Problem & Strategy

Market dynamics, alternatives considered, and the selective friction approach

Grab needs to cut support cost per case while raising policy adherence — without killing CSAT or agent morale. The answer isn't more automation or harder rules. It's selective friction: intervene only where risk exists, and learn from every override.

Market Dynamics

Four forces shaping the opportunity for a deviation-intelligent copilot.

Multi-dimensional complexity

SEA support orgs handle multi-product, multi-language, multi-market cases — a combinatorial challenge that generic AI misses.

AI tool gap

Current AI support tools focus on conversational assist and deflection, with limited support for localized compliance and policy enforcement.

Grab already uses AI

Grab deploys AI in merchant support and driver guidance, creating foundations to extend into agent-facing copilot workflows.

Rising expectations

Regulatory and trust expectations are rising across SEA markets; cost must drop without degrading CSAT or compliance posture.

Alternatives Considered

Three approaches were evaluated before converging on selective friction (WS3).

WS1 only

All-in upstream prevention

Eliminates contact at source

Permanent cost reduction

Slow product-change cycles

Doesn't help complex cases that remain

No learning loop from agent behavior

WS2 only

Self-serve everything

Scales without headcount

Fast resolution for simple cases

Fails on nuanced, multi-step cases

No compliance enforcement on edge cases

Customer frustration on misroutes

Hard gates

Binary compliance gating

Strict policy adherence

Simple to implement

Kills agent autonomy and morale

CSAT drops on edge cases

Zero learning — rules are static

Chosen approach: Selective friction (WS3) — intervene proportionally to risk, preserve agent autonomy on corrective overrides, and learn continuously from outcomes.

Trade-off Analysis

Selective friction is the only approach that improves — or at least holds — every dimension simultaneously.

Approach	Speed	Compliance	CSAT	Automation
Hard guardrails	↓	↑	↓	Limited
Laissez-faire	↑ initially	↓	Mixed	Low
Selective frictionChosen	↔ / ↑	↔ / ↑	↔ / ↑	↔ / ↑

Build vs Buy

Build on the differentiating layer. Rent or extend everything else.

KeepCRM / Chat platform

Keep existing Zendesk / in-house

BuildSuggestion engine

Build on LLM APIs — rent model, build retrieval

BuildDeviation classifier + risk scorer

Build custom — CORE differentiator

Core

BuildSelective intervention router

Build custom — tier routing logic

Build / ExtendSupervisor queue

Build or extend existing escalation tooling

BuildPost-resolution audit

Build on existing QA infrastructure

BuildPolicy repository

Build custom — versioned, multi-market

Buy / AdoptInference gateway

Buy or adopt open-source gateway

ExtendAnalytics dashboard

Extend existing BI / analytics tooling

Competitive Landscape

Seven major players — none offer real-time deviation classification or selective friction.

Vendor	Key Strengths	Gap for WS3
Zendesk AI	Deep CRM integration, large install base, intent detection, agent assist macros	No real-time deviation classification; compliance is rule-based, not risk-scored
Salesforce Einstein	Enterprise trust, vast data platform, case routing and predictive CSAT	Heavy implementation; no selective friction model; override handling is binary
Intercom Fin	Strong self-serve deflection, modern UX, fast deployment for digital-first orgs	Customer-facing focus; limited agent-side copilot; no policy compliance layer
Ada	Automated resolution engine, multilingual, no-code bot builder	Automation-first with no human-in-the-loop deviation intelligence
Forethought	AI triage and suggested responses, workflow automation, knowledge retrieval	No override taxonomy; no risk scoring; compliance is post-hoc not real-time
Freshworks Freddy	Affordable, built-in CRM, auto-triage, canned response suggestions	Shallow AI layer; no deviation-aware intervention; limited multi-market support
Sprinklr	Omnichannel, enterprise analytics, unified agent desktop, sentiment analysis	Platform complexity; no selective friction; compliance is dashboard-only

Key insight: Every competitor optimizes for suggestion accuracy or deflection rate. None treat override classification and selective intervention as the core product — that's the open lane.

Why This Model Wins

Five reasons selective friction is the right bet for Grab.

Aligns with leadership goals

Directly targets the KPIs leadership cares about: policy adherence up, cost per case down, CSAT held or improved.

Converts judgment into data

Every agent override becomes a structured signal — corrective, cost-risk, or policy-risk — feeding continuous improvement.

Minimizes agent frustration

70% of overrides pass with no friction. Only the ~10% that carry real risk get slowed down — preserving autonomy and morale.

Creates measurable proof

Each phase produces quantifiable outcomes (FP rate, override distribution, CSAT delta) that justify continued investment.

Concentrates build effort

The differentiating layer (deviation classifier + risk scorer + intervention router) is a narrow build surface — everything else is rent or extend.

Build narrow, rent wide

The differentiating layer — deviation classifier, risk scorer, and intervention router — is a focused build. Everything else (CRM, LLM inference, analytics) is rented or extended. This keeps scope tight and time-to-value short while creating a defensible intelligence layer no vendor offers today.