Customer Central
Designing a unified agent workspace for a platform serving thousands of small businesses
My role: Product strategy, data architecture, UX design, interactive prototype, and delivery planning. This was a solo project completed as a proposal for a product leadership position.
Six Problems Compounding Into One
A mid-market HR and payroll platform had a support team of roughly 50 agents fielding calls from thousands of SMB customers. On every call, agents were toggling between 5-6 disconnected systems — a CRM, a payroll admin tool, a billing dashboard, a benefits portal, and a ticketing system — just to answer basic questions.
The friction wasn't just inefficiency. It was a cascade of compounding problems, each one making the others worse. Here are the six I identified:
Context-Switching Tax
Agents lose 30-60 seconds per call switching between 5-6 systems. At 40+ calls/day, that's hours of wasted productive time.
Caller Identity Problem
When a call comes in, agents can't quickly determine who the caller is, their role, or their history. Every call starts cold.
Repeat-Yourself Problem
Customers re-explain their issue to each new agent. No shared timeline of interactions means every handoff resets context.
Account Status Blindspot
Agents can't see billing, payroll, or benefits status at a glance. They discover problems reactively, mid-call, in separate systems.
Proactive Blindspot
No alerts for upcoming payroll deadlines, failed payments, or expiring enrollment windows. Agents are always firefighting.
Open Issues Visibility Gap
No quick way to see unresolved tickets for an account. Agents unknowingly re-open resolved issues or miss escalations.
Three Guiding Principles
Rather than building a massive all-in-one tool, I designed the architecture around three principles that sequence risk and maximize learning:
Read Before Write
Ship a read-only MVP first to validate data accuracy before enabling write-back actions. This lets agents verify information without risk of corrupting source systems.
Dependency Sequencing
Phase the architecture so each layer enables the next: data aggregation first, then intelligence, then prediction. No phase depends on unproven technology.
Tiebreaker Principle
When features tie on near-term value, pick the one that builds toward the AI-native future. This ensures the MVP creates optionality, not just utility.
Feature Architecture: 11 MVP Stories in 3 Tiers
Each feature was scoped, categorized by build type, and mapped to a specific problem it solves. The architecture is designed so Foundation features are prerequisites — they provide the data layer that Problem-Specific and Health features build on.
| ID | Feature | Tier | Build Type |
|---|---|---|---|
| FS-1 | Caller recognition & contact card | Foundation | Full Build |
| FS-2 | Interaction history timeline | Foundation | Full Build |
| FS-3 | Account status strip | Foundation | Full Build |
| PS-1 | Pay stub viewer with comparison | Problem-Specific | Deep Link |
| PS-2 | Payroll alerts & status | Problem-Specific | Alert + Deep Link |
| PS-3 | Tax document access badges | Problem-Specific | Deep Link |
| PS-4 | Benefits summary card | Problem-Specific | Deep Link |
| PS-5 | Invoice & billing summary | Problem-Specific | Deep Link |
| PS-6 | Payment health & risk cascade | Problem-Specific | Alert + Deep Link |
| HS-1 | Lightweight account health score | Health | Full Build |
| HS-2 | Leadership portfolio view | Health | Full Build |
A Working Prototype to Validate the Design
I built a functional prototype that demonstrates both the Phase 1 read-only foundation and the Phase 3 AI-powered intelligence view. The Agent Dashboard shows how all 11 MVP features come together in a single workspace. The Intelligence View previews the AI-native future — interaction summaries, churn prediction, and an agent copilot — all built on the same data layer.
Read-Optimized Aggregation Layer
Customer Central doesn't replace source systems — it reads from them through a lightweight aggregation layer that normalizes data from five different sources. This approach avoids the risk of a full data migration while still delivering a unified view.
Aggregation Layer
Read-only data normalization, caching, and pre-computation. No write-back in Phase 1.
Customer Central
Unified workspace — all 11 features render from a single, pre-computed data model.
