The New AI Stage for Cross-Border B2B Platforms: From Connecting Transactions to Completing Transaction Capabilities

Alibaba.com’s Accio Work can be understood as a deployment model for a “seller operating agent”. Its entry point is to organize AI capabilities around the continuous operating process after small and medium-sized sellers enter the global market.

The first step is data and business system integration. Cross-border sellers’ operating activities are scattered across multiple systems and objects, including products, stores, customer inquiries, orders, inventory, logistics, marketing, and compliance documents; if these data and tools cannot be integrated in a unified way, agents can only remain at the level of general Q&A and cannot enter the real operating process. The reason Accio Work can be positioned by Alibaba.com as an agentic business team for global SMEs is that it is able to stand on top of the B2B data, industry knowledge, store operating information, and business tools already accumulated by the platform, and bring operating materials that were originally scattered across different backends and processes into an AI-callable work environment.

The second step is operating context construction and task intent understanding. After completing data and business system integration, the system needs to organize diverse information into an operating context that can support business judgment, and understand the seller’s natural-language goals within this context. The key here is to let the system know the seller’s current operating state, and then judge what executable tasks a vague goal should be decomposed into. For example, with the same request, “help me find ways to improve this product’s performance in the European market”, the tasks the system should break it into will vary across different categories, inventory levels, inquiry performance, and competitive environments: sometimes the priority should be to check product information and keywords; sometimes to compare competitors; sometimes to generate marketing materials; and sometimes to flag inventory or lead-time risks. In other words, the key capability of a seller operating agent is to place the operating goal expressed in natural language back into a concrete business context, and further translate it into a standard task chain that machines can schedule.

The third step is agent orchestration and workflow execution. After tasks are decomposed, the system needs to further determine which agent or capability module should handle which task, how task dependencies should be arranged, which steps need to call platform data, which steps need to connect external tools, and which steps need to wait for human confirmation. The fact that Accio Work is described as an AI agent team rather than a single assistant is important here: it does not hand all operating problems to a generic dialogue box, but compresses the role division of a mature foreign trade team into multiple collaborative agent roles, with the orchestration layer coordinating different tasks such as market analysis, sourcing, store operations, marketing conversion, and inventory monitoring. At this layer, AI truly moves from “operating advice” into “operating execution”: the system not only tells sellers what they should do, but must also be able to push forward concrete actions such as listing optimization, marketing content generation, customer follow-up material preparation, inventory change monitoring, logistics milestone reminders, and compliance document preparation.

The fourth step is permission control and human confirmation. Once agents begin executing actions, they will touch sensitive links such as product information, customer data, transaction documents, payments, and fulfillment. Therefore, deployment must include fine-grained permissions, action approvals, human confirmation, and exception fallback mechanisms. Which actions can be executed automatically, which actions can only generate suggestions, and which actions must be confirmed by the merchant before proceeding must all be defined in advance within the system. Otherwise, the closer agents get to real operations, the more likely they are to cause misoperations, unauthorized access, and unclear accountability.

The fifth step is observability, evaluation, and continuous optimization. After the agentic team goes live, the system needs to continuously track whether each task decomposition is reasonable, whether tool calls succeed, whether generated content is adopted, whether operating metrics improve, whether human intervention is frequent, and whether costs are controllable. Only when these operational data are continuously fed back can the system continue to optimize task decomposition, agent role allocation, prompts, tool calls, and human confirmation nodes.

Therefore, the deployment model of Accio Work can be summarized as a complete technical chain from data integration to operating execution: first, products, stores, customers, orders, inventory, logistics, marketing, compliance, and other data and business systems are integrated; then, the seller’s current operating state is used to build context, and natural-language goals are decomposed into executable tasks; next, multi-agent orchestration assigns tasks to different operating roles, and tool calls push concrete workflow execution forward; finally, permission control, human confirmation, runtime observability, and result feedback ensure that agents can continuously optimize within controllable boundaries. It shows how cross-border platforms can break down the “foreign trade operations team capability” that small and medium-sized sellers originally lacked into a deployable AI workflow.

If mapped to the open-source technology stack, this chain can be broken down into several capability modules. The data and business system integration layer can correspond to Airflow, Dagster, or various connector frameworks, used to synchronize product, order, inventory, customer, logistics, and marketing data and trigger tasks. The business context and knowledge retrieval layer can correspond to LlamaIndex or Haystack, used to connect product materials, store rules, industry knowledge, customer inquiries, and compliance documents, so that agents can understand tasks within a concrete operating context. The task decomposition and agent orchestration layer can correspond to LangGraph, used to build multi-step, stateful, interruptible operating workflows that allow human intervention. Multi-model access and invocation routing can correspond to LiteLLM, used to select suitable models and control costs across tasks such as market analysis, content generation, customer follow-up, and document processing. The tool-calling layer can connect agents to product management, inventory systems, marketing tools, CRM, and file systems through function calling, MCP Server, or custom APIs. The permission governance layer can correspond to OpenFGA, used to define access boundaries across different roles, stores, files, customer data, and transaction actions. The observability and evaluation layer can correspond to Langfuse or Phoenix, used to trace task decomposition, model calls, tool execution, human intervention, costs, and failure nodes. System tracing can correspond to OpenTelemetry, connecting agent behavior with backend system logs, business workflow states, and operating results.

