AI shopping assistants are replacing the familiar "search box plus shelf" model and becoming the new front door to e-commerce. For Chinese sellers operating in Kazakhstan and Central Asia, the contest is no longer about who buys the most keywords, but about whose product information AI can understand and is willing to recommend.
In the United States the shift can already be measured. In its fourth-quarter 2025 results, Amazon said its AI shopping assistant Rufus had about 300 million users over the year and generated nearly 12 billion dollars in incremental annualized sales. CEO Andy Jassy told the earnings call that customers who use Rufus are about 60% more likely to complete a purchase. For a platform built on search and shelves, this shows that conversational shopping guidance has moved from experiment to genuine source of revenue.
Platforms in Kazakhstan are taking the same road. On 1 July 2026, Kaspi.kz unveiled its AI assistant Kasper in Almaty. Users can describe what they need by voice or text in Kazakh or Russian, for example "I have carpets, tiles and a cat at home, help me choose a vacuum cleaner", and Kasper asks follow-up questions, compares products and explains the differences in plain language. It launched first in Almaty and will roll out gradually nationwide. Given the size of the market, this is a weighty step: according to the Bureau of National Statistics, Kazakhstan's retail e-commerce reached about 3.77 trillion tenge in 2025, or 14.3% of total retail, with about 86% going through marketplaces. Such high concentration means that once an on-platform AI assistant becomes a habit, sellers' traffic sources will be rewritten along with it.
A former member of Amazon's Rufus team, speaking on the Chinese platform Biji Xia, offered a useful framework for thinking about this change. Several of his points are worth testing against their own business by anyone selling online in Central Asia.
First, define the problem and its boundaries before deciding whether to build an agent. In the speaker's view, Rufus worked because from the start it did only one thing, shopping guidance: narrowing down a vast catalogue, explaining differences and offering a handful of options. He sees shopping as two problems: too much choice and too little trust. Interestingly, Kaspi's Mikheil Lomtadze described almost the same logic when presenting Kasper: start with one concrete task, helping customers pick the right product on the marketplace, and then expand to other everyday tasks.
Second, an agent can be broken into three parts: a strong model, context, and multi-step execution. The model handles understanding and expression; context is long-term information relevant to the task, such as household situation, purchase history and budget; multi-step execution breaks a goal like "get me fully kitted out for camping" into steps: list, search, compare, filter, combine. The speaker stressed that more context is not better; what matters is retrieving the right information at the right time. Nor are more steps better: each extra step adds cost and a chance of error.
Third, reliability is worth more than cleverness. In his view, it is easy for an AI product to impress on first use; the hard part is getting people to come back. An assistant that often makes things up or simply tells users what they want to hear quickly loses trust, while in shopping people need exactly the opposite: why something fits, why it does not, and what the risks are. For any platform or seller wanting to do AI-guided shopping, this is a simple but easily overlooked standard.
Fourth, the logic of advertising will shift from buying keywords to making AI understand your product. The speaker noted that Amazon is exploring ad mechanisms in conversational settings, and advertised products may in future get extra weight in AI recommendations. Sellers will compete not only for human attention but for AI's understanding of their products, and AI will increasingly look at return rates, reviews and real purchasing behaviour, not just the seller's own description.
Fifth, A2A, or agent-to-agent commerce, is drawing closer. The buyer's agent understands the need, the seller's agent supplies product, stock, after-sales and usage-limit information; the two negotiate first and a human makes the final call. The speaker advises sellers not to rush into building a full system of their own, but to start with the platforms' native tools and keep feeding them product knowledge and service rules.
Platforms' attitude to outside agents shows how fierce the fight for the front door has become. In November 2025, Amazon sued Perplexity in federal court in San Francisco, demanding that its AI browser Comet stop placing orders on Amazon on users' behalf, arguing among other things that Comet did not identify itself as an automated agent. In August 2026, the US Court of Appeals for the Ninth Circuit vacated the lower court's preliminary injunction against Comet, finding that it is the user who accesses Amazon; the case is ongoing. The speaker's explanation is that a platform's most valuable asset is real people's click, dwell-time and add-to-cart data, and machine activity mixed in distorts the recommendation system. That, however, is his personal view. For Central Asia, it is reasonable to expect Kaspi, Wildberries, Ozon and others to promote their own assistants first rather than open their doors to outside agents.
Rules related to this news
- Kazakhstan China investment forum Almaty 2026On 25 September 2026 the Kazakhstan–China investment forum took place in Almaty: over 500 business representatives att…Text and legal basis →
- Investment incentivesFrom 1 January 2026 the former investment priority project regime is replaced by three instruments: (1) investment agr…Text and legal basis →
- How do I set up a logistics company in Almaty and what permits are needed for freight transport?Usually you first register an LLP (TOO): a private-business legal entity is registered within one working day of an e-…Text and legal basis →
Sixth, write product information for AI. Product pages used to win clicks with attractive copy; in front of an AI assistant, vague slogans do not help. The more clearly features, specifications, target users, conditions of use, limitations and service boundaries are spelled out, the more accurately AI can recommend. This matters especially for Chinese cross-border sellers: many listings are still machine-translated into Russian with missing specifications, and Kazakh versions are even scarcer.
