Grab Cuts Mechanical Analytics Work from 44% to 30% with AI Agents
In this article
Grab's analytics engineering team reports that the share of analytics tickets classified as mechanical work dropped from 44% in February 2026 to 30% by June 2026 — a 14-percentage-point reduction achieved in four months through a structured AI agent deployment. For teams asking what the agentic era means for data science, Grab's stack offers one of the most detailed public production accounts available.
Five-Level Autonomy Model
Grab's agents operate under a five-level autonomy framework with explicit human-oversight obligations at each tier. At Level 3, humans frame the question and review the result; agents discover data, write and execute queries, validate results, and draft analysis. At Level 4, agents plan and orchestrate workflows while humans review at defined gates. Level 5 is fully end-to-end: humans set objectives, quality thresholds, and escalation rules, then step back. Across all levels, Grab reserves four responsibilities for humans: metric definition, causal interpretation, business assumption validation, and final decisions. This mirrors the layered control thinking described in four agent control layers, no shared contract, where autonomy expansion requires explicit boundary contracts rather than implicit trust.
Spartan, Scarlet, and Certified Data
The production system, named Spartan, handles natural language queries submitted through Slack, routing requests using more than 50 skills and 120 analysis frameworks. A root-cause query triggers multi-dimensional analysis across certified metrics; an experiment question retrieves an existing scorecard rather than re-querying the data lake, avoiding redundant compute and drift risk.
Reliable agent output depends on what Grab calls certified context. The team maintains more than 5,000 certified tables and metrics, 4,000 context documents, and 2,000 golden records. Its ContextIQ subsystem treats this corpus as a living artifact: context updates when instrumentation changes, and production agent failures feed back into correction cycles.
Scarlet handles the operational side — pipeline failures — by performing automated root cause analysis and either applying a fix or escalating when the failure falls outside predefined runbooks.
Self-Service Throughput
The shift in analyst workload is backed by throughput numbers across three request categories between March and May 2026:
| Request Type | Self-Served (March) | Self-Served (May) |
|---|---|---|
| Metric requests | 53% | 67% |
| Data pulls | 63% | 90% |
| SQL requests | 50% | 81% |
About 75% of all threads originated outside the analytics team, and 85% received a first response within one minute. The BriX portal, which supports analytics workflow development, grew more than tenfold in usage since September. In the first half of 2026, the team logged 31 production deployments, 283 merge requests, and 60 features shipped.
What the Numbers Indicate
The percentage gains in self-service SQL and data pulls — both rising more than 30 points in two months — suggest that once the certified data layer reaches sufficient coverage, throughput scales quickly. The architecture's explicit human-in-the-loop gates at Levels 3 and 4 provide a replicable pattern for teams concerned about agent reliability in high-stakes analytical contexts. As organizations instrument agent autonomy defaults with formal safety constraints, Grab's five-level model offers a concrete reference implementation: the primary constraint on autonomy gains appears to be data certification and context lifecycle management, not model capability.