By Adam Mills
AI Editor
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.

Newsweek
Editor's note: This is AI Impact, Newsweek's weekly newsletter where each week, we will explore how business leaders are unlocking real value through artificial intelligence.
Read More on AI
Tap or click here to get this newsletter delivered to your inbox.
Thanks, as always, for joining me.
Some ideas get repeated in AI circles so often that they start to sound less like arguments and more like laws of nature.
My colleague Gabriel Snyder takes aim at one of them in his latest piece, “Stop Saying ‘Jevons Paradox’ Will Save the AI Boom.”
The shorthand goes like this: As AI gets cheaper, usage will explode, making lower prices a reason to be even more bullish on the infrastructure buildout. Gabriel went back to William Stanley Jevons’ original 1865 argument about coal and found something much less reassuring. Jevons was worried about how long Britain’s advantage could last once cheap energy became more expensive and competitors caught up.
For today’s AI industry, falling costs could benefit companies using the technology even as they squeeze the businesses selling the underlying compute. His piece makes the case that demand can keep growing while the margins move elsewhere in the value chain.
It’s an extremely sharp piece and worth reading in full, especially if you’ve heard “Jevons paradox” invoked as proof that cheaper AI automatically strengthens the case for ever-larger infrastructure investments.
This week’s Signal Capture turns to a different assumption: that bigger AI models are always the right tool for the job. Some network problems may be better handled by smaller, more task-specific models running closer to the data they need rather than sending everything back to a large centralized system.
Read on for more.
Signal Capture
Signals from the frontlines of AI adoption
Calix CPO: Not Every AI Problem Needs a Bigger Model

...
Security threats, network faults and service issues affecting broadband subscribers can often be diagnosed and addressed using data from a particular piece of customer-premises equipment or network location. Shane Eleniak, chief product officer at broadband technology company Calix, argues that those problems do not always belong in the largest general-purpose AI model a company can access.
“When you start to get into reasoning problems, you want a specialist, right?” Eleniak told Newsweek.
He compared the choice to hiring a mechanic. One option is a highly capable generalist with access to manuals; the other is someone trained specifically to diagnose and repair the problem.
Calix is applying that idea to communications work. The company has released an open-weight Communications Language Model on Hugging Face, giving developers a model designed for tasks such as security analysis, network troubleshooting and subscriber support to experiment with and adapt. In 2027, Calix plans to introduce Calix Gems, specialized models built for a narrower set of network tasks.
For Eleniak, “right-sized” means matching the compute, storage and knowledge required to the task. Many network problems rely on information generated in one place, such as activity around a particular endpoint or network location. Eleniak's argument is that those tasks can be handled by smaller, task-specific models running closer to that data instead of sending everything back to a large, centralized model.
Where the model runs can also depend on the task, with some work happening closer to network equipment and other work handled elsewhere in the infrastructure.
“It’s going to be a mix of centralized and distributed. Just has to, right? Given the nature of the problems we’re trying to solve,” Eleniak said.
Calix also plans to use those models alongside AI agents that can recommend or carry out actions. Eleniak said a person should still approve actions that materially change a subscriber's experience or would be difficult to reverse.
Tightly defined tasks can make this architecture easier to deploy, he said.
“Smaller models become much more approachable because they become very finely scoped,” Eleniak said.
The business need also differs by provider size. Smaller broadband companies often lack enough people to tackle everything on their lists, while larger operators contend with many employees, systems and processes that have to work together. He sees models running across different parts of the network as a way for smaller teams to take on more work and for larger organizations to reduce some of the coordination required across systems.
Providers should eventually be able to point to concrete changes such as adding subscribers, increasing average revenue per user, lowering costs or completing more work with the people they already have, he said.
Trust will also influence adoption. Eleniak compared AI with earlier shifts in banking and e-commerce, where people gradually became comfortable moving money digitally or making purchases online.
“Once you build trust in the technology, then you get adoption,” he said.
Routine subscriber tasks could become much simpler. Eleniak offered a hypothetical example of a broadband subscriber asking their provider's AI agent to set up guest Wi-Fi for four hours. The agent could generate a QR code and later ask whether the subscriber wants to extend or end the access.
“Technology is no longer the problem,” Eleniak said. “The enemy is complexity.”
Prompt Injection
What’s one recent insight you’ve learned about AI?

...
“AI made answers cheap, exposing a problem most companies have had for years. They weren't necessarily short on answers. They struggled to turn those answers into action.
Give a marketing team instant answers, and you can still end up with more reports, more meetings and the same number of changes in front of customers. The insight may be better, but the process around it hasn't changed.
That shifted the way I measure AI, both within Amperity and with the brands we serve. I don't just ask how many questions a tool answered. I ask how many good decisions reached customers, how quickly, and whether the team learned something useful from the outcome.
That's why we built Pér to help carry a customer insight into an approved action and then monitor what happens next. The opportunity isn't simply to make analysis faster. It's to shorten the time between understanding a customer and taking meaningful action on that understanding.
The return begins when a customer feels recognized: the brand understands why they bought, responds in a relevant way and gives them a reason to come back. Faster analysis matters when it helps the business earn that next interaction.”
Have your own lesson to share? Email me at: a.mills@newsweek.com
Run Log
AI use case of the week
Conviva AI Finds Costs Hidden in Resolved Customer Chats

