Rakuten Deploys Custom AI for E-Commerce Across Asia

Rakuten, Japan's largest e-commerce conglomerate, is deploying a suite of custom-built AI models designed to hyper-personalize shopping experiences across its platforms in Asia. The initiative marks one of the most ambitious AI rollouts by a non-US tech giant, positioning Rakuten to compete directly with Amazon's and Alibaba's AI-driven commerce engines.

The company's new Rakuten AI Engine (RAE) leverages proprietary large language models fine-tuned on over 1.8 billion transaction records spanning 15 years of consumer behavior data. The system is now live across Rakuten Ichiba in Japan, Rakuten France, and is expanding into Taiwan, India, and Southeast Asian markets throughout 2025.

Key Takeaways at a Glance

• Rakuten has deployed custom AI models across its e-commerce platforms serving over 100 million active users
• The Rakuten AI Engine processes 1.8 billion historical transaction records to generate personalized recommendations
• Early results show a 23% increase in click-through rates and a 14% boost in average order value
• The system uses a multi-modal approach combining text, image, and behavioral data
• Rakuten is investing an estimated $500 million in AI infrastructure through 2026
• The rollout directly challenges Amazon's Rufus AI assistant and Alibaba's Tongyi Qianwen commerce tools

Rakuten Builds Its Own AI Stack From Scratch

Unlike many e-commerce players that rely on third-party AI services from OpenAI, Google, or AWS, Rakuten has chosen to build its AI infrastructure in-house. The company's Rakuten Institute of Technology (RIT), with research labs in Tokyo, Paris, Bangalore, and Singapore, has spent the past 3 years developing foundation models specifically optimized for commerce.

The core of RAE is a family of transformer-based models ranging from 7 billion to 70 billion parameters. These models were pre-trained on multilingual commerce data — including product descriptions, customer reviews, search queries, and purchase histories — across 12 languages commonly used in Rakuten's operating markets.

This approach gives Rakuten a significant edge in understanding the nuances of Asian consumer behavior. Western-trained models often struggle with the complexity of Japanese product naming conventions, the formality levels in Korean customer service interactions, and the visual-first shopping preferences common in Southeast Asian markets.

How the Personalization Engine Actually Works

The RAE system operates across 3 distinct layers, each handling a different aspect of the shopping experience.

Layer 1 — Intent Understanding uses natural language processing to decode what shoppers actually want, even when their search queries are vague or misspelled. The model interprets context from previous sessions, seasonal trends, and regional preferences to surface relevant products.

Layer 2 — Visual Matching employs computer vision models that analyze product images to recommend visually similar items. This is particularly powerful in fashion and home goods categories, where aesthetic preferences drive purchasing decisions.

Layer 3 — Dynamic Pricing and Bundling uses reinforcement learning to optimize product bundles and promotional offers in real time. The system considers inventory levels, merchant margins, competitor pricing, and individual customer price sensitivity.

Key technical specifications include:

• Inference latency under 50 milliseconds for real-time recommendations
• Support for 12 languages with cross-lingual transfer learning
• Processing capacity of 500,000 concurrent personalization requests
• Integration with Rakuten's loyalty point ecosystem (Rakuten Super Points)
• Privacy-preserving federated learning for cross-market insights
• A/B testing framework that evaluates over 200 model variants simultaneously

Early Results Show Significant Revenue Impact

Rakuten reports that early deployments of RAE have delivered measurable business impact. In Japan, where the system has been running in production since late 2024, click-through rates on personalized recommendations jumped 23% compared to the previous rule-based system.

More importantly, average order value increased by 14%, suggesting that the AI is not just driving more clicks but actually helping customers discover higher-value products they want to buy. Return rates in categories served by RAE dropped by 8%, indicating better product-customer matching.

Merchants on the Rakuten Ichiba marketplace are also benefiting. Small and medium-sized sellers using RAE-powered listing optimization tools saw a 31% increase in organic visibility. This is critical for Rakuten's marketplace model, where merchant satisfaction directly correlates with platform growth.

The financial implications are substantial. Analysts at Nomura Securities estimate that RAE could add between $1.2 billion and $1.8 billion in incremental gross merchandise value (GMV) across Rakuten's platforms by the end of 2026.