Seventh, companies can adopt AI internally in three stages. The speaker's route: first use leading closed-source large models to verify whether AI can actually solve the business problem; once it works, switch to open-source large models to control cost and data; finally, fine-tune open-source small models for vertical use, doing just a few things well enough. His analogy: you do not need Einstein to brew a cup of coffee.
For Chinese sellers in Kazakhstan, a few steps can be taken right away.
One, rewrite product cards. In correct Russian and Kazakh, set out specifications, sizes, materials, and the situations a product suits and does not suit; avoid descriptions full of bold selling points but short on facts, so that assistants like Kasper can answer customers' follow-up questions accurately.
Two, manage reviews and return rates as ranking assets. AI takes real feedback into account, and fake reviews or high return rates may quietly push a product out of recommendations.
Three, watch platform tools closely. When Kaspi, Wildberries or Ozon launch new AI or advertising features for sellers, trying them early and building up data is better value than building a system from scratch.
Four, start small with internal AI. Pick tasks with clear boundaries, such as customer-service answers or translating product cards, validate results with a mature large model, and only then think about cutting costs and customization.
Five, do not bet everything on a single entry point. The front door is moving from apps to conversation, and today's traffic rules may not last three to five years. Keep a multi-platform presence and turn product data into a structured, reusable asset.
Sources:
Bijixia (WeChat),https://mp.weixin.qq.com/s/qlZI7AtnMq1K8RrvO_ktbA
MLQ.ai,https://mlq.ai/earnings/highlight/AMZN-jassy-highlights-rufuss-impact-on-custo-33d375/
Modern Retail,https://www.modernretail.co/technology/amazon-says-its-ai-shopping-assistant-is-gaining-traction-with-rufus-users-up-115/
PPC Land,https://ppc.land/amazons-ai-shopping-assistant-drove-12-billion-in-sales-for-2025/
Forbes Kazakhstan,https://forbes.kz/articles/mikhail-lomtadze-unveils-kasper-apersonal-aiassistant-from-kaspikz
Forbes Kazakhstan,https://forbes.kz/articles/why-kasper-was-first-launched-in-the-kaspikz-marketplace-a260bd
GlobeNewswire,https://www.globenewswire.com/news-release/2026/07/14/3327084/0/en/Watch-the-Kaspi-Event-Mikheil-Lomtadze-Unveils-Kasper.html
Qazinform,https://qazinform.com/news/one-million-customers-have-used-kasper-kaspikzs-personal-ai-assistant-7abec8
Bureau of National Statistics of Kazakhstan,https://stat.gov.kz/en/industries/economy/local-market/publications/488906/
Inform.kz,https://www.inform.kz/ru/kazahstantsi-sdelali-bolee-291-mln-zakazov-na-marketpleysah-pochti-vdvoe-bolshe--cd44c785
The Tenge,https://the-tenge.kz/articles/obem-elektronnoi-torgovli-v-2025-godu-sostavil
Bloomberg,https://www.bloomberg.com/news/articles/2025-11-04/amazon-demands-perplexity-stop-ai-agent-from-making-purchases
CourtListener (complaint),https://storage.courtlistener.com/recap/gov.uscourts.cand.459191/gov.uscourts.cand.459191.1.0.pdf
CourtListener (Ninth Circuit order),https://storage.courtlistener.com/recap/gov.uscourts.cand.459191/gov.uscourts.cand.459191.114.0.pdf
Justia,https://law.justia.com/cases/federal/appellate-courts/ca9/26-1444/26-1444-2026-08-04.html
Key facts
- Amazon disclosed that Rufus had approximately 300 million users in 2025 and generated nearly USD 12 billion in incremental annualized sales.
- Amazon stated that customers who use Rufus are approximately 60% more likely to complete a purchase.
- Kaspi.kz launched Kasper in Almaty on July 1, 2026, before gradually expanding it nationwide.
- Kazakhstan’s National Bureau of Statistics reported that the domestic retail e-commerce market reached approximately 3.77 trillion tenge in 2025, accounting for 14.3% of total retail turnover.
- Approximately 86% of retail e-commerce transactions in Kazakhstan are completed through e-commerce platforms.
- Chinese sellers operating in Kazakhstan should avoid relying on a single entry point and should maintain a presence across multiple platforms.
FAQ
Where was Kasper launched, and which languages does it support?
Kaspi.kz launched Kasper in Almaty on July 1, 2026. It supports Kazakh and Russian and was initially rolled out in Almaty before gradually expanding nationwide.
How might AI shopping assistants change traffic sources for sellers?
The article argues that once users form the habit of using platform AI assistants, sellers’ traffic sources will shift accordingly. AI recommendations may also take return rates, seller reputation, and genuine purchasing behavior into account.
What product information should Chinese sellers in Kazakhstan improve first?
The article recommends rewriting Russian-Kazakh bilingual product listings to clearly state specifications, dimensions, materials, applicable and inapplicable use cases, and the boundaries of after-sales service, while reducing descriptions that rely solely on exaggerated selling points without factual information.
Do sellers need to immediately develop their own AI agent systems?
The article cites a contributor who recommends that sellers need not rush to develop an entire in-house system. They can first use the platform’s native tools while continuing to improve product knowledge and service rules.
Should Chinese sellers concentrate their traffic on a single entry point?
The article recommends maintaining a presence across multiple platforms and organizing product data into structured, reusable assets to respond to changes in entry points and traffic rules.