Zenspring · Case Study · 01 / 01 · Telco Home Services · United States
Listen, Learn, and Leverage Customer Intent.
How a US Telco Aggregator transformed zero-party review data into an agentic AI lifecycle engine – reducing churn risk, driving re-engagement, and unlocking revenue hidden in plain sight.
A data-rich brand flying half blind.
The client is a US-based Telco Aggregator operating multiple comparison portals covering mobile, fibre and cable internet providers across geographies. Their top-of-funnel performance marketing was working. Their bottom-of-funnel was invisible.
Significant investment in search and social had driven strong anonymous visitor traffic. But without first-party or zero-party data, the brand had no way to reach, retain, or re-sell to the same users over time. Every visitor interaction was effectively disposable.
Meanwhile, the brand was collecting something more valuable than click data: user reviews. Real customers describing – in their own words – their frustrations, switching intent, plan concerns, and service expectations. That semantic data was being moderated and published. It was not being activated.
"Money lying on the table untouched – users with high intent already in the market, looking for action and help."
No first-party customer identity
Performance marketing drove anonymous traffic. No consent-driven data existed to enable lifecycle re-engagement or personalisation beyond a single session.
Unstructured data, unread and unactivated
User reviews contained rich semantic intent – churn risk, upgrade desire, price sensitivity – but human agents could not read and action each review at scale.
Generic outreach, diminishing returns
Without individual intent signals, any re-engagement campaign was inherently batch-and-blast. Relevance was low, click-through was low, acquisition costs kept rising.
Provider partnerships underserved by insight
Telco and cable providers wanted more than traffic. They wanted customer intelligence. The brand had that intelligence locked in review data – but could not surface or monetise it.
A network of AI agents. One unified lifecycle engine.
Zenspring deployed a two-layer solution: a ZenXLM-anchored agentic lifecycle marketing engine, underpinned by ZenConsult's AI architecture, cost governance, and data security framework. The combination turned inert review data into a real-time intent intelligence system – and that intelligence into personalised, conversion-oriented experiences.
The ZenXLM lifecycle model operates across all five stages – from the moment a user submits a review (data capture and consent) through to an individually orchestrated call-to-action experience. ZenConsult provided the architectural backbone, including data anonymisation before any LLM processing, enterprise-hosted model inference, and human-in-the-loop governance controls at each agent step.
Five signal layers. One composable AI fabric.
Composable by design. The solution was built on top of the client's existing data warehouse and review management infrastructure. No CRM migration. No platform replacement. ZenXLM's Customer DNA profiling and journey orchestration modules were deployed as a composable intelligence layer – reading from existing systems, enriching with GenAI, and surfacing outputs through familiar operational workflows.
Confidence scores derived from semantic analysis of zero-party review data, cross-referenced with transaction and call records. High-intent cohorts (score ≥ 80) prioritised for immediate personalised outreach.
Four agents. One autonomous workflow.
The Zenspring agentic solution instantiated a network of four specialised AI agents, each operating within a defined scope and governed by configurable human-in-the-loop checkpoints. The workflow ran sequentially, with each agent's output forming the input for the next.
Analyse individual user intent signals from reviews and ratings
The first agent processes each review against its associated metadata – geo, provider, rating, date – and any linked transaction or call summary records. It produces a structured intent signal set per user: churn probability, switching intent, price sensitivity, upgrade interest, and advocacy potential. Each signal carries a confidence score. Reviews are classified and ranked to form the prioritised outreach queue.
Generate personalised outreach with intent-matched calls-to-action
Once intent signals are established, a second agent generates a personalised communication per user – empathetic in tone, specific in content, and grounded in the expressed concern. CTAs are dynamically matched to the detected intent: a churn-risk user receives a comparison prompt and a retention offer; an upgrade-interested user receives a speed comparison and upgrade path. Human operators can review, approve, or regenerate before send. Fully autonomous mode is also configurable.
Render real-time personalised landing pages on click
When a user engages with the personalised email, a third agent generates a landing page in real time – assembled from the user's intent profile. Content, design hierarchy, comparison data, and CTAs are all context-specific. A price-sensitive user sees a cost-of-switching breakdown with a savings estimate. A speed-upgrade user sees a performance benchmark for their current provider vs alternatives. The page adapts on subsequent interactions based on in-session behaviour.
Surface cohort insights through natural language queries
A fourth agent provides an NLP query interface for business users – enabling non-technical stakeholders to interrogate the review and intent dataset without SQL or BI tooling. Queries such as "How many customers with a churn score above 90% are based in California?" or "Show me all high-intent upgraders by provider, sorted by confidence score" return structured, formatted results in real time. Output format is configurable per query.
Five phases. Every ZenConsult engagement.
ZenConsult provided the delivery backbone across all five phases. This is Zenspring's standard engagement model – adapted here to the Telco context, with particular emphasis on data security, AI cost governance, and change management for a brand operating across multiple consumer-facing digital properties.
Audited the existing review management system, data warehouse schema, consent and PII handling, and integration architecture. Established baseline metrics for review volume, engagement rates, and provider partner reporting needs.
Designed the composable AI layer: data anonymisation pipeline, on-premises LLM inference for PII-sensitive processing, agent orchestration framework, and the ZenXLM Customer DNA schema calibrated to the Telco aggregator context.
