Telco’s AI Problem Isn’t Tech, It’s the 40 Years Before It

I’ve spent the last decade helping enterprises make infrastructure programmable. When we started Itential in 2014, the bet was simple: networking was going to follow the same path as compute and storage. It was going to become software. And once it did, the teams that figured out programmability first would have an enormous operational advantage.

That bet held up. But I didn’t anticipate how long it would take, or how much friction would come from the industry itself.

Now AI is accelerating everything. And in telco specifically, the stakes are higher than in any other segment I work with. The networks are more complex, the change controls are stricter, the uptime expectations are unforgiving, and the gap between where most operators are today and where they need to be is wider than most people want to admit publicly.

Here’s what I actually think is going on.

Telco Has an Expertise Problem, Not a Technology Problem

The technology to automate telco networks has existed for years. REST APIs, NETCONF, YANG models, orchestration platforms. The tooling got better every year. The problem was never the tools.

The problem was that the knowledge required to use those tools lived in a small number of people, and those people were already running hard just to keep the network up. Automation was a project for when things calmed down. Things never calmed down.

What AI changes is the cost of encoding expertise. For years, taking what a field engineer knew and turning it into an automated workflow required developers, documentation cycles, and months of effort. Most organizations built small tiger teams to own that work. The tiger teams became bottlenecks. The expertise stayed locked.

Now, with agents that can reason through a methods of procedure document, a design spec, a runbook, the transfer cost drops dramatically. The expertise that used to live in one engineer’s head becomes something the whole organization can access and act on.

That’s not a marginal improvement. That’s a structural shift in how telco operations can scale.

The Architecture Question Everyone Is Getting Wrong

When people talk about AI in network operations, they usually mean one of two things: a chatbot that answers questions about the network, or a fully autonomous agent that makes changes without human involvement.

Neither is the right frame.

The chatbot is too limited. The fully autonomous agent is too risky – at least today – and the telco operators I know are right to be skeptical of it. You don’t deploy AI that makes unchecked decisions on infrastructure that carries voice, data, and emergency services for millions of people. That’s not a sensible operating model.

The right frame is layered. Keep deterministic execution where determinism is the right tool. Add reasoning where reasoning adds value. Build the governance layer before you expand what agents are allowed to do.

The happy path should be mostly deterministic. The long tail of edge cases, the unexpected states, the “device is offline but the ticket says it’s up” scenarios that used to require a developer to anticipate and code – that’s where reasoning earns its place.

This isn’t a compromised approach. It’s actually a better architecture than either extreme. And it’s the architecture that earns operator trust, which is the only thing that matters in production.

The Vendor Problem Nobody Wants to Talk About

Here’s the structural issue that comes up in every conversation I have with telco leaders.

For decades, the telco vendor ecosystem treated API access as intellectual property. Closed interfaces, proprietary data models, integrations that required their professional services teams. That was the business model. That was how they maintained leverage.

In an agentic world, that model is incompatible with a functioning AI strategy.

An AI agent cannot operate on a system it cannot reach.

If your orchestration layer can’t talk to your OSS, your BSS, your provisioning systems, your inventory databases through open and documented APIs, your agents are working with incomplete information at best and making decisions blind at worst.

The telcos moving fastest right now have one thing in common: they stopped tolerating vendor lock-in on interfaces. They put API accessibility in procurement requirements. They’re not asking for a vendor’s AI copilot. They’re asking whether their AI can reach the vendor’s system.

That’s the right question. Ask it about every system in your stack.

The Human Side of This Transition

The engineers who built telco networks over the last 30 years did extraordinary work under extraordinary constraints. They built systems that run at scale most industries will never need to match. They developed expertise that took years to accumulate.

Asking those teams to now operate at AI speed, with agents making recommendations and sometimes taking actions, is a meaningful change. It requires trust that gets built incrementally, not declared in a roadmap.

What I’ve seen work: start with agents that observe and recommend, not agents that act. Build the audit trail first so operators can see exactly what the agent did and why. Let the wins accumulate. The trust follows.

The teams that skip that process and go straight to autonomous action because the technology can technically support it are the ones that end up walking it back after an incident and setting the whole program back by a year.

What the Next Three Years Look Like

The operators who invest now in making their infrastructure programmable and API-accessible will have a compounding advantage. Every workflow they build, every domain of expertise they encode, every guardrail they put in place is infrastructure that agents can act on. The returns improve the more you invest.

The operators who wait for the technology to mature before starting to make progress will find the gap has widened in ways that are hard to close quickly.

This isn’t like waiting for a hardware cycle. The learning is the asset, and it accumulates with use.

Telco has always been an industry where infrastructure itself was the competitive advantage. That’s still true. The difference is that infrastructure now includes the automation layer, the data layer, and the governance layer that lets AI act on it safely. Building those three layers is the work.

The good news is that telco teams are not starting from zero. They have the operational knowledge. They have the domain expertise. In many cases they have the automation foundation. What they need now is the architecture that lets AI reasoning act on what they’ve already built.

That’s the work. And it’s more achievable than it looks from the outside.

Watch the Telco in 20 Episode

I covered a lot of this recently on the Telco in 20 podcast with host Danielle Rios – including how Lumen scaled from 16 to 350+ automated workflows, how MCP changes the agent context problem, and what telcos should be demanding from their vendors.

Listen to the full episode here or below.