- Agentic AI is live in telco production, but most deployments are still reactive, not proactive.
- Customer care use cases are maturing fast; network SLA automation is where the real infrastructure gap sits.
- Only two of 64 demos at MWC 2026 tracked by STL Partners delivered true sub-second real-time action.
- Closing the loop requires sub-10ms data currency, ACID-compliant transactions, and predictive escalation built into the data layer.
TL;DR
Walking the floors of MWC this year, one thing was immediately clear to me: Agentic AI has officially moved out of the experimental labs and into active production environments.
But as telco technology leaders shift from reactive tools to autonomous systems, I could see a critical bottleneck emerging. The narrative is no longer about what the LLMs can comprehend, but rather a dual infrastructure challenge: achieving High Data Currency alongside Latency-optimized decisioning.
The Customer Care “Comfort Zone”
It is obvious to see the immediate applicability of agentic AI to customer care. Currently, AI agents are successfully handling complex queries like, “Why is my bill so high?” or “What happened to my balance?” just as well as human agents.
This domain is maturing rapidly because it operates safely on historical event data.
- The Future: We are quickly moving toward agentic account management, where AI doesn’t just pass recommendations to a human, but actively applies account credits or modifies plans in real time to keep a frustrated customer happy.
- The Logic: As long as the agent has visibility into the most recent subscriber events and CRM history before the customer asks a question, it can provide an accurate root-cause analysis.
The Network SLA Reality Check: The Reactive Trap
When we pivot from billing to operations, automated network SLA management requires a fundamentally different architecture. At MWC 2026, most operators showcasing agentic AI network management solutions were running hundreds of agents in distinct layers:
- Ingestion Layer: Gathering real-time telemetry from probes and cell utilization stats.
- Context Layer: Ingesting user experience and historical performance data.
- Orchestration Layer: High-level agents pulling the lot together to perform Root Cause Analysis (RCA).
The catch? It is mostly reactive. For example, T-Mobile demonstrated a highly impressive solution that predicts network congestion on specific interconnects when Bayern Munich plays certain teams in the Champions League. They know in advance to increase capacity, which is a massive win, but it is still fundamentally a response to a known, scheduled event.
The Data From the Floor
A recent analysis of 64 major industry announcements by STL Partners perfectly highlights this current technological divide:
- The Real-Time Gap: Out of all the demos tracked by STL, only two delivered true sub-second, real-time action—and both were focused on in-call fraud detection, rather than network operations.
- Massive Efficiency Gains: Deutsche Telekom’s RAN Guardian Agent (utilizing Google Gemini models) reported a remarkable 95% reduction in network management time during peak traffic events.
The Holy Grail: Proactive, Closed-Loop Networks
To move from “AI-assisted” to “AI-led” operations and move the needle for CTOs, we need to bridge the trust gap. We must move beyond having an AI merely report an anomaly before customers complain, and toward autonomous capacity provisioning.
The barrier to achieving this isn’t the AI’s intelligence; it is the underlying data infrastructure. To allow an AI agent to dynamically provision network resources without a “human-in-the-loop,” the data layer must guarantee:
- Sub 10ms Data Currency: Agents cannot make safe, automated network changes based on data that is already seconds old.
- Transactional Integrity: We must ensure that network-altering actions are entirely ACID-compliant and reversible.
- Predictive Escalation: Systems must be designed where agents autonomously solve the vast majority of anomalies, only escalating complex edge cases to humans with full, real-time context attached.
We might be a little way off from completely trusting AI agents to alter networks without human authorization. However, building the data architecture to support that future is the necessary first step.
Dive Deeper into the Data
How are industry leaders planning to overcome inference latency, resolve data currency issues, and reduce “human-in-the-loop” dependency?
I highly recommend reading the latest research to see the roadmap to self-governing networks.
Read the Full STL Partners Report on Agentic AI Maturity.
What is data currency in agentic AI?
Data currency refers to how fresh the data is that an AI agent acts on. In network operations, agents making autonomous decisions need data that is milliseconds old, not seconds.
Why is agentic AI in telco still mostly reactive?
Most telco AI deployments rely on data layers that cannot deliver sub-second freshness at scale, forcing agents to work from slightly stale context and limiting them to responding to known, scheduled events rather than acting autonomously in real time.
What is the difference between AI-assisted and AI-led network operations?
AI-assisted operations use agents to surface recommendations or flag anomalies for human review. AI-led operations allow agents to take autonomous action, such as provisioning capacity or modifying network configurations, without requiring a human in the loop. The data layer requirements for the two are fundamentally different.
What does ACID compliance mean in the context of agentic AI?
ACID compliance means that network-altering actions taken by an AI agent are atomic, consistent, isolated, and durable. In practice, it ensures that any automated change to network configuration is fully reversible and does not leave the system in a partial or corrupted state if something goes wrong mid-execution.
About Author
Featured Resources
