How the Proximus Group Is Piloting Agentic AI to Speed Network Incident Resolution
Download PDF
Projected Outcomes
Minutes
From hours of manual work to minutes, often under a minute for simpler cases
1,000+
System messages exchanged per incident, once reviewed by hand
2–4 hrs
Prior resolution time for a routine incident, up to days for complex ones
About Proximus Group
The Proximus Group is Belgium's largest telecommunications and digital services group, serving consumer, business, and wholesale customers across the national network. Majority-owned by the Belgian government, it employs around 13,000 people. Its Network IT division runs the operations support systems that activate and maintain services—from internet subscriptions to mobile plans—across a complex, high-volume web of interconnected applications.
The Shift
Resolving a routine service issue once meant manually filtering and analyzing 1,000+ system messages across specialized applications—a process that typically took two to four hours of work and sometimes stretched into days. In early testing, an AI agent orchestrated by Camunda performed the same investigative legwork in just a few minutes, with simpler cases often resolved in under a minute.
Before
- Engineers manually filtered through 1,000+ system messages per incident, often taking two to four hours for a routine issue and days for complex ones.
- Misrouted tickets bounced between application teams, with each handoff requiring fresh analysis before the right owner even started work.
- Institutional knowledge about resolving specific issues lived in individual experts' heads, with no structured way to capture or reuse it.
After: In Testing
- An AI agent orchestrated by Camunda retrieves logs, queries the right systems, and proposes a resolution, cutting the investigative legwork from hours to minutes or less.
- The agent's guardrails route each incident to the correct systems and data from the start, with the aim of reducing unnecessary handoffs between teams.
- When the agent needs help, it routes a question to a human expert, then captures the answer for potential reuse when similar issues arise.
Why It Mattered
The Proximus Group's Network IT division keeps operations support systems moving so customers can get new services activated without delay. For Belgium's national operator, service reliability directly shapes brand trust in a competitive market.
A single routine request, like a new internet subscription, can trigger over a thousand messages across the microservice architecture. That's normal engineering practice, but it makes manual troubleshooting slow and highly specialized—engineers traced logs by hand and pulled in colleagues from other teams before a root cause emerged.
The Proximus Group was already running Camunda in production to orchestrate its network activation processes. Extending that same orchestration layer to coordinate an AI agent was a natural next step.
What They Did
Rather than treat AI as a bolt-on chatbot, the Network IT team built an agent that plugs into the same Camunda process already orchestrating its operations, so deterministic steps and AI reasoning can run side by side within a single model. Camunda's ability to blend fully automated steps with AI-driven decisions in a single process was the deciding factor: the team kept the reliable, straight-through parts of incident handling exactly as they were and added agent reasoning only where judgment was genuinely needed.
The agent pairs a large language model with a defined set of tools: connectors to the Proximus Group's provisioning systems, a knowledge base built on stable documentation and expert-captured feedback, and human-in-the-loop tasks for cases it cannot resolve on its own. A decision table constrains which systems the agent can query, and a fixed query template prevents it from acting outside its defined scope—guardrails the team called essential for trusting the agent in a live network. Structured logging captures the agent's reasoning at each step, giving engineers a clear trail of what it decided and why.
“At the beginning, our operational colleagues made our agent fail. In the end, we realized they are exactly the people who will make it succeed.”
What Changed
The Proximus Group's AI agent is at the proof-of-concept stage, running in a trial environment rather than in production. Even at this stage, the early proof points show what one orchestrated process makes possible:
- Faster investigation, even at prototype stage. In demo runs, the agent moved from an incident description to a proposed root cause in under a minute, against a baseline of two to four hours.
- A knowledge base that learns from experts. When the agent hit a gap it could not resolve, it asked a human expert, then stored the answer for reuse when similar incidents arise.
- Deterministic control over what the AI can touch. A decision table and fixed query templates are designed to restrict the agent to approved systems and defined query patterns—guardrails that the team called essential to trusting it in a live network.
- A clearer audit trail. Structured logging of the agent's reasoning gave engineers visibility into every decision, which the team highlighted as important for troubleshooting and trust.
What's Next
The Proximus Group is testing the agent against real-world incident types in a trial environment to validate the approach before any production rollout decision. The team is considering widening connector coverage and refining the agent's prompts before scaling to more incident categories.
The team is also enriching the knowledge base with context from past conversations and exploring more advanced retrieval. Taken together, this work can help assess whether the approach is relevant to other areas of network operations.