...
Resolution rates looked healthy at one e-commerce brand, but some closed customer-service chats were costing more than the metric revealed. In certain refund conversations, customers received more money back than they had requested or already agreed to accept.
Keith Zubchevich, president and CEO of Conviva, a real-time performance analytics company, said its AI analytics agent reviewed more than 75,000 conversations the brand’s customer care team handled in one week. For refund chats, it compared what the customer requested or accepted with the amount ultimately issued.
The analysis surfaced cases in which a customer who had agreed to a 20 percent refund, for example, received 25 percent instead.
“Over-concession showed up in under 1 percent of chats, but at this volume that’s more than 10,000 overpaid refunds a year,” Zubchevich said.
Conviva said resolution rate alone did not reveal whether the company was giving customers more than they had already agreed to accept.
AI gave the brand a way to test whether a healthy-looking service metric was hiding a cost problem underneath it.
“The resolution rate alone can’t tell you whether a customer care agent is working,” Zubchevich said.
Have an interesting AI use case to share with me? Email me at: a.mills@newsweek.com
Context Window
■ Zendesk introduced an AI service platform that moves away from traditional seat-based pricing, charging customers for eligible resolutions and completed business work while adding tools that connect external AI agents to its service workflows. [Zendesk]
■ A study of 322 managers experienced with generative AI found that higher AI literacy was associated with more AI-related job redesign, innovative work behavior and workplace well-being, with greater autonomy and competence helping explain the relationship. [Nature]
■ A Foundry survey found that 34 percent of organizations have already moved AI or machine-learning workloads out of the public cloud and another 38 percent plan to, as cost, security, performance and sovereignty concerns reshape where enterprises run AI. [CIO]
■ United is restricting unauthorized third-party AI agents from accessing its booking and account systems, citing risks involving customer credentials, unintended transactions and regulatory requirements that still require human acknowledgment. [United]
■ A Gartner survey of 298 infrastructure and operations leaders found that 95 percent said their organizations had experienced at least one AI agent-related incident in the past year, while 57 percent rated their governance as weak and only 15 percent had automatic circuit breakers in place. [Gartner]
Transfer Protocol
Tracking executive moves across the AI landscape
Ilya Bonic, who led Marsh’s Workforce and Rewards practice, has taken on the role of chief AI officer for People and Investments, developing AI-enabled client solutions while expanding automation, process engineering and AI adoption across the business.
Nikki Shaver, co-founder and CEO of Legaltech Hub, will join Davis Polk & Wardwell on Dec. 1 as chief AI and knowledge officer, becoming the firm’s first standalone executive dedicated to AI and knowledge management.
Yann Ravel-Sibillot, co-founder and former CEO of Flowie, is now chief AI officer at Sovos, guiding the use of AI across its agentic tax compliance platform and integrating technology from its recent acquisitions.
Mark McSpadden, most recently head of innovation at USAA Labs, has joined FCM Travel as chief product officer, taking ownership of its end-to-end product and AI strategy, including development and integration of its core capabilities.
Sandeep Aramalla, promoted from enterprise architect to chief AI officer at Groupize, will steer the evolution of its Aime platform, with priorities spanning interoperability, AI governance and measurable business outcomes.
Know someone on the move in AI? Send job change info to a.mills@newsweek.com
Magic Moment
What’s the most fun or unexpected way you’ve used AI lately?

...
“I asked AI to save my favorite shirt.
Some background. In 2006 I bought a Tommy Bahama Hawaiian shirt when I was on vacation. Black, with giant green banana leaves. Subtle it is not. It's also my favorite shirt ever. Twenty years later, the collar finally frayed past the point where I could respectably wear it in public, and I couldn't find a replacement anywhere.
So one morning, in between board decks and product strategy, I uploaded a photo to Claude and asked it to find me a replica.
It turned out to be a very good detective. It figured out the era and the print family and fabric. It taught me the terms vintage sellers use for shirts like mine. It combed eBay and Poshmark one listing at a time. When I told it the first batch was the wrong size, it went back and found Larges. It even turned up the fact that my exact shirt, in my exact size, had recently sold for $18.
Within a week of setting up watchlists on both platforms, my replacement shirt had arrived!
Because the best thing it did was answer a question I didn't ask.
I asked for a replica. It gave me listings. Then it pointed out that the shirt I love is the one from Lanai, not a lookalike from eBay, and that a good tailor can rebuild a frayed collar with fabric harvested from the shirt tail, where nobody will ever miss it. I had no idea that was a thing. ‘Find a substitute’ became ‘save the original.’ That's not search. That's what a friend who knows a lot about menswear does. So in addition to the new purchase from eBay, my tailor actually saved my beloved shirt as well.
Here's the thing. I run an AI company. Most of what I do with these tools is serious: strategy, writing, building software. But the difference between a tool and a collaborator shows up best in the small, personal problems. No playbook. No expert you'd ever hire. Nobody keeps a vintage Hawaiian shirt consultant on retainer. Now everyone has one.
Those of us in the industry love to talk about transformation. For most people, AI won't show up as a transformation. It'll show up as a shirt.”
Experience some AI magic? Tell us about it at a.mills@newsweek.com
Tap or click here to get this newsletter delivered to your inbox.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines