Grab Cuts Mechanical Analytics Work From 44% to 30% with AI Agents

Grab, the ride-hailing company, is leveraging artificial intelligence (AI) agents to streamline its analytics workflows, significantly reducing the amount of routine work performed by its data analysts. According to information from InfoQ, the company observed a 44% reduction in the share of mechanical tickets handled by analysts, dropping from 44% in February to 30% in June. This reduction encompasses tasks like data preparation, alerting, and reporting.

The company's approach is guided by a five-level autonomy model. At Level 3, humans frame questions and review the results, while agents handle data discovery, query writing, and validation. By Level 4, agents are capable of planning and orchestrating workflows, with human oversight at defined gates. Level 5 represents full autonomy, where humans set objectives, quality thresholds, and escalation rules, but retain ultimate accountability.

Maanas Prabhakar, Grab's senior analytics leader, shared insights on the evolving role of analysts in a LinkedIn post. He emphasized that the real challenge lies in determining what analysts do when AI agents take on data preparation, analysis, and other aspects. Grab's system integrates with natural language analytics requests through platforms such as Slack. It utilizes over 50 skills and 120 analysis frameworks to route requests to appropriate workflows.

For example, a query related to root causes can lead to a comprehensive analysis across certified metrics and relevant dimensions. In contrast, an experiment request might retrieve a pre-existing scorecard instead of querying the data lake. This system also benefits from a robust index architecture, drawing from a knowledge base containing more than 5,000 certified tables and metrics, 4,000 context documents, and 2,000 golden records.

Grab has invested considerable resources in ensuring the reliability of its AI agents. The company maintains a comprehensive data context system, ContextIQ, which continuously updates as instrumentation changes and incorporates fixes identified from production failures. Additionally, Grab utilizes AI agents in its analytics operations, with Scarlet being one example.

Scarlet manages pipeline failures by conducting root cause analysis and can either fix the issue or escalate it when predefined gates or runbooks are insufficient.

Beyond operational support, Grab employs AI agents to automate recurring analytics tasks. For instance, these agents handle metric and OKR commentary, analyzing significant movements, breaking them down by countries and segments, and correlating them with operational changes and experiments. The company's BriX portal has seen a tenfold increase in usage since September, with over 31 production deployments, 283 merge requests, and 60 features rolled out during the first half of the year.

In the period from March to May, self-service analytics handled a significant increase in requests without human intervention. Metric requests rose from 53% to 67%, data pulls climbed from 63% to 90%, and SQL requests increased from 50% to 81%. Approximately three-quarters of these inquiries originated outside the analytics team, and 85% received a first response within a minute.

This shift has notably lightened the workload of Grab's analysts, enabling them to focus more on developing self-service workflows and conducting deeper analyses.