Workstream 3 — Copilot for Human Agents

Deviation Intelligence for Grab Support

A copilot that classifies agent overrides in real time, applies friction only where risk exists, and continuously learns from human judgment to strengthen automation.

Deviation handling is the product, not a side dashboard

Agents override copilot suggestions for good reasons and bad ones. The system that distinguishes corrective judgment from risky behavior — in real time, across 6 markets — is the core intelligence layer. Everything else (suggestions, automation, analytics) is built on top of it.

Three Workstreams

Grab's support transformation runs in parallel. This project is Workstream 3.

1

Upstream Prevention

Remove preventable cases at source. Product and process fixes that eliminate contact reasons entirely.

2

Self-Serve Automation

Automate resolved patterns end-to-end. FAQ, chatbots, and guided flows for repetitive inquiries.

This Project
3

Copilot for Agents

Deviation-intelligent copilot for complex cases. Classifies overrides, applies selective friction, and learns continuously.

Workstream 3 feeds learnings intoWorkstream 2 (self-serve graduation)Workstream 1 (upstream prevention)

Target Metrics

Policy Adherence
88%+
from ~75% baseline
CSAT Score
≥4.2
no regression
Cost per Case
<$0.01
LLM cost target
Classification
<500ms
end-to-end latency
Risk Precision
>85%
supervisor-confirmed
Agent Adoption
>70%
within 60 days

Explore the Analysis

Product thinking, technical architecture, and a working prototype.