Your agent is only as good as its intelligence.

AI agents are capable of executing multi-step workflows across operational tools without the need for a human to initiate each step and are now deployed in production at fiber operators.

According to vendors interviewed for this story and analyst research, the factor that determines whether the technology delivers value is not model capability but rather the accessibility and quality of an operator’s data. 

Boston Consulting Group’s 2025 report, The Widening AI Value Gap, found that agentic AI accounted for approximately 17% of total enterprise AI value in 2025, up from a negligible share in 2024. This figure is projected to reach 29% by 2028. The report found 46% of surveyed companies piloting or deploying agentic systems and ranked telecommunications second among all sectors in AI maturity, up 11 points year over year.

Vendor activity has followed. In June, Nokia announced an agentic AI framework for its Network Services Platform, with agents working within operator-defined policy boundaries. Commercial availability is planned by the end of 2026.

Telstra and Ericsson announced a collaboration in October to advance autonomous networks, centered on building a "knowledge plane,” an information layer using data, AI, and reasoning to monitor, analyze, and control the network. 

"Developing and validating ideas and technical possibilities in real environments is essential to closing the gap between aspiration and execution," Telstra chief architect Mark Sanders said in an October 2025 Ericsson press release.

Most operators considering agentic AI are not evaluating it from a clean starting point. They are running legacy operations and business support platforms (OSS/BSS) that were designed before modern application programming interfaces (APIs) were standard, and before data accessibility was considered a product requirement.

Stephen Farnsworth, Vice President of Go-to-Market at gaiia, a Montreal-based OSS/BSS provider that closed a $40 million Series B in May 2026 and reports more than 40 ISP customers, says most operators are trying to retrofit a legacy platform around AI rather than build for it. 

New builds are unburdened by prior vendor decisions and are in a position to select modern software from the outset. The operators moving most urgently toward modernization are legacy carriers. 

"We're spending a lot of our time with companies that have 10,000 subscribers, 100,000, 200,000 subscribers, where they're looking at their stack and saying, 'I'm getting crushed. I have no ability to move quickly,'" Farnsworth says.

Data Access as the Threshold Condition

Farnsworth states that agentic AI cannot function for operators whose data is inaccessible. 

“People think this agent thing is real, and that they can benefit from agentic AI—but they’re on on-prem software where they have no access to the data,” he says. “I’ll tell you: agentic AI is real, but not for you.”

Farnsworth cites the historical practice of OSS/BSS vendors charging operators to export their own data. 

“If you wanted access to your data, you had to pay them. You go to your vendor and say, ‘Hey, I want to do an export here,' and they say, 'Okay, great, happy to do that. It’ll cost you $20,000.’ Or ‘We have a professional services group that’ll get it for you.’ It’s an odd part of how this market has existed for a long time.”

He identifies open data and modern APIs as prerequisites: “How does an AI agent do anything unless it can understand everything?”

Gaiia's architecture is built around that premise, and the platform exposes data through open APIs, continuously updated in real time or pushed to a database on a regular cycle, so that any agent or external system can access it without a professional services engagement. 

Farnsworth also describes a Model Context Protocol (MCP) layer that allows AI agents from different systems to communicate with each other, which he considers a requirement for any platform that claims to be AI-native rather than AI-adjacent. 

The company also allows customers to bring their own AI models rather than locking operators to a single AI provider. 

"Being AI native requires you to be kind of open data," Farnsworth says. "It requires you to have modern APIs that can speak to other systems."

He describes his own use of agents built in Claude Cowork, drawing on Google Calendar, Gmail, Slack, a call recorder, and Gaiia’s customer relationship management (CRM) and states that any system an agent cannot access leaves the agent without context.

He applies the same test to vendor software. 

“If I’m asking AI to do certain things and it fundamentally can’t access the system, I immediately question that system,” he says.

Farnsworth identifies two additional constraints beyond data access.

The first is data quality. "You can move data from one system to another," he says, "but if you're moving bad data into a new system, it's still bad data."

The second is network complexity: for legacy operators running mixed plant, automating across that mix is significantly harder than automating a pure-fiber environment.

Farnsworth reports the same limits within Gaiia. The company uses AI extensively, and he states that queries against its own systems sometimes return answers that are “directionally true” but missing context until source priorities are specified.

Deployment Status and Operator Readiness

Reilly McClure, Senior Product Marketing Manager at Sitetracker, reports the same data-readiness gap from the deployment-management layer. 

Founded in 2013, Sitetracker reports more than 350 customers operating in over 100 countries. The company announced Scout, an agentic AI platform for critical infrastructure, in limited release on April 8, 2026, following customer pilots that began in late 2025; the product reached general availability this spring, and the company reports agents running in production at customer organizations.

McClure says the data-readiness question is one operators are still working out. 

"That's what a lot of operators are trying to figure out right now: 'How ready are we really?'" he says.

He also distinguishes data structure from data quality. 

"There's going to be things to clean up. Details left out or lost in translation, or sitting in somebody's email inbox somewhere."

Scout is designed as a three-layer platform. At the foundation is a data layer with connectors to external systems (Sitetracker, CRM, ERP, GIS, file storage) that provides access to operational knowledge across a customer's environment. In the middle is an intelligence layer powered by Compass, the platform's operational context engine. At the interface layer, Scout provides a conversational UI for natural-language data queries, as well as pre-built agents for common workflows. Customers can also build and maintain their own agents within the platform. 

"We're not going to be able to come up with every use case under the sun," McClure says. "Ultimately, we see customers building a lot of those and being able to maintain them on their own."

