Redis Real-Time Decisioning Demo | AI Customer Service Agent

Stage 1: The Architecture

Sub-15ms context assembly for AI-powered customer service decisioning

AI agents cannot act on customer context they do not have. CRM, billing history, and interaction records exist across systems that do not assemble fast enough for the live conversation. Redis Iris — built on Context Retriever and Agent Memory — sits between those systems and the AI agent, assembling the Customer 360 and surfacing prior session context before the first response.

Data Sources

CRM + Account System

Customer profile, product holdings, relationship tier

Billing + Payment System

Payment history, outstanding balances, fee records

Interaction History

Call transcripts, chat logs, resolution records, sentiment scores

Product Catalog + Eligibility

Available offers, upgrade paths, promotional rates

AI Model Registry

Intent classification, churn propensity, CLV models

Kafka / Event Stream

Real-time session events, channel signals, escalation flags

Ingest Layer

RDI

Syncs CRM, billing, and interaction history into Redis

Redis FeatureForm

Serves customer value, churn, and intent features online

Unified Context Layer

Redis RAM

Hot session state, live churn triggers, active offer eligibility, and real-time intent signals

Redis Flex

Warm interaction history, product holdings, resolution records, and customer embeddings

Redis FeatureForm

Feature store — Customer value, churn propensity, intent, and empathy response features

Redis Context Retriever

Assembles the Customer 360 — account state, product holdings, and interaction history — and exposes it as structured MCP tools for the AI agent

Redis Agent Memory

Stores session-scoped working memory for the current interaction and long-term memory of prior sessions — the agent knows what was discussed, resolved, and promised before

Decision Engine

Intent Classifier

Session intent classification and sub-signal detection

NBA Ranker

Next-best-action ranking weighted by CLV, churn risk, and fit

Policy Arbitration

Retention policy, offer constraints, and compliance rules

Redis Search

Semantic intent matching across resolution and offer vectors

Output Actions

Resolve in session

Issue handled by AI agent with full context, no escalation needed

Retention offer

Proactive retention action with personalized terms

Escalate to specialist

Warm handoff with full context pre-loaded for the human agent

Update CRM + memory

Outcome recorded in CRM and Agent Memory for future interactions

Decision Target

<15ms

Primary Goal

Retention + CSAT

Redis Role

Customer context + agent memory

Stage 2: Contact Event

A high-value customer calls about a late fee — with a prior service history the agent already knows

Marcus Chen has been a ClearPath Financial customer for 5 years. He's never missed a payment until this month. Sixty days ago he called about a similar fee — that interaction was resolved with a waiver after he mentioned a job transition. He's calling again. The AI agent has 15ms to assemble everything before the first response.

Authorization Request

Marcus Chen

5-year customer | Visa Preferred | $4,200 annual revenue

RETENTION RISK

Eventinbound_contact

ChannelPhone — routed to AI agent

Issue reportedLate fee — $35 charge on checking account

Customer tierPreferred — top 20% by revenue

Payment history47 consecutive on-time payments — first miss

Days since last contact61 days (prior: fee dispute, resolved, waiver granted)

Current churn signalModerate — first payment miss + prior fee complaint

Why This Moment Matters

Marcus has been a customer for five years and has never been late. He called 60 days ago about a similar fee and the interaction ended positively — the agent waived it after he explained a job transition.

Redis Agent Memory means the AI agent knows all of this before saying hello. The decision is not whether to waive the fee. The decision is how to handle this customer in a way that deepens the relationship.

Stage 3: Ingest

CRM, billing, and interaction history flow into Redis

RDI synchronizes account state, payment records, and interaction history. Redis FeatureForm serves the online features used by the AI agent. Redis Agent Memory persists the context from Marcus's last session and surfaces it in this one.

