How AI is transforming analytics at Grab

Maanas Prabhakar1 Aug 2026

Problem

Grab's analysts spent much of their time on mechanical work: data preparation, alerting and reporting. Stakeholders such as ad sales staff, campaign managers and country teams waited in a backlog for metric questions, data pulls and SQL. An agent that does not know a metric's grain, exclusions and caveats guesses and answers wrong with confidence. Hand-written context goes stale.

Example from the article: a commercial manager asked in Slack why revenue fell in the Philippines mid-market segment in late June, and a product manager asked for an ads experiment summary.

Solution in brief

Grab defines an autonomy ladder (L2 to L5) that states who owns each step of an analytics loop and where a human reviews. Workflows climb only when five capabilities exist: execution, certified knowledge, control checks, review gates and a learning loop. A Slack agent routes each question to a skill and an analysis framework instead of answering from model memory.

Questions this case answers
  • How to let non-technical users get trustworthy answers from an agent when metric definitions are ambiguous and data is spread across many domains
  • How to keep the knowledge an agent relies on current when every production failure should become a permanent fix
  • How to decide which parts of a recurring analytical workflow an agent may own and where human review stays mandatory
  • How to let non-analysts get trustworthy answers to metric, data-pull and root-cause questions through a chat agent?

Tools and methods

  • Router with skills and analysis frameworks — picks the diagnostic or experiment-scorecard path (50+ skills, 120+ frameworks)
  • Certified metric context and index — tell the agent what to search and how to query
  • ContextIQ — refreshes context documents from instrumentation specs and failures
  • Golden-dataset test cases and gates — make each fixed failure a repeatable check
  • RCA agent — writes scheduled metric and OKR commentary
  • Monte Carlo simulation — gives the probability of hitting a target
Continue reading — it's free

Google or a one-time email code. No password.