In April, Grab unveiled 13 AI-powered products built on what it calls the Grab Intelligence Layer — trained on insight from more than 20 billion rides and orders. Several are already live in Malaysia: the driver AI assistant, Group Ride, GrabMaps for consumers, with more rolling out through the year.
That's the consumer story. The builder story is better: Grab's engineering team documents how they build this stuff, publicly, for free, at a level of honesty most companies reserve for internal wikis. If you run or build for a business in this region, their blog is the closest thing to borrowing the homework of a company operating in your exact market conditions.
What Grab Shipped This Year
Beyond the consumer features: a partnership with OpenAI to teach practical AI skills to 30,000 drivers and merchants across the region — Malaysia's turn comes in 2027 — and, this month, a framework agreement with SoftBank and PETROS to explore AI and digital infrastructure in Sarawak. One clarification worth making because it gets misreported: Grab's AI Centre of Excellence is in Singapore, not Malaysia — though they do run an R&D centre in KL.
The Engineering Blog Is the Real Gift
Big tech engineering blogs are usually recruitment ads. Grab's reads like working notes — including one post whose URL still carries a typo, which tells you nobody's polishing these for PR. Below are the ones I'd actually assign as reading, each with the thing a smaller business can steal.
The Reading List, Annotated
- [Grab AI Gateway](https://engineering.grab.com/grab-ai-gateway) — one door for every AI call in the company: one set of keys, one cost ledger, one place for rate limits, 300+ internal use cases behind it. The SME version is a fifty-line proxy. If your business uses AI in more than one place, build the small version of this first.
- [How Grab builds and runs AI agents at scale](https://engineering.grab.com/how-grab-builds-and-runs-ai-agents-at-scale) — the architecture of agents that actually do work, with the guardrails spelled out.
- [Grab Bench](https://engineering.grab.com/grab-bench-evaluating-ai) — my pick for the most transferable idea of the year: before trusting AI with your work, test it on your real tasks, not demos. Any shop can build a ten-item bench from its own invoices, messages, and reports, and score models against it before betting a workflow on one.
- [Jarvis Pro: route first, answer later](https://engineering.grab.com/jarvis-pro-route-firsr-answer-later) — send each request to the cheapest model that can handle it, escalate only when needed. This is exactly how you keep an AI feature from eating its own margin. (Yes, the typo in the URL is Grab's.)
- [Automated merchant review summaries](https://engineering.grab.com/automated-merchant-review-summary-system-with-integrated-feedback) — reads a flood of customer reviews, produces summaries a busy owner will actually read. A directly sellable feature for any business drowning in Google reviews; they published how it works end to end.
- [How AI is transforming analytics at Grab](https://engineering.grab.com/how-ai-is-transforming-analytics) — natural-language questions over business data, and the honest friction they hit.
- [Price sensitivity profiling using LLMs](https://engineering.grab.com/price-sensitivity-profiling-using-llm) — fresh this month: using models to understand which customers respond to which pricing. Read it for the method, not to copy the use case.
- [GrabMart data for disease outbreak detection](https://engineering.grab.com/grabmart-data-for-faster-disease-outbreak-detection) — the most unexpected one: grocery baskets as an early-warning signal. A reminder that the data you already collect probably answers questions you haven't asked it.
Copy the Pattern, Not the Scale
You don't have 20 billion orders. Doesn't matter. The patterns are scale-free: one gateway for AI calls. A bench of your own real tasks before trusting a model. Route to the cheapest capable model. Summarise what nobody has time to read. Every one of those works at five staff the same way it works at five thousand — the fifty-line version just costs a weekend instead of a platform team.
The Sarawak Footnote
The SoftBank-Grab-PETROS agreement is exploratory — framework agreements are intentions, not data centres. But the direction is hard to miss: serious AI infrastructure conversations are happening about Malaysia, not just near it. The companies that will benefit first are the ones whose operations are already systematised enough to plug in. That part — being ready — is a choice available at any size.