Grab's analytics team cut the share of mechanical work handled by human analysts from 44% to 30% between February and June 2026. The 14-percentage-point drop was driven by agentic delegation across data preparation, alerting, and reporting. The result is a smaller analyst bottleneck and a cleaner separation between what machines should own and what still requires human judgment.

The architecture runs on a five-level autonomy ladder borrowed from AI coding frameworks. At Level 3 — where most of Grab's production workloads sit — humans frame the question and review the result. Agents handle everything in between: data discovery, query authoring, execution, sanity checking, and draft write-up. Level 4 shifts planning and orchestration to the agent; humans review at defined gates rather than every step. Level 5 is end-to-end autonomy with humans setting objectives, quality thresholds, and escalation rules. Human accountability for metric definitions, causal interpretation, and final decisions is explicitly preserved at every level.

LevelHuman RoleAgent RoleGrab Production Status
L1Full control over all stepsProvides suggestions only
L2Reviews every individual stepAssists with step execution
L3Frames the question; reviews final resultData discovery, query authoring, execution, sanity check, draft write-upMost production workloads
L4Reviews at defined checkpoint gatesPlanning, orchestration, and full executionTargeted workloads
L5Sets objectives, quality thresholds, escalation rulesEnd-to-end autonomous operationAspirational target
FIG. 02 Grab's five-level agent autonomy ladder: human vs. agent responsibilities at each level — Grab Engineering / InfoQ, 2026

The central system is Spartan, an end-to-end agentic analytics workflow embedded in Slack. It routes plain-language requests through 50+ skills and 120+ analysis frameworks. A root-cause question—"why did revenue fall in the Philippines mid-market segment in the last two weeks of June?"—triggers a diagnostic path through segment, market, and campaign type against certified metrics. An experiment question pulls a pre-computed scorecard from the experiment platform rather than re-querying the data lake. The routing distinction matters: the system avoids unnecessary compute by reading the intent of the question before touching any data.

Spartan request-routing architecture: plain-language Slack query to final answer via intent classification and ContextIQ
FIG. 03 Spartan request-routing architecture: plain-language Slack query to final answer via intent classification and ContextIQ — Grab Engineering, 2026

The data foundation underpinning Spartan is deliberately heavy. Grab maintains 5,000+ certified tables and metrics, 4,000 context documents, and 2,000 golden records. ContextIQ — the platform managing this layer — treats context as having a lifecycle rather than as a static document. A skill reads an incoming instrumentation spec, proposes the SQL changes that follow, and updates the context document in the same pass. The team stated the rationale: "An agent that does not know a metric's grain, its exclusions, and its caveats will guess and confidently produce wrong outputs at speed and at scale." Context is the ceiling on agent quality; every production failure feeds back into it.

Asset TypeCountRole in Agent Quality
Certified tables & metrics5,000+Ground-truth data layer; prevents agents guessing metric definitions
Context documents4,000Captures metric grain, exclusions, and caveats; lifecycle-managed, not static
Golden records2,000Reference records used for output validation and sanity checking
FIG. 04 Grab's ContextIQ data foundation: certified assets underpinning Spartan's agent layer — Grab Engineering, 2026

Scarlet, a separate agent, covers the operations side. It handles pipeline failures by running root cause analysis, applies documented runbook fixes automatically, and escalates when no defined gate or runbook covers the problem. Exceptions surface in Slack. Cycle times across analytics operations dropped approximately 33% comparing Q1 to Q2 2026.

Self-service numbers reveal the sharpest part of the story. Between March and May, the share of analytics threads resolved without any human involvement rose from 53% to 67% for metric questions, 63% to 90% for data pulls, and 50% to 81% for SQL requests. Three-quarters of threads now originate outside the analytics team—from ads salespeople, campaign managers, and country teams. 85% receive a first response within a minute. The BriX portal, which hosts analytics workflow development, grew more than 10× in usage since September and shipped 31 production deployments, 283 merge requests, and 60 features in H1.

Self-service resolution rates (% of threads resolved without human involvement): March vs. May 2026
FIG. 05 Self-service resolution rates (% of threads resolved without human involvement): March vs. May 2026 — Grab Engineering / InfoQ, 2026

Grab's analytics lead Maanas Prabhakar posed the core tension: "The harder question is what an analyst does when an agent handles the data preparation, the analysis, and everything else." Grab's answer is that analysts own problem framing, metric definitions, the causal story behind a movement, business assumptions, and the go/no-go decision — the loop, not the steps inside it.

For ML platform teams evaluating agentic delegation: the 44-to-30 result requires a 5,000-certified-metric foundation. Instrument your data layer before instrumenting your agents.