Seen this way, the insight Accio Work offers to cross-border platforms is not simply to copy a seller AI assistant, but to understand the system structure behind it: for a seller operating agent to work, it must simultaneously possess data integration, context understanding, task orchestration, tool execution, permission governance, and operational evaluation capabilities. Only when these modules are organized into the same deployment chain can AI truly move from “answering seller questions” to “undertaking the seller operating process”.

SourcingAI, launched by Made-in-China.com, can be understood as a deployment model for a “buyer procurement agent”. Its entry point is to organize AI capabilities around the procurement decision-making process by which overseas buyers express demand, identify suppliers, compare solutions, and advance orders.

The first step is procurement data and supplier ecosystem integration. Key information in cross-border procurement is scattered across product specifications, supplier profiles, production capacity, business credentials, quotations, lead times, historical transactions, market changes, and platform verification data; if these data cannot be integrated in a unified way, AI can only perform keyword matching and cannot truly enter the procurement judgment process. The reason SourcingAI can be defined as an “AI global sourcing assistant” is that, based on the supplier resources, product data, and transaction ecosystem already accumulated by Made-in-China.com, it can bring information originally scattered across different platform pages, supplier profiles, and procurement communications into an AI-analyzable, comparable, and recommendable procurement environment.

The second step is procurement demand context construction and intent understanding. After supplier and product data integration is completed, the system needs to organize the buyer’s demand description, product specifications, application scenarios, purchase quantity, quality requirements, and delivery conditions into a context that supports procurement judgment. SourcingAI’s official page mentions that it can conduct intent-driven search through text, images, and documents, and transform sketches and technical drawings into solutions that can be directly used for procurement. In other words, the key capability of a buyer procurement agent is not merely to identify what keywords the buyer has entered, but to understand what the buyer truly wants to purchase, what technical conditions this demand corresponds to, which suppliers may match, and which solutions are closer to real procurement execution.

The third step is supplier matching and procurement workflow advancement. After procurement demand is understood by the system, SourcingAI needs to further complete supplier discovery, product matching, quotation comparison, and procurement solution generation. Traditional search often hands a large number of results to buyers for manual screening, while the value of a procurement agent lies in its ability to narrow the range of choices based on demand intent, product specifications, supplier capability, and platform verification information, and to form more comparable procurement alternatives. Made-in-China.com’s official announcement also emphasizes that SourcingAI 2.0 covers intelligent matching, supplier background checks, intelligent price comparison, qualification verification, and order management. This shows that it has moved beyond the “search entry point” and further into “procurement workflow advancement” — the system not only tells buyers which suppliers exist, but also helps buyers judge which suppliers are more worth continuing communication with.

The fourth step is supplier verification and risk control. The closer a buyer procurement agent gets to real procurement decisions, the more it needs to deal with trust and risk. In cross-border B2B transactions, what buyers are most concerned about is often not whether they can find suppliers, but whether supplier qualifications, production capacity, product quality, delivery stability, and cooperation risks are controllable. SourcingAI’s related announcements mention that it provides a more comprehensive view of supplier reliability around business credentials, production capacity, operational scale, and cooperation experience, and supports supplier risk identification in transactions. This design shows that procurement agents cannot only make recommendations; they must also bring supplier verification, risk alerts, and decision evidence into the same chain. Otherwise, AI-generated procurement solutions will struggle to translate into trustworthy orders.

The fifth step is procurement feedback, evaluation, and continuous optimization. After the procurement agent goes live, the system needs to continuously track whether buyer demand is accurately understood, whether recommended suppliers are clicked, inquired about, or adopted, whether quotation comparisons are effective, whether qualification verification reduces communication costs, and whether order advancement becomes smoother. Made-in-China.com has emphasized a 35% improvement in procurement efficiency in the SourcingAI 2.0 release, and related announcements also mention that early users reduced a large amount of manual screening and repetitive inquiry time. For the platform, these feedback signals are not merely promotional proof points, but the basis for subsequently optimizing demand understanding, supplier ranking, risk alerts, price comparison logic, and procurement process design. Only when these operational data continuously flow back can the buyer procurement agent move from a one-off recommendation tool into a continuously optimized procurement decision-making system.

Therefore, SourcingAI’s deployment model can be summarized as a complete technical chain from procurement data integration to procurement decision support: first, products, suppliers, qualifications, quotations, capacity, transactions, and market data are brought into an AI-analyzable procurement environment; then, demand context is built around the buyer’s natural language, images, documents, and technical drawings; next, intent-driven search and intelligent matching transform demand into concrete procurement actions such as supplier discovery, solution comparison, qualification verification, and order advancement; finally, risk identification, procurement feedback, and result evaluation continuously optimize recommendation quality. It shows how cross-border platforms can break down the “procurement judgment team capability” that overseas buyers originally lacked into a deployable AI workflow.