Source Systems → Redis

CRM

CRM + Account System

Account profile, product holdings, relationship tier, contact preferences

→

BILL

Billing + Payment System

Payment history, fee records, outstanding balances, autopay status

→

INTR

Interaction History

Prior call transcripts, resolution records, sentiment scores, escalation flags

→

PROD

Product Catalog

Available offers, upgrade eligibility, rate options, promotional terms

→

KFK

Kafka / Event Stream

Live session events, IVR signals, wait time, channel handoff state

→

Pipeline Status

CRM state syncContinuous

Interaction history syncSub-second

Agent MemoryActive — prior session loaded

Redis FeatureForm parity100%

Intent signal pathServed from Redis

Decision modeInline session context

Stage 4: Context

Redis assembles the Customer 360 — and Agent Memory surfaces what happened last time

Context Retriever builds the live customer picture. Agent Memory adds what no other system tracks: what Marcus said 60 days ago, how it was resolved, and what was promised.

Redis RAMRedis FlexRedis Context RetrieverRedis Agent Memory

Customer 360 — Live State

Account standingActive — Preferred tier, 5-year relationship

Current balance$2,847 checking — healthy

Outstanding fee$35 late fee — first in 47 months

Payment history score0.98 — one miss in 5 years

Product holdingsChecking, savings, Visa Preferred card

Churn risk signalModerate — fee complaint pattern

Agent Memory — What Redis Knows

Prior session (61 days ago)Fee dispute — waived after job transition explanation

Session resolutionPositive — customer expressed relief, thanked agent

Commitment recordedNo further fee for 90 days if autopay set up

Autopay statusNot yet enrolled — commitment not fulfilled

Sentiment trendPositive → Neutral (this call)

Long-term memory flagHigh-value customer, responds well to proactive empathy

Context signal: Redis Context Retriever assembles the Customer 360 via MCP tools from the CRM schema. Redis Agent Memory surfaces the prior session context — Marcus's job transition, the waiver granted, and the autopay commitment — so the agent opens with full context instead of asking him to repeat himself.

Stage 5: Feature Serving

Customer value and intent features hydrate in milliseconds

Redis FeatureForm serves the online features that support next-best-action for retention: CLV score, churn risk, payment health, intent classification, and session sentiment.

customer_ltv_score

Predicted lifetime value relative to segment median

0.870.3 ms

churn_risk_score

Probability of attrition in next 90 days given current signals

0.410.4 ms

payment_health_score

Long-term payment reliability adjusted for recent miss

0.910.2 ms

session_intent_score

Classified intent — fee dispute with retention sub-signal

0.780.3 ms

empathy_response_score

Predicted positive response to proactive empathy and offer

0.880.4 ms

retention_propensity_score

Likelihood of retention given appropriate action

0.830.5 ms

Feature Serving Performance

Features Hydrated

P99 Lookup

1.9 ms

Train / Serve Parity

100%

Stage 6: Decision

Waive the fee, enroll in autopay, and check upgrade eligibility — one ranked action

The NBA ranker evaluates available actions using the Customer 360, Agent Memory context, and live session features. This is not a script. It is a ranked action grounded in everything Redis knows about Marcus.

#1 Winning Action

WAIVE + RETAIN

Waive fee proactively — offer autopay enrollment and rate review

High CLV + payment history + Agent Memory context all point the same direction

Agent Memory confirms prior waiver was granted with positive outcome. Repeat proactive empathy. Waive the fee before Marcus asks, enroll in autopay, and check Preferred upgrade eligibility.

Action confidence0.91

#2 Escalate

RETENTION OFFER

Escalate to retention specialist with context pre-loaded

Valid if churn risk were higher or CLV signal were mixed

Not needed here — the AI agent has enough context to resolve in session. Escalation would add friction for a customer who wants a quick, respectful resolution.

Action confidence0.67

Suppressed

STANDARD RESPONSE

Apply fee with standard messaging

Suppressed — ignores Agent Memory and CLV signal entirely

This is the action a generic IVR or scripted agent would take. It misses the relationship context, violates the implicit commitment from 60 days ago, and increases churn risk.

