Rakuten RMS AI Assistant: merchant data helps, but approval stays human

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A marketplace seller can spend hours writing product copy, preparing images, answering customers and interpreting store data. Rakuten’s RMS AI Assistant addresses those repetitive tasks inside the Rakuten Merchant Server, the operating console for Rakuten Ichiba stores in Japan. The useful lesson is not an autonomous seller. It is structured merchant input making preparation faster while the seller keeps authority to check, edit and send.

A bounded merchant-console assistant

Rakuten introduced RMS AI Assistant as a beta from 28 March 2024. The launch described product-description drafting, background processing for product images, drafted customer-inquiry replies, explanations of year-over-year sales trends and merchant help/chat. Some beta features used OpenAI GPT-4; Rakuten did not say every feature did. Its reference to 57,000 merchants described the target population, not adoption or productivity. 1

The workflows remain narrow. R-Storefront drafts copy from a product name, images and basic product information, and proposes contextual image backgrounds. R-Messe combines a shopper’s inquiry with a short merchant-written response outline to create a reply draft. R-Karte explains patterns in visitors, average order value and conversion rate. The retained evidence does not show the assistant sourcing authoritative facts, changing price or stock, publishing a listing, sending a commitment or executing payment. 1

For merchant help, Rakuten retained a scenario-based bot, added embedding-based semantic search over “Store Operation Navi” documentation and added generative dialogue. The engineering account says design, testing, quality assurance and security review took about three months, but discloses no retrieval accuracy, model version, corpus-refresh policy or live error rate. 2

Preparation is not authority

A practitioner walkthrough describes description output being checked and edited, and an R-Messe reply being checked, edited, copied into the response field and then sent by the merchant. This is secondary corroboration, not a contractual product policy. Rakuten’s later merchant guidance explicitly says generated text and images need human checking, particularly for incorrect facts and regulated health, cosmetics and advertising claims. 3 4

The product-manager interview reported that backgrounds can fail and recommended one complete, unobscured product with margin around it and no overlaid text. It described seven background options and roughly 30-second generation. These are operational observations, not a quality benchmark. 5 Group-level responsible-AI governance describes an AI and Data Committee and a 2024 ethics code, but does not establish RMS-specific retention, training-use, approval-log or human-override terms. 6

Adoption and merchant-reported outcomes

Rakuten’s adoption disclosures indicate use but do not form a clean time series. Around ten months after launch, a product manager said more than half of stores had used the tool and over 20% were established users. A March 2025 corporate article reported about 64% ever used and more than 20% daily; an April interview reported approximately 65% ever used and about 35% almost daily. Definitions, eligible-store denominators, windows and feature mix were not published, so these are directional usage disclosures rather than comparable productivity KPIs. 5 7 8

Named cases are more concrete but remain field testimony. Yuhido reported using RMS image backgrounds on non-hero images and comparing a prior two-to-three-week photographer and preparation process with about five minutes using AI. No image count, quality or rework rate, cost calculation, test period or sales counterfactual was published. 9

Momofuku’s Petit Gift momo-fuku reported reducing inquiry work from roughly 10 minutes to 1–3 minutes per case and from more than 60 hours to about 20 hours per month; another report gives 67 to 20 hours. It also reported expanding registrations from 1,000 to more than 10,000 using description generation, without a stated period. These figures have no control group, quality audit, staffing context or revenue result. 10 11

A December 2025 trade report attributed to Rakuten said 17,000 stores used the tool routinely and about half of 6,000 surveyed users felt efficiency had improved. It relayed anonymous claims including 60 and 17 hours saved per month, fivefold description efficiency and more than six hours saved on replies. Field dates, recruitment, response rate, task denominator and calculation method were not published. 9

By July 2026, coverage called the expanded suite “Rakuten AI for RMS,” reporting use rising from about 25% in December 2024 to 50% in May 2026. A data-analysis agent, available from 30 April 2026, answers merchant questions over store sales data and proposes next questions. A cited 72.9% engagement figure has an ambiguous denominator and is not an accuracy, revenue or autonomous-action metric. 10 11

An illustrative governed test

For a distributor or retailer, start with one product family. Provide canonical identity, approved attributes, images, locale, price and stock references, plus the merchant’s content or response intent. Let the assistant draft, but keep the product owner authoritative for facts and the seller authoritative for publication and communication. This is an illustrative application, not a reported deployment.

Measure approved-page rate, unsupported-claim rate, image rework, reply handling time, escalation and correction separately. Preserve source inputs, generated draft, edits, approver and send or publish time. Test missing attributes, regulated claims, near-duplicate variants and misleading images. The credible progression is a supervised action queue and merchant-initiated analysis—not autonomous catalogue publication, negotiation, pricing, customer messaging or payment.

Sources and next reading

• RMS AI Assistant in Rakuten Merchant Server — Public launch description for Rakuten Ichiba merchants.
• Rakuten AI University — Merchant learning destination; content may require access and is Japanese-language.
• Rakuten EC AI operations guide — Public guidance, not a merchant-console login or full specification.

Explore OfferSignal · Focused Business & AI Workshop · All research

References