Telco
Agentic AI for telecom operators — network assurance, SLA breach prediction and customer care agents that resolve rather than route, deployed in EU infrastructure.
What makes telecom hard for AI
Scale changes the economics
At tens of millions of events per day, per-transaction inference cost decides whether a design is viable. Model selection and caching strategy are architectural decisions, not tuning.
Care systems are deeply fragmented
Billing, provisioning, CRM and network systems rarely share identifiers cleanly. An agent that cannot resolve a customer across them cannot resolve their problem.
Churn is decided in the interaction
Care quality directly drives retention, so degradation is expensive and shows up in revenue before it shows up in monitoring.
Where the value concentrates
- Care that acts. Agents that read the account, diagnose and execute the fix rather than routing a ticket onward.
- SLA breach prediction. Forecasting breaches and escalating resources before penalties trigger.
- Network assurance. Correlating alarms into probable cause and reducing alert volume.
- Field operations. Predictive maintenance and automated work-order dispatch.
What the economics look like
Cost per contact is the obvious measure; first-contact resolution is the more meaningful one, because it moves both cost and churn together. SLA penalty avoidance is often the fastest single payback in enterprise segments.
How we build it
We start with identity resolution across systems, because everything downstream depends on it, then layer agents that can take governed action rather than only summarise.
Related work
Starting smaller
Regional operators and MVNOs usually begin with care summarisation and knowledge retrieval before granting agents the ability to act on accounts.
Questions we get from operators
Is per-interaction inference cost viable at our volume?
Only if it is designed for. Route by task rather than sending everything to a frontier model, cache aggressively, and keep context windows tight. Cost per interaction is an architectural constraint at telco scale, and retrofitting it usually means rebuilding.
Our systems do not share a customer identifier. Where do we start?
There. Identity resolution across billing, CRM, provisioning and network systems is the prerequisite for everything else, and it delivers standalone value in reporting even before agents are deployed.
Can agents make changes to accounts?
Yes, within policy and with an audit trail. Start read-only, measure resolution quality, then extend authority to the action types where the model performs reliably. Granting broad write access on day one is how trust gets lost early.
How do we protect service levels during incidents?
Design for the spike, not the average. Capacity planning, graceful degradation and a defined fallback to human handling matter more than peak accuracy, because the failure that hurts is the one that happens under load.