Analysis of 3,096 fiber optic subscribers shows a direct correlation between latency spikes (>30ms) and churn events within 30 days. Enugu TX degradation currently places 340 subscribers in the high-risk queue. Proactive intervention (speed guarantee offer in local language) could prevent ₦25.3M monthly revenue loss.
Design intervention ↗
Month-to-month + no tech support = 41.6% churn — structural vulnerability
2,341 subscribers on month-to-month contracts with no TechSupport add-on represent 33.2% of all churners despite being 23.5% of the base. Bundling tech support into month-to-month plans could reduce this segment's churn from 41.6% to an estimated 15.2%, saving ₦68M monthly.
Model impact ↗
OVERALL NPS SCORE
+41
▲ 4 points this month
NEGATIVE SIGNALS
23%
Hausa & Pidgin dominate
MULTILINGUAL COVERAGE
6 langs
WAXAL-powered ASR
CALLS ANALYSED TODAY
14,820
▲ 8.4% vs yesterday
REAL-TIME SENTIMENT BY LANGUAGE — NLP ENGINE
Positive · Neutral · Negative breakdown per language stream
Hausa · 6,840 calls today
Positive 41%Neutral 34%Negative 25%
Pidgin English · 4,210 calls today
Positive 38%Neutral 28%Negative 34%
Yoruba · 2,480 calls today
Positive 54%Neutral 30%Negative 16%
Igbo · 980 calls today
Positive 49%Neutral 32%Negative 19%
TOP COMPLAINT THEMES — NLP EXTRACTED (ALL LANGUAGES)
TOPNetwork speed (Pidgin): "My data dey finish before time" — 34% of Pidgin complaints. Correlates with fiber churn segment.Live
#2Billing confusion (Hausa): "Kuɗin data ya kare da wuri" (data finished early) — 28% of Hausa calls. Electronic check users most affected.Live
Google WAXAL ASR → Multilingual Churn Signal Extraction
The WAXAL dataset (2,000+ hours West African speech: Hausa, Yoruba, Twi, Pidgin) fine-tunes a Whisper ASR model that transcribes customer support calls in real time. A Hugging Face multilingual BERT classifier then extracts dissatisfaction signals — "My network dey do anyhow" (Pidgin, frustration), "Kuɗina ya tafi" (Hausa, billing complaint) — and feeds them directly into the churn risk score. Today, 847 calls triggered churn risk elevation after multilingual NLP flagged high-dissatisfaction language patterns. Without multilingual processing, these signals would be invisible to English-only models.
View pipeline architecture ↗
MONTHLY REVENUE
₦456B
▲ 11.2% YoY
ARPU
₦4,768
▲ ₦384 vs last quarter
REVENUE AT RISK
₦222M
1,869 predicted churners
AI RETENTION SAVES
₦38M
↑ 340 auto-interventions
MONTHLY CHARGES DISTRIBUTION BY CHURN STATUS
Retained (avg ₦61.3/unit)Churned (avg ₦74.4/unit)
CHURN BY PAYMENT METHOD — REVENUE RISK
Electronic check 45.3%Auto payments ~16%
AI-GENERATED REVENUE RECOVERY OPPORTUNITIES
₦68M/month — TechSupport bundling
Bundling TechSupport into month-to-month plans reduces churn from 41.6% to est. 15.2% in that segment. 2,341 subscribers, ₦74.4k avg charge.
Model ROI ↗
₦45M/month — e-check to auto-pay migration
Electronic check users churn at 45.3% vs 15.2% for auto-pay. 2,365 e-check subscribers represent 19.5% of revenue at risk. Auto-pay incentive campaign target.
Design campaign ↗
₦34M/month — 0–12 month early intervention
New subscribers (0–12 months) churn at 47.4% — nearly 1 in 2. Early-tenure onboarding programme with 90-day engagement plan could halve this rate.