What Mystore actually did

India's Mystore marketplace helps small sellers list products through the ONDC network. In a Google Cloud customer story, co-founder Rajiv Kumar says product approval, previously taking one to two days, now takes less than an hour. He also estimates a 60–70% improvement in the operational efficiency of that process. These are company-reported figures in a cloud vendor's case study, not independently audited results or a forecast for another retailer.

The task is broader than producing attractive copy. Sellers upload uneven photographs and incomplete information; the marketplace needs a title, category, attributes and usable image, followed by a check against catalog rules. According to Google Cloud, Mystore uses Gemini through Vertex AI to propose titles, descriptions, categories and Indian HSN product codes from a single image. A separate tool removes distracting backgrounds and centers product photos. Quality and ONDC compliance checks are part of the approval workflow.

There is an important boundary to the evidence. The case study does not disclose the exact auto-approval rate, how the 60–70% figure was calculated, the cost per listing or error rates for mandatory attributes. “Less than an hour” is Mystore's reported process time, not a guaranteed latency for every item. A Russian retailer should borrow the process design rather than copy these numbers into its budget.

Break the work into testable steps

A product listing is not a single block of prose. Some fields can be checked with rules, some extracted from a picture or supplier document, and some require an accountable employee's judgment. A practical workflow for a smaller catalog is:

  • Collect original photos, SKU, barcode, supplier data, existing category and the marketplace's required fields. Preserve the originals and the version of the rules.
  • Match the item against existing SKUs to prevent duplicates. Check SKU, barcode and exact manufacturer name deterministically; use similar images or descriptions only as a review signal.
  • Ask a model to propose a category, concise title and attributes in a strict structure. Record evidence for every value: a document excerpt, supplier field or visible feature in the photo. Leave uncertain fields blank.
  • Apply validation rules for units, ranges, required fields, prohibited claims, category compatibility and marketplace-specific restrictions.
  • Route doubtful listings to an editor. A human approves publication or changes to price, inventory and legally consequential characteristics.

This pipeline differs from asking a chatbot to “write a description from this photo.” A model may confidently name a material, wattage or included accessory that cannot be seen. The GAVEL study published by the ACL treats attribute extraction as a separately measurable problem; its results apply to its own products and languages and do not establish Mystore's accuracy or that of any local pilot. A retailer needs a labeled test set drawn from its own catalog.

Where local AI fits

Mystore uses Google cloud services; its case does not show that the reported outcomes can be reproduced on premises. The workflow itself can be moved into a controlled environment when supplier photos, purchasing terms or internal reference data cannot be sent to an external API. That requires access to the catalog and listing queue, storage for original files, an attribute taxonomy, a decision log, role-based access and a runtime for a multimodal model. Hardware should be selected after measuring real images and expected throughput, not from a single benchmark.

RAG is useful for retrieving current listing policies, supplier instructions and category restrictions, not for guessing an item's color from its photo. Attach the retrieved rule version to each decision. If the catalog is modest and its rules fit a structured table, a conventional reference database and validator are simpler and more reliable than RAG. An autonomous publishing agent is unnecessary at the start: the model prepares a draft, rules check it, and an employee approves it.

The integration with a PIM, CMS or marketplace should separate reading from writing. Initially the system creates a proposal in a separate queue; only after approval does it send permitted fields through the normal API. Use an operation identifier to prevent repeats, log original values and provide a way to revert a listing change. For products with mandatory labeling, certificates, medical claims or other regulated properties, filling fields automatically without documentary evidence is especially risky.

Economics without borrowing someone else's percentages

The key pilot metric is cost per accepted listing that needs no correction, not cost per model response. Measure staff time for the original process, review time for drafts, the share returned for manual rework, and computing, integration and maintenance costs. Track errors discovered after publication separately: returns, marketplace moderation and catalog corrections.

Illustrative calculation, not a Mystore result: if a worker takes eight minutes to create a listing manually and three minutes to review an AI draft, 1,000 accepted listings save about 83 hours before rework, infrastructure and error costs. If half the drafts need another six minutes of work, that adds 50 hours and sharply reduces the benefit. Use your measurements rather than the 60–70% in another company's case.

Compare three approaches on the same sample: today's manual workflow, simple rules and templates, and a model plus rules. If deterministic rules already cover most routine listings, adding an LLM may not pay back. If volume is low and data is not sensitive, a cloud call may be cheaper than an always-on local server. If volume is steady and data must remain inside, calculate hardware utilization and the cost of maintaining the local environment.

A pilot that can finish in two weeks

Pick one product category and 200–300 previously processed listings with original photos and verified attributes. Reserve some items for final testing and do not use them while tuning prompts. Define critical fields and unacceptable errors in advance. During week one, test extraction and validation without writing to the production catalog; during week two, let an editor compare proposed drafts with the source material.

The result should be a table, not just attractive examples: accuracy for each mandatory attribute, share of listings accepted without manual correction, average and worst-case processing time, duplicates, cost per accepted listing and reasons for rejection. Measure image changes separately: making a photo prettier must not conceal the actual condition of a product. Decide whether to expand using these metrics and the error log.

The lesson is straightforward: speed comes not from a model alone but from combining extraction, verifiable rules and accountable approval. Mystore shows this can materially improve a workflow. A small Russian business is better served by starting with one category and controlled drafts than by immediately entrusting its entire catalog to AI.