Agentic CEI · Concept overview

A living intelligence platform
for telecoms operators

Not a dashboard. Not another data pipeline. A lean, queryable model of your entire subscriber base — built from data you already have — that any team can interrogate in plain language.

fields from the CSP

80–90%

of behaviour is stable week-on-week

6 weeks

from data handover to live society

The core insight

Why this works when 300-field CEI systems don't.

Stability

Networks and people are boring

A Netflix subscriber who watches in the evening is still doing that next week. A commuter who streams Spotify on the 7:42 keeps doing it. 80–90% of behaviour doesn't change week-on-week. We exploit this.

Leanness

You don't need 300 features

Traditional CEI systems request 300+ fields, cell-level granularity, and hourly Hadoop feeds. We use 28 fields from 5 sources every CSP already exports. No new infrastructure. Deployable in weeks.

Intelligence

The agent does the reasoning

Instead of building complexity into the data pipeline, we move intelligence into the reasoning layer. The model stays lean. The agent answers hard questions by thinking across the society.

The architecture

Three layers. Lean by design.

Layer 1

The broad CEI model

Five weighted components, computed weekly per subscriber. Every number traceable to its raw input.

throughput_score    = avg(DL, UL bands)    × 0.25
service_score       = service quality band × 0.25
reliability_score   = time on 4G %         × 0.20
distress_score      = 1.0 − penalties      × 0.20
stability_score     = 1 − abs(usage delta) × 0.10
─────────────────────────────────────────────
cei_score           = weighted_sum × 100

Weights calibrated quarterly against aggregate NPS. No per-subscriber survey labels ever needed.

Layer 2

The artificial society

Subscribers are clustered into named archetypes — living personas that have CEI trends, NPS distributions, top services, and migration patterns.

Evening Streamer

Promoter

CEI

The Commuter

Passive

CEI

Remote Worker

At risk ↓

CEI

What makes it a society — not segments: archetypes track migration (subscribers moving between personas), influence scores, and the whole evolves as behaviour evolves.

Layer 3

The agentic reasoning layer

Any team. Any question. In plain English.

Network Eng.

Q · Which regions are driving the most detractors?

A · Remote Worker archetype in North Wales — CEI down 9pts, 334K subscribers. Primary signal: Teams uplink quality.

Marketing

Q · Which subscribers are ready for a 5G upgrade?

A · Evening Streamers, 25–34, premium handset, CEI trending up. 12,400 subscribers. Est. £340K revenue uplift.

Senior Exec

Q · What's our NPS trajectory next quarter?

A · Predicted +2pts if trends hold. Remote Worker is the primary risk. Model confidence: 78%.

How CEI is calculated

From raw weekly fields to a NPS-calibrated score.

CSP sends 7 raw fields per subscriber — weekly

DL throughput band, UL throughput band, top service quality band, time on 4G %, data usage delta, care contact flag, complaint flag, speed test flag.

We convert bands to scores (0–1)

Poor → 0.25 · Fair → 0.50 · Good → 0.75 · Excellent → 1.0. Penalties: complaint −0.25, care −0.15, speed test −0.10.

We apply the weighted formula → CEI score 0–100

The CSP never sends a CEI score. This is our output, not their input.

We assign NPS predicted category

Top 38% → Promoter · Middle 44% → Passive · Bottom 18% → Detractor. Matched to the CSP's known NPS distribution.

Quarterly — CSP provides aggregate NPS → model recalibrates

Weights adjust until population distribution matches reality. Only time NPS data is used — at society level, never subscriber level.

Data sources

Five standard CSP exports. No new infrastructure.

Source	Fields provided	Cadence
BSS / billing	Contract, tenure, billing tier, spend	Monthly
CRM	Age band, handset, region, community type	Quarterly
Network KPI summary	Throughput bands, 4G %, usage delta	Weekly
DPI summary export	Top service name, category, quality band	Weekly
NPS survey platform	Aggregate NPS, sample size (calibration only)	Quarterly

All five sources are standard CSP exports. No new infrastructure. A CSP data engineer can have all five feeds running in two weeks.

Why "society"?

Segments freeze people. A society lets them move.

Traditional segmentation

Subscribers as isolated buckets. Segments don't interact. No concept of movement, influence, or evolution. A dashboard shows a number. That's it.

Artificial society

Archetypes have migration paths. High-influence subscribers carry churn risk to their network. The whole population evolves and the model tracks it.

Migration tracking

1,200 subscribers migrated from Student to Commuter this quarter — their network sensitivity just doubled. That's a life event, not a network metric.

Influence scoring

A high-influence detractor carries 5× the churn risk of a low-influence one. Segments treat them identically. A society doesn't.

Honest limitations

The model is weekly batch, not real-time. No cell-level data. No live outage correlation. What it does — predict NPS trajectory, identify at-risk personas, surface intervention opportunities — it does with precision and transparency.

What's in this prototype