Data curation: the key to intelligent use of AI in network operations
5G networks are scaling fast, and with 3GPP Release 18 and 5G Advanced on the horizon, operations need more than dashboards. They need clean, timely data that AI can act on. Industry forecasts point to strong AIOps growth through 2029, and the winners will be the teams that curate data at the source and feed models with precision.
A network with ten million subscribers can generate up to nine petabytes of data per day. That flood overwhelms traditional observability, especially when the data is late, unstandardized, and buried in data lakes. Raw packets and CDRs are granular, noisy, and expensive to prep for AI. The insight you get is limited by the quality of the data you send in.
Fix the data at the source
Smart monitoring, curation, and pipelining are the foundation for cost-effective AIOps across RAN, Core, Transport, and MEC. Build pipelines that normalize, enrich, label, and anonymize data where it is generated. That cuts noise, increases fidelity, and keeps compliance teams happy.
Telecom domain expertise matters. You need context to turn mixed data sources into clear, actionable signals. Define what data is relevant, collect it consistently, then clean, validate, and label it. Keep humans in the loop to design models and deliver high-quality feedback when AI gets it wrong.
Tokenization: from raw to curated
A single raw 5G event can carry ~180 tokens. Curation can reduce that to ~25 tokens-an ~85% cut in compute and GPU cost, especially in public clouds like AWS Bedrock. Precision goes up, storage goes down, and AIOps becomes financially viable.
Once curated, merge signals with subscriber demographics, cloud infrastructure metrics, geospatial or environmental data, and even aggregated social analytics. With the right domain intelligence, these combined features drive better detection, prioritization, and automation.
Packet-level precision
Curate for the use case. Deep Packet Inspection (DPI) shows what moved, when it moved, and how the stack responded. Enrich with control plane metadata and identifiers such as IMSI/SUPI to produce metrics per cell, slice, handset, or subscriber.
This level of precision lets you train AI on real behavior tied to real customers. Curated feeds are often one hundredth the size of the raw data while keeping maximum analytical value. You get sharper AIOps signals that improve NPS, strengthen SLA management, and expose new monetization angles.
Driving down AI costs with configurable streams
NETSCOUT Omnis AI Streamer delivers curated, high-fidelity metadata from packet flows. It correlates observability trends, automates analysis, and flags hidden issues and risks before they hit customers. In production, teams have seen up to 93% reductions in data volume, lower GPU memory and processing time, and higher throughput from fewer GPU instances.
Use cases include network optimization, predictive maintenance, real-time slicing analytics, and digital twin scenarios. Flexible feed configuration lets you define a playbook of feeds, intervals, metrics, dimensions, and filters so only necessary curated data reaches AIOps engines.
Example: aggregate QUIC transport latency to monitor a premium YouTube slice. If performance drops, trace to the specific cell or node and generate a focused dataset for rapid triage. User Plane Data-TEID, QoS Flow ID, IP addresses, latency, app signatures-feeds traffic analysis, SLA breach detection, QoE estimation, and app-level performance monitoring.
What operations teams should do now
- Set clear outcomes: SLA adherence, churn reduction, MTTR, GPU hours per insight, and storage spend.
- Map sources: packet sensors (DPI), control plane, RAN counters, MEC, transport, device telemetry, and customer context.
- Build at-source pipelines: normalize schemas, enrich with domain context, label events, and anonymize sensitive fields.
- Reduce tokens: keep the smallest feature set needed per use case to cut compute and noise.
- Keep humans in the loop: establish labeling standards and feedback to correct model errors.
- Measure impact: A/B key automations and track NPS, SLA breaches, MTTR, false positives, and cloud cost.
- Create feed playbooks: schedules, critical metrics, dimensions, and filters for each AIOps workflow.
- Control cost: cap egress and GPU budgets, and choose execution venues wisely (on-prem, edge, or public cloud).
Aligned with industry moves
Data curation supports the shift to 3GPP Release 18 capabilities and the push toward autonomous operations. It also fits with the TM Forum Autonomous Networks initiative, where clear data contracts and closed-loop control are essential.
Bottom line
High-quality, high-value, low-volume curated data is the decisive input for AIOps. Do the heavy lifting at the source, deliver gold-standard feeds, and your AI will act faster, cost less, and improve customer experience. That's how operators move from firefighting to predictable performance and new revenue.
If your team is upskilling for AIOps, explore practical pathways on Complete AI Training.