Instant (<500ms)
Caller ID, account status, health score
Fast (~2s)
Interaction timeline, open tickets
Background (2-5s)
Pay stubs, invoices, benefits
On Demand
Tax docs, payment history
Three Phases, One Foundation
The delivery plan sequences risk: Phase 1 is read-only and validates the data layer. Phase 2 adds intelligence once we trust the data. Phase 3 introduces AI capabilities once we have the behavioral data to train on.
Foundation
Read-only MVP. All 11 features. Aggregation layer, unified workspace, health scores. Validates data accuracy with real agent usage.
Intelligence
ML-powered health scores, write-back actions, proactive alerts. Builds on Phase 1 behavioral data and validated data layer.
Prediction
AI copilot, churn prediction, cross-functional signals. Requires the intelligence layer and sufficient training data from Phase 2.
Team Composition (Phase 1)
Product Manager
Owns roadmap & stakeholders
Backend Engineer
Aggregation layer & APIs
Frontend Engineer
Workspace UI & components
Designer
UX research & design system
QA / Support Lead
Agent testing & feedback loops
From Data to Intelligence to Prediction
The MVP isn't the destination — it's the foundation layer of a three-tier system. Each phase unlocks capabilities that are impossible without the one below it. The data aggregation layer makes intelligence possible; the intelligence layer makes prediction possible. Here's what each layer adds.
Layer 3: Prediction
AI copilot, churn prediction, cross-functional signals
Layer 2: Intelligence
ML scoring, write-back actions, proactive alerts
Layer 1: Foundation
Aggregated data, unified workspace, rules-based health
Phase 2: Intelligence Features
ML-Adaptive Health Scoring
Segment-aware weights that learn from outcomes. The model detects churn indicators that static rules miss — declining login frequency, support escalation patterns, payment delays.
Proactive Alert System
Surface issues before agents hear about them: missed payments, declining NPS trends, approaching payroll deadlines, and expiring enrollment windows.
In-Place Actions
Create tickets, update CRM records, and trigger workflows without leaving the workspace. Write-back to source systems with audit trails and rollback safety.
Benefits Life Events
Marriage, new baby, divorce — life events that cascade across payroll, benefits, and tax systems. Detect the trigger once, coordinate the response everywhere.
Phase 3: AI-Native Features
AI Interaction Summary
NLP processes 90+ days of interaction history — calls, emails, tickets, chats — and generates an instant narrative. Agents get full context in seconds, not minutes of scrolling.
Churn Prediction
Combines interaction velocity, sentiment trajectory, and escalation frequency into a predictive risk score. Flags at-risk accounts before renewal conversations, not after cancellation.
Agent Copilot
Real-time AI assistance during live calls: suggested responses based on account context, next-best-action recommendations, and auto-drafted follow-up emails.
Cross-Functional Intelligence
The AI layer doesn't just help support agents — it routes insights to the teams that can act on them. Every customer interaction becomes a signal that feeds product, sales, content, and leadership decisions.
AI Signal Router
Cross-functional insight engine
How We'd Know It's Working
Metrics were designed in three tiers: leading indicators to validate adoption within weeks, outcome metrics to confirm business impact at 90 days, and guardrails to ensure we don't optimize one thing at the expense of another.
Leading Indicators (Weeks 1-2)
Outcome Metrics (90 Days)
Guardrails
Tradeoffs Worth Explaining
Why read-only first?
Write-back actions (creating tickets, updating records) carry data integrity risk. By shipping read-only first, we validate the aggregation layer's accuracy with zero risk to source systems. Write-back becomes a Phase 2 unlock once agents trust the data.
Why rules-based health score over ML?
ML models need training data we don't have yet. A weighted formula (billing 30%, payroll 25%, support 20%, satisfaction 15%, activity 10%) is transparent, debuggable, and immediately useful. ML upgrades to this in Phase 2 once we have behavioral baselines.
Why 7 weeks, not 12?
A shorter timeline forces scope discipline and earlier feedback loops. The read-only constraint actually makes this possible — we're building a presentation layer, not modifying core systems. Each sprint ships usable functionality.
Why "NPS per headcount dollar" as north star?
Traditional support metrics (AHT, FCR) can be gamed. Tying customer satisfaction to staffing efficiency creates a metric that only improves when you genuinely help agents help customers better — not when you rush them off calls.