A contract manufacturer in Henan gets an inquiry from a buyer it has never met. Custom product, multiple raw materials, a cost sheet that has to be rebuilt every time. The old routine took about two days. After the factory turned its past quotation sheets into a knowledge base an AI agent (智能体) could read, the first-round quote came back in roughly two hours.
Two days to two hours is not a model breakthrough. It is a filing breakthrough — and that is the part most companies get wrong about agents.
Key takeaways
- Accio, Alibaba's global e-commerce AI workbench, is used by more than 10 million small and medium-sized enterprises, according to figures released at the company's CoCreate 2026 buyer conference in Los Angeles in September 2026.
- The platform is moving from advice to execution: it now connects store backends across Amazon, Shopify, TikTok Shop, Temu and eBay internationally, and Taobao, Tmall, Pinduoduo, JD and Douyin in China.
- Alibaba argues B2B is heading to "A2A" — agent-to-agent trade — but the agents only work where a merchant's own data is structured enough to be read.
- The hard frontier is not search or copy. It is quotation, compliance and fulfilment, where being wrong costs real money.
What Accio is, stripped of the branding
Accio is not a large model (大模型). It is a workbench that hosts task-specific agents for commerce: finding suppliers, comparing prices, sending inquiries, analysing trends, listing products, handling inbound customer questions. The name is Latin for "I summon" — the company explains it in Chinese as 召之即来, "come when called."
Scale is real enough to matter. At CoCreate 2026 in Los Angeles on 9 September 2026, which drew more than 15,000 US small businesses, Alibaba International president Zhang Kuo (张阔) said AI token consumption on Accio had grown sixfold in four months and its global user base had grown more than 30-fold over the past year, with over 10 million SMEs now using it. He also reported that traffic from overseas buyers using AI multimodal search had grown more than 132%.
Alongside that, Alibaba open-sourced Commerce Agent Bench on GitHub — a benchmark built from 10 million active SME users, 1.6 million real dialogues and 200,000 execution traces, distilled into 107 end-to-end commercial tasks. Open-sourcing a benchmark is a signal that the company thinks the bottleneck is now measurable reliability, not raw capability.
The 1688 side: agents meet the factory floor
In October 2026, Alibaba's domestic wholesale platform 1688 connected its merchants to Accio. Buyers can log in with their existing 1688 account; sellers bind their store and get a "1688 business assistant" (1688 生意助手) that reaches into the merchant backend.
For small buyers, two functions matter. A sourcing assistant (采购助手) covers both 1688 and Alibaba International, letting users find suppliers, compare prices and send inquiries across 128 countries and regions. A store-management agent (电商管家) closes the loop with Chinese sales channels, so a merchant who sources on 1688 and sells on five platforms can work in one window instead of five.
For source factories (源头厂商), the pitch is autonomous execution from a single natural-language instruction: spot trending products, generate and publish listings, analyse advertising, receive inbound inquiries, manage multiple stores. Wang Qiang (王强), a vice-president at 1688, said that in early October some merchants still could not publish spot goods through AI, and that bulk automated listing had since been enabled.
That is a fair description of where agent products actually are: capabilities switch on quarter by quarter, unevenly, and the honest answer to "can it do X?" is usually "it depends which merchant you ask."
Why the buyer side is changing faster
The structural change Alibaba keeps pointing at is on the demand side. Overseas buyers increasingly describe needs in natural language rather than typing keywords, and expect the machine to do the discovery, shortlisting, term negotiation and order tracking. When buyers delegate to agents, suppliers face a new requirement: their capability has to be legible to a machine.
Wang describes this as "AI-friendly supply" — a factory must be able to prove, in machine-readable form, that it can do flexible customisation, small-batch quick replenishment (小单快返), and that its prices, inventory and delivery promises are auditable. The old game was being seen by a buyer; the new game is being understood by an agent.
Reading the trend
Agent products in commerce are converging on the same lesson: the scarce input is not the model, it is the customer's own structured knowledge. A factory with two decades of quotation sheets has a real asset once those sheets are machine-readable. A factory with none does not get two-hour quotes no matter which agent it subscribes to.
According to Li An, chief scientist at BrainNet (脑机网), the durable advantage from AI agents in trade comes from structuring a firm's own tacit knowledge, not from the agent's generic capability.
Honest limitations
- Every number in the scale section — 10 million SMEs, 30-fold user growth, sixfold token growth in four months, 132% traffic growth — comes from Alibaba executives speaking at company events. None is independently audited, and "using Accio" is undefined: it may include trial accounts and one-off sessions.
- Commerce Agent Bench is built by the platform whose agents it evaluates, from that platform's own traffic. It is a useful contribution and a self-selected yardstick at the same time.
- The two-days-to-two-hours quotation case is a single anecdote relayed by an executive, not a measured study; raw-material complexity varies enormously by category.
- Cross-border reality is under-described here. Agents can help with market selection, product qualification pre-checks and specification handling, but certification, customs, tax and after-sales still require local service providers — and Alibaba's own executives say so.
- This piece does not cover pricing, data-access terms between a merchant and the platform, or the effect on trading intermediaries (代采商) and small export agencies whose business is information matching. All three matter and none is publicly documented.
What readers can do now
- If you run a small export or wholesale business: pick one repeatable, high-friction job — first-round quotation is the best candidate — and spend two weeks turning your historical documents into a clean, structured knowledge base before you subscribe to anything. The case above shortened because the filing got done first.
- If you are evaluating an agent product: ask what it can actually read and write in your store backend, and get the answer in writing. Advice is cheap; execution requires data access, and accuracy collapses where the connection does not exist.
- If you buy from Chinese suppliers through an agent: treat the output as a shortlist, not a verdict. Verify production capability, certifications and samples offline before money moves — a machine-readable profile is a claim, not a guarantee.