Grab Shifts Analytics From Busywork to Strategy Using AI Agents

Grab is reshaping how its analytics teams operate by deploying AI agents to handle mechanical, repetitive work—from data prep to reporting. The results are tangible: the percentage of routine tickets analysts manage has dropped from 44% in February to 30% in June, and self-service answers are now generated without human intervention far more often.

How the Autonomy Model Works#

The company has structured its approach around a five-level autonomy framework that clarifies which parts of an analytics workflow humans control and which agents can own.

• Level 3: Humans pose the question and check the result; agents handle data discovery, query writing, execution, validation, and draft reporting.
• Level 4: Agents plan and coordinate multi-step workflows; humans review at defined checkpoints.
• Level 5: End-to-end automation, with humans setting objectives, quality bars, and rules for escalation.

Throughout all levels, humans retain final responsibility for metric definitions, causal reasoning, business assumptions, and strategic decisions. This boundary keeps accountability clear while letting agents handle the mechanical lifting.

Spartan: The Analytics Backbone#

Grab's Spartan system brings the model to life, fielding natural language requests—including questions dropped in Slack—and routing them to the right workflows. The system draws on more than 50 specialized skills and 120 analysis frameworks. A question about root causes triggers cross-metric analysis, while an experiment inquiry can resurface an existing scorecard without hitting the data lake again.

Behind the scenes, Grab has built a substantial knowledge foundation: over 5,000 certified tables and metrics, 4,000 context documents, and 2,000 gold-standard records. A system called ContextIQ treats this context as living data, refreshing it as infrastructure evolves and learning from production failures to improve agent reliability.

Agents Beyond User Queries#

Grab extends agents beyond answering questions. Scarlet, for instance, investigates pipeline failures, performs root cause analysis, and either fixes them or escalates when the situation falls outside documented runbooks. For recurring work, agents automatically draft commentary on metrics and OKRs, flagging significant shifts and correlating them to operational events or ongoing tests.