As AI tears through the communications industry, telco operators face a stark reality: they cannot experiment at hyperscaler speed without risking real-world consequences. Networks are not sandboxes. They carry emergency calls, financial transactions and the digital lifeblood of entire communities.
In a conversation with Joshua Turiano, chief innovation and AI officer at Blue Stream Fiber, a broadband provider operating in Florida and Texas, he outlines a pragmatic deployment model rooted in operational discipline rather than hype. His philosophy is simple: “Replace the task, not the job.”
Turiano argues that AI should accelerate human decision-making, not remove it from the loop. Deterministic systems remain critical for infrastructure control, while large language models (LLMs) are best applied to ambiguous, text-heavy workflows such as ticket review and data interpretation, he says. Turiano also warns that AI introduces new attack surfaces if not carefully constrained behind structured API layers.
Most importantly, he reframes the industry debate: The challenge facing telecom operators, he argues, is not a lack of AI capability — it is a translation gap between impressive demonstrations and scalable, real-world implementation.
Keep reading for an edited version of my conversation with Turiano.
Interview with Joshua Turiano, Chief Innovation and AI Officer at Blue Stream Fiber
Steve Saunders: You’re responsible for the AI strategy at Blue Stream Fiber so you’re the guy everybody wants to interrogate.
John Turiano: I feel like the dog that caught the car.
Saunders: What have you learned moving from theory to deploying AI on live networks?
Turiano: We started our AI journey in 2024, and it was really just experimentation. Some of the features that we have now, we couldn’t have predicted would have been here six months ago. I had a vision in 2024 that we deployed in 2025. What I’m envisioning right now isn’t what’s going to be there in less than three months. By the time I can deploy something, someone’s going to come out with something that may completely blow that away.
It’s really hard to predict what the next six months are going to look like. It’s hard to think more than a year in advance.
Saunders: How do you decide whether to use hyperscaler orchestration or telecom vendor orchestration?
Turiano: All the different network vendors now are bringing AI into their platforms and they’re all creating their own little micro AI agents. Whether you go to Cisco or you go to Juniper or Ciena or Calix or Nokia, they have these tiny little agents.
I don’t necessarily want five tiny different little agents. I want all these agents feeding into a single pool of data so that our [management] layer can give broad answers on any of that equipment in less than 10 seconds. You see this divergence again. Eventually, it’s going to come back together.
Saunders: Is there a standard for trading orchestration information between telecom systems and AI agents?
Turiano: Not today, not in the way that I would like to see it. You have to normalize all of the data into certain lakes, give your model context protocol instructions on how to read and what to read and why it’s important. There’s still a lot of human intervention on building all of those LEGO blocks and putting the puzzle together.
Saunders: What is AI’s function inside a telco network? Is it autonomy? Is it augmentation?
Turiano: My strategy at Blue Stream Fiber is human first. Replace the task, not the job. Augment your people to give them access to data faster so that they can make the important strategic decisions, versus just letting an autonomous AI run the show. There are certain tasks AI can do all day, every day — checking a work order, looking at light levels, alerting. But when it comes to human touch and tone, I would much rather prefer to augment the people that we have than replace them with just an AI. You have to keep that human in the loop.
Saunders: Where do probabilistic LLMs fit versus deterministic systems?
Turiano: Where the LLMs really shine is when you hand it fuzzy data. Reviewing tickets. Deciding where something should go based on a description. Super critical items probably need a more traditional ML [machine learning] Ops, traditional robot software that just runs, looks at data, and makes a decision, but hands it over to someone to alert upon. LLMs work great for ambiguity. But when you try to push them into tone detection or critical control, they’re not quite there.
Saunders: What about security once AI is introduced into the network?
Turiano: It does become another attack plane. Anytime you’re exposing a chatbot to the public web or connecting it to data sources, because of that pleasing bias that exists inside these models, you run a risk of it going off the rails. That’s why we don’t just give AI full database access. We give it an API layer. We define what it can read and what it can use. If you structure that correctly, it can’t hallucinate beyond what you’ve defined. But I think it’s dangerous to let an AI autonomously sit in your network core. It’s better reacting and elevating predictive analysis than taking direct actions without constraint.
Saunders: Is innovation moving faster than telcos can operationalize it?
Turiano: In some regards, yes, and in some regards, no. The gap that I see across the industry isn’t so much a technology gap. It’s a translation gap.
We see these demos. They look great. But somewhere between flying the plane and landing it, that signal kind of gets lost on AI.
Sometimes people get paralyzed because no one showed them a version of AI that really works for them at their scale, with a lean team, with an existing tech stack. Maybe it’s not as photogenic. But it does actually move the needle without having to be at the bleeding edge of what’s going on today.