Network automation is having an AI moment. Don’t forget the plumbing.
What does it really take to scale AI-driven network automation? Gabriele Di Piazza recaps his recent conversation with Sebastian Barros, exploring why trusted data, agent orchestration, and measurable business outcomes matter more than simply deploying more AI agents.
Communications service providers (CSPs) don’t have a shortage of AI ideas. The harder problem is turning those ideas into automation that can safely operate across a live, multi-vendor network and produce an outcome worth paying for.
That distinction was at the heart of a recent conversation I had with Sebastian Barros, my friend and former Google colleague.

Sebastian and I have lived through several telecom technology cycles together, which means we share both an enthusiasm for what’s possible and a healthy skepticism for any technology that promises to magically erase decades of complexity.
AI is no exception.
As CSPs move from building agents to operationalizing them, three questions are becoming much more important:
- Can AI understand the relationships hidden in fragmented network data?
- How will CSPs orchestrate and govern agents across vendors and domains?
- And where does the ROI actually come from?
Here’s where we landed.
Before agents, get the data right
Telecom has been trying to solve the data problem for years. Data lakes, warehouses, and aggregation platforms have all played a role. But AI changes what we need from that data.
Collecting it isn’t enough.
Automating bad data faster is not autonomous operations.
For automation to make increasingly autonomous decisions, it needs operational context. It needs to understand the relationships between sites, physical devices, services, topology, capacity, SLAs, and operational state.
That’s where data fabric and ontology become so important.
A normalized, structured representation of the network turns raw data into knowledge that automation can reason over. It creates a foundation not just for AI agents, but for deterministic automation, applications, and workflows as well.
It can also affect the economics of AI. If a model has to repeatedly search and reason across large volumes of poorly structured data to find the context it needs, token and infrastructure consumption can grow quickly. Good automation should deliver the right context to the right model or workflow at the right time.
As Sebastian joked, telcos have historically thrown people at problems. We probably don’t want to replace that habit by simply throwing agents at them.
Automating bad data faster is not autonomous operations.
Get ready for an agent orchestration problem
Once the data foundation is in place, another issue appears.
CSP networks are inherently multi-vendor and multi-domain. RAN, IP, optical, OSS, BSS, IT, and security environments will increasingly incorporate their own agents.
A single agent solving a single problem is relatively straightforward. Thousands of agents operating across vendors, systems, and domains are something else entirely.
Sebastian described the future challenge as needing an “HR department for agents.” It sounds funny, but the analogy works.
Agents need identities, permissions, policies, lifecycle management, security, and observability. They also need coordination so that one agent’s action doesn’t conflict with another’s.
We have seen a version of this before with APIs. As APIs proliferated, organizations needed gateways, catalogs, security, and lifecycle management. AI agents could create a similar requirement for agentic orchestration and governance.
That doesn’t mean every automation problem must suddenly become cross-domain. Quite the opposite. If a problem can be efficiently solved within one domain, solve it there.
But customer and service outcomes increasingly span multiple domains. Fault isolation, for example, may require understanding dependencies across RAN, IP, and optical infrastructure. In those situations, optimizing each silo independently won’t optimize the outcome.
The goal shouldn’t be more agents. It should be agents working together, with the right context and guardrails, to deliver the intended result.
Don’t measure ROI one agent at a time
Then comes the inevitable and entirely appropriate question: What’s the return?
“Look, we built an agent” is not a business case.
The temptation is to calculate the ROI of each AI agent or use case independently. But the real value often emerges across an end-to-end process.
AI-assisted equipment modeling, infrastructure planning, and root-cause analysis can target operational efficiencies. Faster service provisioning can address something different: speed and time to revenue.
So rather than asking only how much an individual agent saves, ask what happens to the overall process.
- Can we reduce manual effort?
- Can we resolve problems faster?
- Can we avoid unnecessary field activity?
- Can we shorten the journey from customer order to an operational service?
- Can we turn a process measured in weeks into one measured in minutes?
That is where the ROI conversation becomes far more interesting.
Automation can be a growth story, too
This brings us to one of my favorite parts of the discussion.
Network automation has traditionally been framed as an OpEx play. And there is meaningful value in taking cost and complexity out of operations. But CSPs shouldn’t stop there.
Automation can also become an engine for differentiation.
We’re already seeing pressure to deliver and provision services faster, particularly in B2B. CSPs aren’t only competing with other CSPs. They may also find themselves competing for enterprise business with hyperscalers and newer, more nimble providers.
That makes the experience around the network increasingly important.
Imagine services that can be understood, configured, and provisioned dynamically. Network-as-a-Service (NaaS) and more automated B2B experiences become considerably more powerful when the network underneath can respond quickly to customer intent.
And there’s another interesting possibility ahead: the customer may increasingly be an AI agent itself.
Sebastian and I had some fun with that one. But it isn’t difficult to imagine agents eventually discovering, comparing or requesting connectivity services on behalf of people or enterprises. That raises the bar again for how consumable and automated CSP services need to become.
Scale the outcome, not the AI
At the end of our conversation, Sebastian asked what I would tell a CTO trying to move from successful AI experiments to automation at scale.
I’d start with three questions:
- Where is AI already creating meaningful value?
- What prevents those successes from being replicated across the organization?
- Do we have the trusted data and governance required to scale them?
Because ultimately, CSPs don’t have a competition to deploy the most AI agents.
The opportunity is to create an environment where AI and automation can operate at scale, with trusted context, appropriate governance, and measurable business value.
Get that foundation right, and “Mission Possible” starts looking a lot more possible.
Want to hear Sebastian and me debate the rest, including legacy modernization, humans in/on the loop, and the future of autonomous operations? Watch the full LinkedIn Live session here.