August 19, 2026
The Agent Levels Up, But the Decision Never Changes Hands
Grab measured how much analytics work was mechanical, put agents on those steps, and defined four points that never get delegated: metric, premise, causality, decision.
Grab, the Southeast Asian superapp, measured how much of a data analyst's work was actually analysis.
In February, 44% of tickets closed by the analytics team were mechanical tasks: preparing data, triggering alerts, writing reports. After placing AI agents on those steps, the company measured again in June. The number dropped to 30%.
Grab named its own autonomy levels, inspired by Dan Shapiro's five-level concept created for code, without inheriting his numbering or labels. At level 2, the agent only suggests a chart or a first draft query; at level 5, it is nearly fully autonomous. At every level, four points remain the analyst's: metric definition, business premise, causal interpretation, and final decision. Autonomy climbs step by step, and those four points stay where they are. At level 4, the agent already runs the entire flow on its own, from data discovery to publication, and the human steps back in when the model triggers a gate, an anomaly, or sensitive data. These are the points we do not delegate, in product or in analytics.
Cycle time dropped 33%. In a Slack channel opened for self-service, the share of questions answered without an analyst queue rose from 50% to 81% for SQL queries and from 63% to 90% for raw data requests. The analytics team reinvested that capacity in deeper analysis.
The deeper analysis happens at the level 4 gate: the agent flags, the human decides. At every step of the ladder, the final decision still belongs to us.
Sources: Grab Engineering Blog · InfoQ