but’, there’s the ongoing concern that something targeted at opex reduction will end up breaking a critical tech service, feature, or application.
Part of the reason for that risk adversity is feedback from subordinate managers. Concerns about the risk of opex reduction, particularly involving new technology like AI, rises with every management level you descend through. I think the obvious reason for that is fear of personal job impact, because the perception of risk is highest among operations personnel below the management layer, but the specific issues still need to be looked at.
The top problem cited by both operators and enterprises is unproven benefit claims. Only 18% of enterprise and only 11% of operator ops teams said that they were offered “convincing” proof of benefits. Of the rough quarter pf both groups that said they’d tried to get contractual steps to prove out promises, the great majority said that vendors had backed off claims when told there would have to be validation strategies included in the deal. This “vendors won’t back claims with benchmarks” group were the most vocal in disputing vendor promises, not surprisingly.
Vendors hedge a bit in their comments on this point. I had nine cases where I got both seller and buyer comments on the same deal, and in those the vendor said that the problem was that there were too many variables introduced by the buyers’ own operations practices and team skills to be able to make firm promises. Buyers, to a degree, understood that, but their point was that where specific practices and skill levels were needed, fulfilling the need had to be incorporated in the project to deploy the new opex tools, and so did the costs.
The second-most-sited problem was fear of losing key personnel for job-security reasons. “As soon as I launch a project that’s justified by cutting jobs or hiring, the most mobile of my team start looking for another job,” one senior manager told me. No amount of promises, they say, will prevent this from happening because everyone seems to know of a case where those kinds of promises were broken.
Specific risks of a major technology-induced failure is a very close third, and when you focus specifically on an AI tool, this risk rises to second place. A similar shift happens where the operations focus is a business-critical process. “I might reduce my operations costs by twenty percent, but if I take down all my stores I’ve killed my job and maybe the company,” one CIO said. The problem here is that, like benefit claims, stability claims are hard to prove as far as buyers are concerned. Because right now there’s a very polarized view of AI stability/suitability, it’s a greater risk as far as the technical decision-makers are concerned.
With AI, though, there’s considerable variation in buyer attitude depending on the source. Right now, Juniper customers are more willing to accept that AIops tools won’t blow their network out of the water than users of tools from other vendors, and AI agents supplied by someone other than the provider of the infrastructure whose operation is targeted get the least love of all. Self-hosting of AI tools doesn’t do much better, because senior management doesn’t see/read much about the future of self-hosted AI.
Next on our list is an issue that everyone cites, but not at the top of their list; there’s a difference between improving operations group productivity and reducing opex by cutting labor costs. As one manager put it, “You can’t cut fractional employees”. Another said “If people get a job done faster, it doesn’t matter if there’s no other task they can take on.” How many humans are needed in an ops organization? Unless you’re willing to trust AIops unsupervised, the number is greater than zero. You probably have three shifts you need to cover. You have to allow for sickness and vacations and likely training. So many businesses can’t really reduce opex without eliminating that human oversight.
A related problem here is simply who to target. Both enterprises and operators say you can’t target senior people because you can’t lose them; they’re the ones who know systems and “how to keep the lights on”. But targeting junior staff or even limiting hiring reduces the unit value of labor and the available savings, and “can you expect AI to grow up to be a senior operations specialist?”
The final issue cited is perhaps the most telling; general fear that impacted personnel throughout the company will blame those responsible for any changes they don’t like. Decades ago, a wise CIO told me “the worst project you can propose is a conversion; it’s all cost and no benefit because the best you can hope for is that nobody will ever know you did anything.” There have been many stories about how disappointment needs a focus of a responsible party to punish. AI is no exception.
This issue plays in concert with all the rest, because of something a senior ops manager said “AI is the employee you can’t read a resume for, can’t qualify for a job, and probably can’t get rid of.” Ops management will surely be held accountable for anything that anyone feels has gone wrong because of ops automation. One operations executive at a joint banking network consortium put it this way: “If I mess this up, there’s thirty banks I can never work for.”
Right now, AI usage (says the Economist) is growing faster than AI revenue, suggesting that people are fleeing one free model for another when they run out of free tokens. That’s not the behavior of someone who’s found value, it’s the behavior of someone despairingly seeking it. It argues that demonstrating an AI business case is a major issue. That’s true overall, and it’s surely true for operations missions. Both enterprises and operators believe AI can be helpful, but they’re not sure whether the help can justify the cost.
