Grab Deploys Custom AI for Food Delivery

Grab, Southeast Asia's leading superapp, is deploying a suite of custom-built AI models designed to overhaul its food delivery operations across the region. The initiative marks one of the most ambitious applications of proprietary machine learning in the food delivery sector outside of Western markets like DoorDash and Uber Eats.

The company, headquartered in Singapore and serving over 700 cities across 8 countries, is leveraging these models to tackle challenges unique to Southeast Asian markets — from hyper-dense urban traffic in Jakarta to fragmented merchant ecosystems in Ho Chi Minh City. Unlike off-the-shelf solutions from Google Cloud or AWS, Grab's approach centers on models trained specifically on regional data, accounting for local driving patterns, weather disruptions, and cultural dining habits.

Key Facts at a Glance

• Custom AI models are being deployed across Grab's food delivery vertical in 8 Southeast Asian markets
• The system targets 3 core areas: route optimization, demand forecasting, and merchant-rider matching
• Grab processes over 20 million food delivery orders per month, generating massive proprietary datasets
• Initial internal benchmarks reportedly show a 15-20% improvement in estimated delivery times
• The AI stack is built on a combination of transformer-based architectures and reinforcement learning
• Grab's R&D investment in AI exceeded $150 million in 2024, with further increases planned for 2025

Why Grab Built Its Own Models Instead of Using Third-Party AI

Southeast Asia presents a fundamentally different operating environment compared to North America or Europe. Road networks are less predictable, address systems are often informal, and traffic conditions can shift dramatically within minutes due to monsoon rains or motorbike congestion.

Third-party routing APIs — including those from Google Maps and Mapbox — provide solid baseline performance. However, Grab found that these general-purpose solutions consistently underperformed in dense urban corridors like Manila's Makati district or Bangkok's Sukhumvit area, where last-mile navigation requires hyper-local knowledge.

By training models on its own historical data — encompassing billions of completed trips, rider GPS traces, and order timestamps — Grab can capture patterns that external providers simply cannot. This mirrors a broader trend seen at companies like DoorDash, which invested heavily in its own logistics AI, and Meituan in China, which built proprietary delivery algorithms that now handle over 40 million daily orders.

How the AI System Works Under the Hood

Grab's AI deployment operates across 3 interconnected layers, each addressing a different aspect of the delivery pipeline.

Demand Forecasting Engine

The first layer uses time-series transformer models to predict order volume at granular geographic levels. The system forecasts demand in 15-minute intervals across hexagonal grid cells, allowing Grab to pre-position riders in high-demand zones before orders even come in. This approach reduces idle time for riders and cuts average wait times for customers.

Dynamic Route Optimization

The second layer handles real-time route optimization using a combination of graph neural networks and reinforcement learning. Unlike traditional shortest-path algorithms, this system considers live traffic feeds, road surface conditions, and even building entry points — a critical factor in Southeast Asian cities where many restaurants are located inside multi-story shopping malls or narrow alleyways.

The model updates route recommendations every 30 seconds, adjusting for sudden traffic changes or road closures. Internal testing shows this dynamic approach outperforms static routing by approximately 18% in average delivery time.

Merchant-Rider Matching

The third layer optimizes how orders are assigned to delivery riders. Rather than simple proximity-based matching, the AI considers:

• Estimated food preparation time at each merchant
• Rider's current trajectory and upcoming order queue
• Historical rider performance at specific restaurants
• Vehicle type (motorbike vs. car) and cargo capacity
• Real-time weather conditions affecting travel speed

This multi-factor matching system reduces the likelihood of riders waiting idle at restaurants — a common pain point that degrades both rider earnings and customer satisfaction.

The Data Advantage Grab Holds Over Western Competitors

Data scale is the critical differentiator in Grab's AI strategy. Operating across Indonesia, Malaysia, Thailand, Vietnam, the Philippines, Singapore, Myanmar, and Cambodia, Grab has accumulated one of the richest mobility and commerce datasets in the developing world.

