Grab
Grab's analytics agents answer more data requests without an analyst
Grab runs AI agents inside Slack and an internal portal so business teams can ask data questions in plain language. Grab reports that, in one self-serve channel, the share of requests answered with no human involvement rose across three request types between March and May 2026.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
53%
to 67%
Share of metric questions in a Slack self-serve channel answered with no human involvement (March to May 2026)1
63%
to 90%
Share of data pulls in the same channel answered with no human involvement (March to May 2026)1
50%
to 81%
Share of SQL requests in the same channel answered with no human involvement (March to May 2026)1
44%
to 30%
Share of closed analyst tickets that were mechanical work such as data preparation, alerting and reporting (February to June 2026)1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Almost every decision at Grab depends on data. Business teams used to wait for analysts to prepare data, write queries and build reports, so output was limited by how fast analysts could work. Grab wanted anyone making a business call to get trustworthy answers right away.
What they built
Grab built several AI agents for analytics. Spartan answers questions in plain business language inside Slack, from root-cause questions such as why revenue fell in a market to summaries of experiments. Scarlet reacts when a data pipeline fails: it finds the cause, then fixes it or hands it to the owning team. BriX is an internal portal where analysts set up a purpose-built analytics tool for a team once and run it in Slack, in a coding tool or on a schedule.
How they did it
The agents answer from certified metrics, meaning metric definitions Grab has approved, rather than from the model's memory. Spartan routes each question to one of more than 50 skills and 120 analysis frameworks. A system called ContextIQ keeps the written context for each metric up to date, and each time an agent fails in production Grab patches the context behind it. Data engineers own BriX, the platform and the support queue without a product manager or designer.
Results
In one Slack channel where self-serve agents are enabled, the share of metric questions answered with no human involvement rose from 53% to 67% between March and May 2026. Data pulls went from 63% to 90% and SQL requests from 50% to 81%. Grab also reports that the mechanical share of closed analyst tickets fell from 44% in February to 30% in June, and cycle times fell by about a third.
What to copy
Measure automation by request type, not as one average. Invest in definitions and context first. Plan what analysts do with the time they get back.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Share of metric questions in a Slack self-serve channel answered with no human involvement (March to May 2026) | 53% to 67%1 | |
| Share of data pulls in the same channel answered with no human involvement (March to May 2026) | 63% to 90%1 | |
| Share of SQL requests in the same channel answered with no human involvement (March to May 2026) | 50% to 81%1 | |
| Share of closed analyst tickets that were mechanical work such as data preparation, alerting and reporting (February to June 2026) | 44% to 30%1 | |
| Reduction in ticket cycle time (Q1 versus Q2, per Grab's chart) | ~33%1 | |
| Growth in BriX portal usage since September 2025 | more than 10x1 |
Lessons for your project
Track the share of requests closed with no human involvement by request type. Grab reports metric questions, data pulls and SQL separately, which shows where the agents are strongest.
Certify metric definitions before you add agents. Grab says unclear definitions cause most poor analytics output, and it fixes the context documents whenever an agent fails.
Put agents where people already ask for help. Grab's Slack channel is used mostly by people outside the analytics team.
Redirect freed analyst time on purpose. Grab moved it to self-serve workflows and deeper analysis, and measures the mechanical share of tickets.
Caveats
- All figures are Grab's own, from one engineering blog post. No outside party confirms them.
- The post does not name the AI models behind Spartan, Scarlet or BriX. It mentions a design tool called Claude Design for the portal interface only, which does not show which model runs the agents.
- The post says some explorations run in production and others are still experiments. It does not say which parts are which, and gives no launch date. The deployed date shown is the post date.
- The cycle-time figure is stated as ~33% and is based on a Q1 versus Q2 chart that could not be read in text. The channel figures cover one Slack channel, not all analytics work.
- Grab also gives an estimate of 230 to 470 business days of avoided backlog. It is an assumption of 1 to 2 days per ticket, so it is not used here.
- No cost, team size or build time is given for the agents themselves.
Sources
- How AI is transforming analytics at Grab
Grab Engineering Blog · 2026-08-01 · Company blog · engineering.grab.com
Summarized in our own words from the sources above. Spotted an error? See corrections.