Any carrier can say it uses AI. These are the rules ours answers to, written down so you can hold us to them. They apply to every model, every threshold and every automated decision in our network, today and as the stack grows.
Our AI runs under eight principles: intelligence sits in decision paths, learns from real traffic, stays accountable to named engineers, gets measured honestly, admits its limits, protects subscriber data, explains itself to customers, and is treated as attack surface. AI-native means the principles are load-bearing, not decorative.
Models feed routing, fraud scoring and quality supervision. If the output only ever reaches a dashboard or a slide, it is not AI-native, it is analytics. Every model we deploy consumes its own output in production.
Our systems learn from production wholesale traffic, where a routing decision proves itself within the hour and a fraud hypothesis lives or dies in days of CDRs. No lab data, no demo benchmarks. The feedback loop is the asset.
Machines propose, engineers confirm. A named human owns the outcome of every automated decision, and the escalation path to that human is defined before the model goes live, not after it goes wrong.
False-positive rates, detection rates and time-to-escalation are measured and tracked for every detector. A fraud score without a measured error rate is an opinion. Ours publish their numbers internally, and we publish the direction of travel.
We say where AI does not help: interconnect negotiations, novel fraud investigations, strategy without feedback loops. Machines run where the loop exists, engineers where it does not. No AI theatre.
We do not train on subscriber communication content. Traffic analytics run on metadata patterns under GDPR-aligned controls, with minimisation and retention limits designed in, not bolted on.
When our systems affect your traffic, a route change, a fraud flag, a quality alert, you can ask why and get a human answer with the evidence. Automated decisions are never a black box you just have to trust.
Models, pipelines and thresholds are attack surface. They receive the same hardening, monitoring and change control as the network itself, because a manipulated detector is worse than no detector.
These principles govern how we build. This is what AI-native means, and this is how we run.
