When you talk to networking and security folks, the consensus is that 2026 is the year not of the horse, but of the AI agent. But if you asked anyone in the pure-play enterprise software space, it’s more like the last three years or so have each been the breakout year for all things agentic AI, only to be proved wrong each and every time.

Over in the telecom space, the truth remains equally murky. Deployments have been exercised, not at a particularly grand scale, but enough for one exec to say that this year telecoms can already look forward to the next stage of agentic AI, with 2026 the year of so-called AI ontology.

That was the view of Mark Sanders, Telstra’s chief architect, who talked up the emergence of a structured, explainable knowledge plane that removes silo barriers between agents, freeing them up to become the workhorses of network automation. “We think for the autonomous network to reach level four or five is going to require a standardized, ontology-driven approach on the knowledge plane,” said Sanders at a recent Ericsson conference, touting this approach as the ultimate driver in next-level autonomous networks.

That same conference, attended by SDxCentral and focused on operations support systems (OSS) and business support systems (BSS), gave a more practical, here-and-now overview of AI agent uptake in telecoms, along with what agentic AI is actually being used for.

For operator BT, agentic AI is already yielding tangible value in IT service desks, especially as organizations shift from assistance to execution. This was according to Girish Mahajan, senior leader for mobile AI data/automation, who explained agents have been assistive to executional AI, which reduced ticket resolution times.

“It has reduced the time of the manual effort, and it has also increased efficiency of the service desk,” he said.

Mahajan, though, warned that the same autonomy that drives value also introduces unpredictability.

“The outcome of agentic AI is something unpredictable because it's continuously adapting during execution,” he said, adding a call for better design principles. “We need reflection-based architecture, and we need better AI/human collaboration. AI agents should learn from their actions and should work along with humans in their day-to-day.”

For fellow U.K. operator Vodafone, work has revolved around lighthouse projects: small-scale efforts to demonstrate the value of a larger business use case.

“It's quite a mundane use case around energy cost recovery. So obviously, energy is a huge operational expense for our industry,” said Simon Norton, digital/OSS engineering director, Vodafone Group. “It’s very complex, especially when you're working in that multi-market environment, to manually compare line by line with energy bills against your own data sets.”

Vodafone’s AI agents, therefore, have been automatically ingesting bills and comparing them to identify any tariff anomalies.

“It’s mundane but actually super valuable,” said Norton, who stressed operators should find a project with a clear value proposition and get it out into production quickly.

“You build the credibility, you start to get the funding into the system, and it buys you the time to work on that longer-term strategy.”

Norton added that for operators to get to this stage, they need deeper collaboration and flexibility with hyperscalers and vendors like Ericsson around open protocols and frameworks, avoiding black‑box designs and ensuring interoperability. Similarly, Mahajan called for scalable, modular AI platforms that integrate cleanly with existing OSS/BSS to engage in true co‑innovation rather than just handing over tools.

Ericsson’s horse in the race

Hassan Iftikhar, head of product domain data & analytics, Ericsson, meanwhile, called for better hyperscaler collaboration on scale, foundational cloud, and AI capabilities.

“The AI tooling, the security framework, we use those to industrialize and put agents into production… It’s pretty much an ecosystem that works together,” he said. At the panel, the data head revealed the vendor’s role in the agentic ecosystem through the use case of one operator needing help with catalog management, as well as scarce developer skills.

“They wanted to take the pain out of product configuration. So we designed a multi-agentic system where it basically helps product managers and marketers to configure and publish new instances through an actual language. So very complex catalog engineering, which can take weeks, is reduced to hours where you can search for reuse and launch.”

Iftikhar also revealed an OSS tool to help one operator’s engineers to diagnose and resolve issues within their operational instances – resulting in an agent that was seemingly too autonomous for the client.

“We put this use case together, basically taking an intent from an operations engineer, such as data diagnostics, and into it, we built the ability to take remediation actions automatically. What we sort of decided from that was a bit of a step too far to just throw that to an operations department for it to autonomously take steps. So we actually had to go in and build guardrails to limit that capability to a human oversight capability.”

“I think what we learned is that we have to sort of build that confidence in the team step by step before we can actually go to fully autonomous operation. Our learning from adjusting that use case was to be practical and adapt very quickly to what the business really needs.”

A two-horse hyperscale race

The same panel also revealed the hyperscaler status on AI agents, one that is perhaps not so surprisingly more advanced, with how the Big Three are servicing telecom clients with agentic workhorses. Aeneas Dodd-Noble, EMEA MD for telcos, media, and entertainment industries, Google Cloud, described how agents are already embedded internally for Google, and are being “used all the time.” But like with Ericsson, the MD was keen to demonstrate that challenges with agentic AI actually prove fruitful in the long run.

When some agents started querying multiple back-end systems at scale, he revealed, query volumes and access patterns overwhelmed those systems, exposing unexpected access-rights issues and performance bottlenecks.

“What often happens is that the back-end systems, they're not designed to be queried that way,” said Dodd-Noble, explaining that this forced teams to rewrite software before proceeding. “We're learning. The industry is learning how to use these things, but internally, we've done a lot to actually improve our efficiency.”

Dodd-Noble also implored operators to trust Google with their production data as “we have your best interest at heart.”

“We need to use that data to make agents accessible,” he reminded the providers in the room.

The issue of trust and governance came in another hyperscale expose via Antonello Arpino, head of operations simplified at Amazon Web Services (AWS), who said the first wave of generative AI created huge hype and “hundreds” of proofs of concepts, but operators and AWS alike lacked the operational practices, guardrails, and compliance frameworks needed to take these experiments into production.

“This raised hard questions of how you actually operate this, how it is going to be compliant, and what about security?” said Arpino.

To fix this, AWS has focused on building the right platform foundations and common rules, including traceability, so that solutions can scale safely in the “real world,” while also coping with country-specific regulations.

“We take governance and regulation very seriously. We signed up to the European Union AI act because we need to build trust, with the foundation being how you should be securing the data, securing the actual agents themselves, where they get the information from, ensuring their output is compliant with requirements, and then making sure that closed loop is again audited and secured. And also making sure the access to that data and the agent is again, fully audited and controlled by your IT team.”

This would likely be music to the ears of BT’s Mahajan, who asked for much closer involvement from hyperscalers on telecom security regulation, such as the U.K. Telecommunications Security Act, so that compliant architectures are “baked in” rather than left for each operator to figure out alone.

“This is what telco expects. If you come up with that kind of architecture, then it will make our lives much easier.”

This article first appeared in the SDxCentral Telecom Supplement.

To read more on

• Network APIs
• Free-space optics
• Wireless backhauls

Simply register for free