The benefits of curated data: three use cases For CSPs

Communications service providers (CSPs) are sitting on a resource that should be their greatest competitive advantage: data. A 5G Standalone network with 10 million subscribers generates approximately nine petabytes of raw user plane data per day – close to one million transactions per second. This volume, common across mobile, fixed, and low earth orbit networks, represents the essential fuel for AIOps. 

However, turning this data into actionable insights represents a significant operational challenge to any CSP. Not only is it a tremendous amount of data to collect, but it also must undergo a process of adding contextual intelligence, starting at the data acquisition layer. This handles data-schema normalisation, KPI creation and cross-plane correlation. It also provides metrics by subscriber, application, handset, market and 5G Slice, among others.  

The alternative to this – feeding raw data directly into AI models – is not a path to intelligence, as it results in inefficient performance and high economic costs due to wasted compute, tokenisation costs and models prone to hallucination.  For CSPs, the key to unlocking genuine network automation lies in transforming raw data into a curated, high-octane fuel. 

The Raw Data Problem 

Scale and signal-to-noise 

Raw 3GPP packet data is fundamentally unsuitable for direct AI consumption. At network scale, the volume overwhelms models; critical alerts are buried beneath irrelevant events, slowing response times and increasing the risk of subscriber impact.  According to a recent IDC forecast, 78 percent of stored data is unstructured, meaning it lacks a predefined format or organisation. Feeding this into an AIOps engine forces platforms to waste up to waste up to 50 percent of compute on basic data preparation – costing organisations an average of $12.9 million each year in poor data quality. 

Why traditional sources lie by omission 

A common misconception is that CSPs already have sufficient visibility through alarms, logs, and network equipment manufacturer (NEM) event streams. They do not. These sources share a structural flaw: they only report what they were explicitly designed to report. Everything else is silence – and that silence is where novel failure modes, slow-burn degradations and cross-domain anomalies live. 

Firstly, alarms are threshold-driven and binary. The gradual erosion of session quality that precedes a threshold breach produces no alarm and leaves no trace. 

Secondly, logs are domain-scoped and retrospective. A core network log has no visibility into the radio access network (RAN) during the same session; cross-domain correlation requires manual effort and often fails when timestamps or formats differ. 

Finally, NEM event streams reflect the vendor’s interpretation of network state. In multi-vendor networks, these streams are heterogeneous and always incomplete from the perspective of end-to-end session behaviour. Events outside a vendor’s taxonomy are never reported.  

A model trained on these sources inherits every blind spot and every vendor-imposed category. Packet data is different: every transaction is captured regardless of whether any element raised an alarm or wrote a log. That continuity is what makes packet data the ground truth. 

What data curation actually means 

Data curation transforms raw packet-level traffic into structured, contextualised information that AI systems can reason about. This is not filtering or compression – it is adding contextual intelligence to the data itself. Effective curation tracks every subscriber across their full session lifecycle, correlating location, timing, session flow, network domain (RAN, core, or transport), and KPIs that characterise normal and abnormal behaviour.  

This enrichment creates an ontology – a structured representation of network reality that gives AI models the situational awareness they need to draw accurate conclusions. The key is to curate at the source via deep packet inspection (DPI), producing a compact, high-fidelity dataset without requiring raw storage in a data lake. 

When this dataset is fed into an AIOps platform, it enhances human judgment by not only providing alerts for known scenarios – such as network congestion during a major sporting event – but also by anticipating issues like latency spikes or throughput degradation before they affect subscribers. The AIOps platform can also recommend corrective actions, for instance horizontally scaling out 5G Standalone user plane function (UPF) instances to handle surges in demand and even trigger those actions automatically – improving network performance without waiting for manual intervention. 

MCP: the integration layer for network AI 

Once network data has been curated, the next challenge is delivering it to AI agents in a consistent, scalable, and maintainable way. Model Context Protocol (MCP) is designed to solve exactly this. In a mobile network architecture, the MCP server acts as a structured gateway between the data curation layer and any AI agent or large language model querying network state. 

Rather than each AI application building its own custom connector to proprietary data stores, MCP provides a standardised interface that any compliant agent can use – lowering integration cost, improving reliability, and enabling operators to expose curated intelligence to third-party applications in a controlled, low-latency manner. 

Three high-value use cases for curated data 

Curated data, delivered via MCP, enables CSPs to shift from reactive management to proactive intelligence – eliminating incidents before subscribers are affected. To illustrate, there are several essential use cases where curated data can dramatically improve outcomes for network operations and service assurance teams: 

Predictive maintenance 

Reactive strategies – waiting for equipment to fail – result in costly emergency repairs, downtime, and subscriber churn. Curated data reveals micro-patterns that precede alarms: abnormal state transitions, registration failures, traffic degradation, and quality-of-service drifts. An AI agent with access to this contextualised data can identify early warning signs and recommend corrective actions – for instance, scaling up 5G Standalone UPF’s ahead of a predicted demand surge – transitioning operations from reactive to genuinely predictive. 

Capacity planning 

Accurate forecasting requires a single, correlated view of historical signaling trends and traffic volumes by market, month, and event type. Curated data provides exactly this, enabling AIOps models to predict when and where resources will become constrained. Operators can scale infrastructure proactively, optimise performance, and align capital expenditure precisely with anticipated demand – rather than over-provisioning on guesswork.  

Smart banking and third-party applications 

Consider a mobile banking application needed to support real-time international money transfers, requiring verification of subscriber location and session integrity at each transaction. Initial approaches – querying the home location register (HLR) directly, or ingesting call detail records – failed. The databases could not handle real-time query volumes; raw packet streams overwhelmed the application. 

The solution was a data curation layer exposed via an MCP server. The banking application’s AI queried subscriber session state – location, session continuity, network domain – through a low-latency MCP interface, seeing only what it needed: a structured, queryable representation of subscriber context. The result was transaction processing at second-level resolution with measurably better user experience and no excessive overhead. 

The same architecture unlocks a broad range of use cases; these are just a few examples for curated network data. Others include digital twins, energy-efficient network operations, dynamic slice management, closed-loop automation, and security and fraud detection. This also sets the framework for various agentic AI agent use cases where this curated data pipeline is on demand via each agent’s request. In each case, success depends on relationship-aware, contextualised data – not raw traffic. 

Intelligence is about understanding, not volume 

The path to AI-driven mobile network operations is not paved with more data – it is paved with better-understood data. Curation, correlation, and contextual enrichment are not pre-processing steps to be deferred; they are the foundation on which everything depends. MCP adds the final piece: a standardised interface connecting curated intelligence to AI agents across service assurance, security, and third-party applications. 

For CSPs, the data deluge must be transformed before it can fuel the intelligent network of tomorrow. Without this critical first step, the promise of AI becomes nothing more than a recipe for wasted compute and inaccurate models. In mobile networks, the question is never how much data you can collect. It is how well you can make an AI understand it.