Each completed delivery generates dozens of data points — GPS coordinates, timestamps, merchant preparation speeds, customer rating patterns, and payment method preferences. Multiplied across 20 million monthly food orders, this creates a training corpus that is both massive and deeply localized.

Compared to Uber Eats, which exited several Southeast Asian markets in 2018, or Deliveroo, which pulled out of the region entirely, Grab has the advantage of uninterrupted data collection spanning nearly a decade. This continuity allows its models to capture long-term seasonal patterns, such as demand spikes during Ramadan in Indonesia or Lunar New Year across the region.

The company has also invested in building its own data labeling infrastructure in Vietnam and the Philippines, where teams annotate edge cases — unusual delivery scenarios, atypical traffic events, and merchant-specific quirks — to improve model accuracy on tail-end situations that generic AI systems often mishandle.

Industry Context: AI Arms Race in Food Delivery

Grab's move aligns with a global trend of food delivery platforms investing heavily in proprietary AI. DoorDash has publicly discussed its use of machine learning for delivery time estimation and dasher dispatching. Uber Eats leverages models from Uber's broader AI research division. In China, Meituan operates what is arguably the world's most sophisticated delivery AI, coordinating millions of daily orders with sub-minute precision.

The stakes are enormous. McKinsey estimates the global online food delivery market will reach $1.2 trillion by 2027, with Southeast Asia representing one of the fastest-growing segments. In this market, even a 5% improvement in delivery efficiency can translate to hundreds of millions of dollars in annual savings and significantly higher customer retention rates.

What sets Grab apart is its superapp model. Unlike single-purpose delivery platforms, Grab combines ride-hailing, food delivery, grocery shopping, payments, and financial services into a single ecosystem. This means its AI models can draw on cross-vertical data — for example, using ride-hailing traffic patterns to improve food delivery routing, or leveraging payment data to predict order frequency.

What This Means for Developers, Businesses, and Users

For developers and AI practitioners, Grab's approach offers a case study in building domain-specific models rather than relying solely on foundation model APIs. The company's use of reinforcement learning for real-time dispatching and graph neural networks for routing demonstrates that traditional ML techniques, when combined with rich proprietary data, can outperform general-purpose large language models in specialized operational tasks.

For merchants on the platform, better AI means faster deliveries and higher order volumes. Restaurants that previously lost customers due to long wait times may see improved conversion rates as estimated delivery times decrease. Grab has indicated it plans to share AI-driven insights with merchant partners, including optimal menu pricing suggestions and peak-hour preparation recommendations.

For end users, the impact is straightforward: shorter delivery times, more accurate ETAs, and fewer cold meals. Grab's internal data suggests that improving ETA accuracy by just 3 minutes correlates with a measurable increase in customer reorder rates.

For riders and drivers, the AI promises more efficient shift utilization. By reducing dead miles — the distance traveled without an active order — the system aims to increase rider earnings per hour by an estimated 10-12%.

Looking Ahead: Grab's AI Roadmap for 2025 and Beyond

Grab has signaled that food delivery optimization is just the beginning of its custom AI rollout. The company is reportedly exploring several adjacent applications:

• Generative AI for customer support, using fine-tuned LLMs to handle order disputes and refund requests in local languages including Bahasa, Thai, and Vietnamese
• Computer vision for food quality verification, analyzing delivery photos to detect order accuracy issues before customers report them
• Predictive maintenance models for its growing fleet of electric delivery vehicles
• Dynamic pricing AI that adjusts delivery fees based on real-time supply-demand conditions while maintaining fairness constraints

The company is also expanding its AI research team, with active hiring for machine learning engineers in Singapore, Beijing, and Bangalore. Job postings indicate a focus on large-scale recommendation systems and causal inference models — suggesting Grab is building toward even more sophisticated personalization across its superapp ecosystem.

As Western tech giants like Google, Meta, and OpenAI dominate the global AI narrative, Grab's deployment serves as a reminder that some of the most impactful AI applications are being built far from Silicon Valley — in markets where operational complexity demands solutions that no general-purpose model can provide out of the box. The company's bet on custom, region-specific AI could well define the next chapter of the food delivery wars in one of the world's most dynamic digital economies.