Rakuten Takes on Amazon and Alibaba in the AI Commerce Race

The deployment of RAE puts Rakuten squarely in competition with the world's 2 largest e-commerce AI initiatives. Amazon launched its Rufus AI shopping assistant in early 2024, which uses conversational AI to help customers research and compare products. Alibaba has integrated its Tongyi Qianwen large language model across Taobao and Tmall, offering AI-generated product summaries and smart customer service.

Rakuten's approach differs from both competitors in a fundamental way. While Amazon and Alibaba focus heavily on conversational AI interfaces — essentially adding chatbots to the shopping experience — Rakuten is embedding AI deeper into the recommendation and discovery infrastructure itself.

'We believe the future of e-commerce AI is not about adding a chatbot to the search bar,' said a Rakuten spokesperson in a recent press briefing. 'It is about making every pixel on the page intelligent and personalized to the individual shopper.'

This philosophy aligns with research from McKinsey showing that 71% of consumers expect personalized interactions, but only 15% want to interact with a chatbot during their shopping journey. Rakuten appears to be betting that invisible, ambient AI will outperform the conversational approach.

Privacy and Data Governance Take Center Stage

Operating across multiple Asian markets with varying data protection regulations presents unique challenges. Japan's Act on Protection of Personal Information (APPI), South Korea's PIPA, and the EU's GDPR (applicable to Rakuten France) all impose different requirements on how customer data can be collected, processed, and used for AI training.

Rakuten addresses this through a federated learning architecture that keeps raw customer data within each market's borders. The AI models learn from aggregated, anonymized patterns rather than individual-level data, allowing cross-market insights without cross-border data transfers.

The company has also established an AI Ethics Board comprising external academics, consumer advocates, and privacy experts. This board reviews all AI deployments for potential bias, particularly in pricing algorithms that could inadvertently discriminate against certain customer segments.

These governance measures could become a competitive advantage as regulators worldwide tighten oversight of AI in consumer-facing applications. Companies that build compliance into their AI infrastructure from the ground up will face fewer disruptions than those retrofitting governance after deployment.

What This Means for the Global E-Commerce Industry

Rakuten's AI deployment signals several important trends for the broader e-commerce and AI industries.

First, the era of one-size-fits-all AI solutions for commerce is ending. Regional players with deep local data are proving that custom models can outperform general-purpose AI tools from Western tech giants. This validates the 'small but specialized' approach to AI model development.

Second, the competitive battleground in e-commerce is shifting from logistics and pricing to AI-driven personalization. Companies that fail to invest in intelligent recommendation systems risk losing market share to platforms that understand customers better.

Third, the investment required is substantial. Rakuten's estimated $500 million AI infrastructure spend through 2026 is significant even for a company with $15 billion in annual revenue. Smaller e-commerce players may need to rely on AI-as-a-service platforms rather than building proprietary systems.

For developers and AI engineers, Rakuten's multi-modal, multi-lingual approach offers a compelling case study. The combination of NLP, computer vision, and reinforcement learning into a unified commerce engine represents the kind of integrated AI system that is increasingly demanded across industries.

Looking Ahead: Rakuten's AI Ambitions Beyond E-Commerce

Rakuten's AI investments extend well beyond shopping. The company plans to leverage RAE technology across its broader ecosystem, which includes Rakuten Mobile (Japan's 4th mobile carrier), Rakuten Bank, Rakuten Travel, and Rakuten Viber (a messaging platform with 820 million users).

The cross-ecosystem play is where Rakuten's strategy becomes particularly interesting. A unified AI engine that understands a customer's shopping preferences, travel habits, financial behavior, and communication patterns could deliver personalization at a scale that even Amazon and Google would struggle to match.

The company has indicated that a next-generation RAE model, expected in early 2026, will incorporate generative AI capabilities for automated product description creation, virtual try-on experiences, and AI-powered merchant analytics dashboards.

As the AI commerce race intensifies across Asia and beyond, Rakuten's custom model approach offers a blueprint for regional players looking to compete with global tech giants. The key question is whether proprietary, locally-trained AI can sustain its performance advantage as foundation models from OpenAI, Google, and Meta continue to improve. The next 18 months will be decisive.