Action confidence0.03

Stage 7: Business Impact

The value is retention protected, relationship deepened, and churn risk reduced

Proactive context-driven handling converts a fee complaint into a retention moment. The cost of the waiver is far below the cost of losing a Preferred-tier customer.

Operational Value

Fee waiver cost$35 — vs. $4,200 annual CLV at risk

First contact resolutionIssue resolved in session — no escalation

Autopay enrollmentReduces future payment miss risk

Churn risk delta0.41 → 0.18 — significant reduction

Agent Memory updateSession outcome recorded for next interaction

Relationship signalCustomer leaves the interaction with positive sentiment

Per-Interaction Outcome

complaint unresolved

generic policy
no prior context

→

retained + deepened

proactive waiver
relationship protected

Stage 8: Outcome

Same customer. Same fee. Completely different relationship outcome.

Without Redis, the agent sees a fee complaint with no prior context. With Redis, the agent opens with the full customer picture and Agent Memory of what happened last time.

Generic Agent Response

Marcus Chen

Inbound — fee complaint

What the agent knows

Account number and fee amount only. No history, no prior interactions.

Action taken

Standard fee policy applied. Customer asked to repeat context.

Outcome

Escalated. Fee not waived. Churn risk increases.

repeated

history ignored

escalated

no resolution

at risk

churn signal

Redis Iris-Powered Agent

Marcus Chen

Preferred — 5yr — Agent Memory loaded

Winning action

Waive fee + enroll in autopay

Agent Memory: fee waived 61 days ago after job transition. Positive outcome. Proactive waiver + autopay = relationship protected.

Why this wins

Prior context + high CLV + empathy signal

Customer 360 + Agent Memory assembled in <15ms. 91% confidence. Churn risk: 0.41 → 0.18.

91%

confidence

<15ms

context ready

retained

relationship

Stage 9: Architecture Recap

Redis Iris becomes the context engine between existing systems and the AI agent

CRM, billing, and interaction history systems stay in place. Redis Context Retriever provides governed, schema-defined access to the customer data. Redis Agent Memory provides what no other system captures — the context of what actually happened in prior conversations.

Data Sources

CRM + Account System

Customer profile, product holdings, relationship tier

Billing + Payment System

Payment history, outstanding balances, fee records

Interaction History

Call transcripts, chat logs, resolution records, sentiment scores

Product Catalog + Eligibility

Available offers, upgrade paths, promotional rates

AI Model Registry

Intent classification, churn propensity, CLV models

Kafka / Event Stream

Real-time session events, channel signals, escalation flags

Ingest Layer

RDI

Syncs CRM, billing, and interaction history into Redis

Redis FeatureForm

Serves customer value, churn, and intent features online

Unified Context Layer

Redis RAM

Hot session state, live churn triggers, active offer eligibility, and real-time intent signals

Redis Flex

Warm interaction history, product holdings, resolution records, and customer embeddings

Redis FeatureForm

Feature store — Customer value, churn propensity, intent, and empathy response features

Redis Context Retriever

Assembles the Customer 360 — account state, product holdings, and interaction history — and exposes it as structured MCP tools for the AI agent

Redis Agent Memory

Stores session-scoped working memory for the current interaction and long-term memory of prior sessions — the agent knows what was discussed, resolved, and promised before

Decision Engine

Intent Classifier

Session intent classification and sub-signal detection

NBA Ranker

Next-best-action ranking weighted by CLV, churn risk, and fit

Policy Arbitration

Retention policy, offer constraints, and compliance rules

Redis Search

Semantic intent matching across resolution and offer vectors

Output Actions

Resolve in session

Issue handled by AI agent with full context, no escalation needed

Retention offer

Proactive retention action with personalized terms

Escalate to specialist

Warm handoff with full context pre-loaded for the human agent

Update CRM + memory

Outcome recorded in CRM and Agent Memory for future interactions

Decision Target

<15ms

Primary Goal

Retention + CSAT

Redis Role

Customer context + agent memory