Model selection follows a similar logic. Both gaiia and Sitetracker allow customers to bring their own underlying LLMs rather than locking operators to a single provider. 

"Every other week, Anthropic releases something, and Claude is taking on a whole new industry," McClure notes. "You just can't limit yourself to one path right now."

McClure also describes a shift in how operators are approaching the technology over the past year.

“A year ago, there was some skepticism around AI, and certainly around the security angle,” he says. “Now the conversation is more pragmatic. It’s: ‘We’re going to do due diligence around security, we know we have to do something.”

Current production use cases are administrative: estimating permitting timelines, checking permit statuses across jurisdictions, and enriching incomplete job and work-order records. McClure characterizes the demand from operators and contractors as “How can I get more time back in my day?”

McClure distinguishes platform approaches from single-purpose tools. He observed numerous narrow AI products at Fiber Connect 2026, including computer vision tools for field photo validation, and considers them useful but limited.

“So many of the problems in fiber deployments happen upstream: in that planning phase, in those handoffs from engineering and design to construction, permitting, as-builts, and service turn-up.”

Deterministic Versus Probabilistic Execution

McClure raised a distinction between probabilistic and deterministic execution that has received limited attention in the industry.

Large language models are probabilistic. “Every time they interpret a prompt, they might look at all the same sources and give you a slightly different output,” McClure says. “And none of them are that great at math.”

Inconsistent outputs, he says, are unacceptable for operational workflows.

“When you think about dates and times and permit compliance and invoice reconciliation, you don’t want variable output for these things. When you ask a question, you want to get the same answer every time, not different results on different days. That’s really what creates trust.”

Farnsworth raised the same point: because AI probabilistically weights available data, input data quality determines output quality.

McClure is skeptical of operators who assume their existing enterprise AI licenses for Claude or ChatGPT are sufficient to build operational tools. 

"What we're hearing from a lot of operators is, 'Hey, our company bought Claude for all of us, or ChatGPT, or Microsoft Copilot, and we think we can go vibe code a CRM or some field production tool.' Okay,” he says. “Good luck."

Scout uses deterministic agent execution: knowledge graphs mapping an operator’s data model and business rules, with workflow steps and routing logic that are explicitly defined, conditional, and repeatable, and a record of every action taken. The mapping is performed by Compass, the platform’s operational context engine, which McClure describes as understanding “the complete ontology and relationships between objects” in a customer’s environment: the data model, terminology, and business rules teams use to manage work. 

The practical consequence is that a Scout customer can query their Sitetracker environment in plain language on day one, without manually providing operational context files or building integrations. Compass has already mapped the relationships. An operator asking about permitting status across multiple jurisdictions gets the same answer each time, derived from the same explicit workflow logic, with a documented record of what the agent checked and when.

Beyond efficiency, repeatability is the mechanism by which an operator can explain to a regulator, a funder, or an auditor what an agent did and why. McClure notes that most agentic tools currently on the market do not offer this: audit trails are being added after the fact rather than built into the architecture from the start.

The knowledge plane at the center of the Telstra–Ericsson collaboration appears to be driven by the same principle, giving agents a structured, explainable map of the network before trusting them to act on it.

For operators evaluating agentic products, the distinction is an important question: Is the agent’s behavior repeatable, and can the vendor produce a record of its actions?

Governance and Accountability

In BCG's research, 72% of organizations reported unmanaged AI-security risks, frequently tied to poor data quality and unclear oversight. The controls operators should expect from an agentic platform follow directly from that gap.

McClure is direct about the limits: "We're not going to suggest to anybody that you should automate all these processes and then take your lunch break." 

Scout addresses that through Watch, its dedicated governance and security layer: virtual private cloud infrastructure, customer data and personally identifiable information kept out of third-party LLM systems, and a full audit trail on every action. Customer data, McClure says, is never used to train or fine-tune Sitetracker's models—the company technically cannot access it.

For BEAD subgrantees, it is not optional. NTIA mandates that subgrantees maintain records and financial documents for five years after BEAD funds are expended or returned, stored wherever practical in open, digital formats. Semi-annual reporting obligations require project-level and location-level data, with subgrantees responsible for compiling supporting documentation on demand. An operator relying on agentic AI to manage permitting timelines, work orders, or deployment milestones without a system that can document what actions were taken, by which agent, and on what basis, is carrying federal compliance exposure with a tool that cannot account for itself.

Audit capability has compliance implications for operators answering to regulators, funders, and federal reporting requirements for BEAD-funded builds.

On autonomy, McClure reports more ambition among customers than anticipated: “I don’t know that we’ve found the line. We’ve been pleasantly surprised at how aspirational some of the goals are.” No customer is pursuing human-out-of-the-loop automation. The adoption model he describes is crawl-walk-run: a conversational data-query layer first, then agents on high-value, low-risk workflows.

Anis Khemakhem, Clearfield's Chief Commercial Officer, describes a similar posture on the manufacturing side. He encourages internal AI use with a stated condition: “Security is, to me, the biggest concern. We want to make sure it’s secure—especially because we’re publicly traded.” On expected value, he says, “It’s going to give us more efficiencies. It’s going to help us do our job a little bit better and faster. But I don’t view it as a cost-cutting tool. It’s an optimization tool.”

Outlook

The sources interviewed for this story identified the same starting point for operators: data that is structured, accessible, exportable, and clean enough to support automated action. Farnsworth states that access alone changes the calculation: “That right there makes AI possible.”

Stay Connected with ISE Magazine 

Subscribe to our newsletters and magazine for the latest telecom insights, explore the current issue for in-depth features and strategies, and register for upcoming webinars to learn directly from industry leaders.