With AI, operators can move from simply seeing the network to working with it. As network data becomes accessible through open interfaces, they can build their own AI-assisted workflows. But doing that successfully requires more than adding an AI model to existing tools. It requires the right data, context, specifications and controls.
At Connected Britain earlier this year, Luke Saunders, principal network engineer at FibreNest, and Cameron Warren, director of engineering at HBNL, joined our Ryan McCowan for a Keynote Stage session on practical AI implementation. Their discussion highlighted several lessons for operators putting AI to work in their networks.
From isolated data to intelligent workflows
Network data has traditionally been divided across technical domains. Access, Wi-Fi, optical transport and other parts of the infrastructure can each generate their own telemetry and expose it through different management systems.
The first step toward more intelligent operations is bringing that information together. Cross-domain platforms can present data from different sources in a common view, helping engineers identify relationships and investigate issues.
AI adds another layer. Models can analyze information across domains, support root-cause analysis and provide guidance on what to do next for customer care, technical support or network operations.
The next step is making that information available to AI agents in a structured way. Operators don’t need their vendor to build every tool. They do need structured data supplied via APIs, knowledge bases, configuration guides, agent skills and protocols like the Model Context Protocol (MCP) to allow these resources to be discovered and used. That gives operators the building blocks to develop custom AI tools around their own operational requirements using the on-demand expertise of their vendors and ICT partners.
Better AI requires better context
Simply giving an AI model access to network data doesn’t automatically produce reliable results.
On the Connected Britain stage, FibreNest’s Luke Saunders described agents working through hundreds of pages of RFCs and standards in minutes, work that would normally take hours of reading.
But he also provided a caveat. An unconstrained model can make assumptions, interpret a task differently from one attempt to the next or “hallucinate” answers. Without clearly defined context, AI can accelerate the work without necessarily improving the outcome. Luke called it “human error at computer speed.”
FibreNest’s experience points to three practical principles for making AI outputs more reliable:
- Constrain task sizes: Keep tasks small enough to maintain a focused context window for the model to reduce the risk of degraded outputs.
- Leverage industry standards: Align prompts closely with the Broadband Forum, ITU and TM Forum so agents ground their logic in existing frameworks rather than inventing rules.
- Provide explicit parameters: Give agents structured information, defined parameters and explicit rules rather than expecting the model to infer everything from an open-ended prompt.
From prompting to spec-driven development
As AI-assisted engineering matures, the emphasis can move from asking AI how to do something to defining what needs to be done. This is the basis of spec-driven development.
Rather than giving an agent a loosely defined task and allowing it to determine the implementation, engineers provide detailed requirements, parameters and rules. The agent then uses those specifications alongside relevant technical documentation to produce a reliable output.
FibreNest now works this way for software and network designs, with agents reading vendor documentation to produce high-level designs – aligned with their documented best practice – in a matter of hours.
This changes the role of the engineer: expertise increasingly lies in defining the requirements, constraints and validation processes that govern what the AI produces. As Luke put it, the work has shifted from how to build to what to build.
Automating installation assessment
Installation jobs often require engineers or scheduling teams to assess factors such as complexity, estimated duration and equipment requirements.
Cameron Warren of HBNL described exactly this challenge. His scheduling team used to spend around 10 minutes per installation job judging duration and difficulty. That assessment is now fully automated. A scoring matrix generates a PDF report for every property in the ready-for-service database, with expected duration and complexity.
This saves time, supports more consistent job evaluations and frees teams to focus on routing installation crews efficiently.
The real value of AI in network operations comes from connecting it to the right data, giving it clear context and defining what it can safely do.Connecting customer experience with network reality
There can be a significant difference between how a network problem is experienced by a customer and what network telemetry actually shows.
HBNL tackles this by combining call data from its CRM platform, ticket data and network data from our Mosaic One Fabric platform. This lets the team compare what the customer experienced with what the agent recorded and what actually occurred on the network, helping to distinguish genuine network faults from issues with other causes.
Because AI handles data integration that was previously manual, teams can analyze the results more quickly, helping them address root causes sooner and improve customer communications.
Testing with digital twins
Digital twins provide virtual representations of network environments against which software and integrations can be tested. Luke gave a recent example: a new device claimed to behave one way, but its data model contained small differences that only showed up during integration. Catching that in a twin is clearly preferable to catching it in production.
Combining digital twins with AI-assisted development creates a feedback loop in which agents help produce software or configurations, test them against a representative environment and surface issues before deployment.
The importance of an open data layer
None of these use cases depend on AI operating in isolation. The common requirement is access: AI becomes useful when it can access network telemetry, technical documentation, customer systems and operational databases.
FibreNest’s next step is building its own fabric that brings internal knowledge, ticketing, billing and subscription data alongside network state, enabling stakeholders to query why something behaves in a certain way and receive an answer in seconds or minutes instead of hours.
This makes underlying data architecture critical. Open APIs, structured data and standardized interfaces give operators the flexibility to build custom workflows. Normalizing data from disparate systems remains a hurdle, but it’s also the kind of repetitive, structured work AI can help address.
From visibility to action
The evolution of network operations isn’t simply about adding AI to existing tools. The more significant change is the move from systems that primarily show engineers what’s happening toward systems that can help determine what events mean and what should happen next. They can then take action in a safe and accountable way, or notify the responsible team where risk demands a human in the decision loop.
That requires accessible data, well-defined context, clear specifications and appropriate controls. The operators that get the most value from AI won’t necessarily be those that deploy the most sophisticated models. They’ll be those who give AI the right information, connect it to the right systems, and clearly define what it’s expected and permitted to do.
Listen to the full Connected Britain keynote session on our podcast.