Forward-deployed engineers built the four-agent network, the intent scoring model, the email generation workflow, the dynamic landing page engine, and the NLP BI interface – all integrated with the client's existing infrastructure.
Launched on a sample review cohort with human review enabled. Calibrated confidence thresholds with brand and compliance teams. Activated the automated outreach workflow with configurable autonomy levels. Onboarded business users to the NLP query interface.
Established ongoing cost monitoring dashboards, model drift tracking, weekly outcome reports against the baseline, and the roadmap for expanding data inputs – speed test data, ad engagement, app behaviour – into the intent engine.
Governance designed in. Not bolted on.
The Telco client operates consumer-facing properties, holds consent-driven user data across multiple states, and serves as an intermediary for regulated service providers. AI governance was not optional. ZenConsult embedded security, cost control, and compliance architecture from week one.
PII Anonymisation Before LLM Processing
All review data was anonymised before transmission to any language model. Name, email, location identifiers were stripped and tokenised within the enterprise perimeter. No personal data left the client's infrastructure. Re-identification mapping was held exclusively on-premises.
On-Premises & Cloud-Isolated Model Inference
Model inference for intent scoring and response generation ran within the client's cloud environment. LLM training data was retained within the enterprise. The solution architecture supports both cloud-isolated and fully on-premises deployment for regulated data environments.
Human-in-the-Loop Governance Controls
Every agent workflow included configurable human oversight checkpoints. Response generation and landing page creation could be set to fully autonomous, semi-autonomous with human approval, or human-initiated modes – per channel and per intent category. Confidence thresholds gated autonomous outreach.
Consent and Brand Compliance Guardrails
All outreach operated exclusively within the consent scope captured at review submission. Brand voice, tone, and content guardrails were embedded in every agent prompt architecture. Output filtering blocked off-brand messaging, unsupported provider claims, and any content exceeding the agreed compliance boundary.
AI Infrastructure Cost Governance
Model selection was calibrated to task complexity. Intent extraction used a lightweight, fine-tuned classification model rather than a frontier LLM – significantly reducing token cost at scale. Response generation used a frontier model only for high-confidence, high-intent users. Cost per engagement was tracked and reported weekly against the defined budget envelope.
Audit Trails and Model Transparency
Every AI-assisted decision – intent score, response selection, CTA assignment, landing page variant – was logged with a full provenance record. Business users could trace any outreach communication back to the originating review, the intent signals extracted, and the model version that generated the response. Essential for provider partner reporting and regulatory review.
Measured against the baseline you set in week one.
Outcomes below reflect results from the initial pilot cohort and the projected model for full-scale deployment across the client's active review database. All metrics are reported against the pre-engagement baseline established during the ZenConsult Assessment phase.
Users with churn confidence above 80% identified from review cohort – previously invisible without semantic AI analysis.
Intent-matched personalised communications vs generic batch outreach. Behaviour-triggered messaging converts at multiples of campaign average.
Modelled conversion of users scoring above 80% churn risk – users already actively evaluating alternatives and ready to be guided to an action.
Full data security compliance. All personal identifiers anonymised before any LLM inference. On-premises model training data retention throughout.
The roadmap extends the signal surface. Phase two integrates speed test data, survey responses, referral patterns, ad engagement, app activity, email interaction signals, and CTA-based click tracking – building a progressively richer Customer DNA profile that compounds intent accuracy and personalisation relevance with every additional data source activated.
The same model. Every retail brand serving end consumers.
The architecture and methodology deployed for this Telco Aggregator is a direct expression of the ZenXLM platform – built to be composable across B2C verticals. The core pattern is constant: capture zero-party and first-party data at natural touchpoints, extract individual intent with GenAI, orchestrate personalised lifecycle experiences, and govern the AI layer with enterprise-grade security and cost controls.
Any retail or consumer services brand that collects reviews, survey data, support interactions, purchase history, or app behaviour has the same opportunity. The data exists. The AI infrastructure to activate it is composable and deployable without rip-and-replace.
Post-purchase reviews and return reasons carry churn intent and upsell signals. Zero-party preference data from product quizzes drives personalised next-purchase journeys. Loyalty re-engagement triggered by engagement decline signals before lapse.
Review and survey data from claim experiences, service interactions, and renewal conversations surface switching intent and product fit signals. Intent-matched outreach timed to key lifecycle moments – renewal dates, life events, rate changes.
Patient reviews, satisfaction surveys, and app engagement patterns signal health journey intent and care gaps. Personalised communication that guides individuals to the right service or product at the right moment in their wellness journey.
Post-stay reviews and in-stay feedback contain loyalty, upgrade, and advocacy signals. GenAI intent extraction enables hyper-personalised re-engagement at the optimal booking window – not on a campaign calendar.
Usage patterns, feedback surveys, and support interactions surface engagement health and churn risk. Lifecycle orchestration across the student or subscriber journey – from activation through renewal and advocacy.
Service reviews, warranty interactions, and connected vehicle data carry intent signals for upgrade, service upsell, and brand loyalty. Zero-party data from preference and configuration surveys drives individualised ownership experience programmes.
The revenue is already in your existing data.
Schedule a complimentary 30-minute diagnostic. We will show you what individual intent orchestration surfaces from the data you already own – before you commit to anything.