Grab has slashed the proportion of routine analytics tasks handled by human analysts from 44% in February to 30% in June by deploying AI agents that automate data preparation, alerting, and reporting workflows, according to a report published by InfoQ. The company's approach centers on a five-level autonomy framework that determines how much of an analytics workflow an agent can control while preserving human oversight at critical decision points. The shift has allowed analysts to spend less time on mechanical work and more on building self-service tools and conducting deeper analysis.
Between March and May, the share of self-service analytics requests completed without human involvement jumped from 53% to 67% for metric questions, from 63% to 90% for data pulls, and from 50% to 81% for SQL queries, the report states. Approximately three-quarters of requests came from outside the analytics team, while 85% received an initial response within 60 seconds. Grab's BriX portal, which supports analytics workflow development, experienced usage growth exceeding tenfold since September, with the team shipping 31 production deployments, 283 merge requests, and 60 features during the first six months of the year. The company maintains more than 5,000 certified tables and metrics, 4,000 context documents, and 2,000 golden records to provide agents with the data context needed for reliable outputs.
Maanas Prabhakar, who leads analytics at Grab, framed the challenge in a LinkedIn post: "The harder question is what an analyst does when an agent handles the data preparation, the analysis, and everything else." Grab's Spartan system processes natural language analytics requests submitted through channels like Slack, using over 50 skills and 120 analysis frameworks to direct requests to specialized workflows. A root cause inquiry can trigger analysis across certified metrics and relevant dimensions, while an experiment question can retrieve an existing scorecard rather than querying the data lake. Grab also uses AI agents for analytics operations—Scarlet, for instance, handles pipeline failures by conducting root cause analysis and can either fix failures or escalate them when predefined gates or documented runbooks don't cover the issue.
The five-level autonomy model defines escalating degrees of agent independence. At Level 3, humans formulate the question and review the outcome while agents discover data, write and execute queries, validate results, and draft analysis. At Level 4, agents can plan and orchestrate workflows while humans review defined checkpoints. Level 5 represents complete end-to-end autonomy, with humans setting objectives, quality thresholds, and escalation rules. The company retains human accountability for metric definitions, causal interpretation, business assumptions, and final decisions. Grab's ContextIQ system treats data context as a lifecycle, updating information as instrumentation changes and incorporating fixes identified from production agent failures. For recurring analytics, Grab uses agents to automate metric and OKR commentary, assessing significant movements, breaking them down across countries and segments, and correlating them with operational changes and experiments.
The report indicates that Grab's strategy has freed analysts from mechanical workload, enabling greater focus on self-service workflows and deeper analytical investigation. The company's investment in certified data infrastructure—spanning thousands of tables, metrics, and context documents—provides the foundation agents need to generate trustworthy results without constant human intervention. By retaining human judgment over metric definitions and causal reasoning while automating data handling and query execution, Grab has created a division of labor that accelerates response times and expands the volume of questions the analytics function can address. The model suggests that analytics teams willing to codify expertise and invest in structured context can dramatically reduce time spent on routine requests while preserving human control over interpretation and strategy. Organizations pursuing similar automation will need to decide which analytical decisions require irreducible human judgment and which can be safely delegated to systems that operate within predefined boundaries.