If mapped to the open-source technology stack, this chain can likewise be broken down into several capability modules. The procurement data and supplier ecosystem integration layer can correspond to Airflow, Dagster, or connector frameworks, used to synchronize product libraries, supplier profiles, quotations, qualifications, transaction records, and market data. The multimodal demand understanding layer can correspond to LlamaIndex, Haystack, and multimodal model invocation capabilities, used to process buyers’ natural-language demands, images, documents, and technical drawings, and convert them into structured procurement requirements. The supplier matching and procurement workflow orchestration layer can correspond to LangGraph, used to build multi-step processes from demand parsing, supplier discovery, solution comparison, and inquiry advancement to order management. Multi-model access and invocation routing can correspond to LiteLLM, used to schedule different models and control costs across tasks such as text understanding, image recognition, technical drawing parsing, supplier ranking, and risk analysis. The supplier verification and risk control layer can combine OpenFGA with rules engines, used to define access boundaries for different procurement information, supplier profiles, quotations, and transaction documents, and to control which risk alerts require human review. The observability and evaluation layer can correspond to Langfuse or Phoenix, used to trace demand understanding accuracy, supplier matching quality, price comparison results, human intervention, procurement conversion, and failure nodes. System tracing can correspond to OpenTelemetry, connecting AI recommendation behavior with platform search, inquiries, supplier communication, and order status.

The insight SourcingAI offers to cross-border platforms is to understand the system structure behind it: for a buyer procurement agent to work, it must simultaneously possess supplier data integration, multimodal demand understanding, intelligent matching, procurement workflow orchestration, supplier verification, risk control, and operational evaluation capabilities. Only when these modules are organized into the same deployment chain can AI better move from “helping buyers search for products” to “assisting buyers in forming procurement decisions”.

The most noteworthy budget shift in this issue is that AI investment by cross-border B2B platforms is moving from “improving platform connection efficiency” toward “completing the capabilities buyers and sellers need to complete transactions”. In the past, platform budgets were more often spent on search and recommendation optimization, payment and fulfillment, merchant management, and transaction trust construction, essentially improving the efficiency with which buyers and sellers meet and complete basic transactions; now, as supply becomes increasingly abundant and buyer demand becomes increasingly complex, platforms must further address the capability gaps of both sides of the transaction if they want to continue improving transaction conversion: sellers do not know how to operate continuously, and buyers struggle to judge efficiently. As a result, budget is beginning to land on two categories of capabilities that are closer to transaction outcomes: seller operating capabilities and buyer procurement decision-making capabilities.

The first category of budget becoming real is seller operating capability budget. Alibaba.com’s Accio Work shows that platform investment on the seller side is moving from product publishing, merchant backends, and marketing entry points toward agentic operating capability construction. For many small and medium-sized sellers, entering the platform is only the first step. What truly determines whether they can form sustained transactions is whether they can subsequently complete market judgment, customer follow-up, product operations, inventory management, fulfillment coordination, and compliance processing. In the past, these capabilities often relied on the merchant’s own team, experience, and external tools; now, platforms are beginning to try to compress these capabilities into callable seller operating agents. When platforms invest in this type of capability, what they are essentially buying is a higher seller success rate: enabling more merchants not only to “list products”, but also to operate continuously, respond continuously to inquiries, optimize products continuously, and advance orders continuously.

This type of budget is easier to justify because it can be translated into operating metrics that both platforms and merchants understand. For merchants, it corresponds to listing efficiency, product information quality, inquiry response speed, customer conversion rate, inventory turnover, and operating labor savings; for platforms, it corresponds to the number of active sellers, seller retention, product supply quality, inquiry conversion rate, order conversion rate, and value-added service revenue. In other words, the seller operating agent is not a generic AI assistant budget, but a reorganization of merchant success budget, operating efficiency budget, and seller value-added service budget. Behind it, of course, are data integration, operating context, agent orchestration, tool calling, and permission governance, but the reason these infrastructure costs can be accepted is that they ultimately serve a clearer operating goal: enabling small and medium-sized sellers to turn platform traffic into real orders more effectively.

The second category of budget becoming real is buyer procurement decision-making capability budget. Made-in-China.com’s SourcingAI shows that platform investment on the buyer side is moving from search matching and supplier display toward procurement judgment support. Cross-border B2B platforms have long been solving the problem of “how buyers find suppliers”, but once the number of suppliers and products becomes large enough, the new bottleneck shifts to “how buyers judge who is more suitable”. If buyers have to repeatedly search, inquire, compare specifications, verify qualifications, and assess lead times by themselves, the procurement cycle will be lengthened, and a large amount of potential demand will remain at the browsing and inquiry stage. When platforms invest in procurement agents, what they are essentially buying is higher buyer decision-making efficiency: enabling buyers to express demand faster, narrow down supplier options faster, form comparable procurement solutions faster, and move forward with inquiries and orders with greater confidence.

This type of budget can also be translated into clear operating language. For buyers, it corresponds to sourcing efficiency, supplier matching quality, inquiry cost, procurement cycle, and first-time procurement trust; for platforms, it corresponds to buyer retention, inquiry quality, request for quotation (RFQ) conversion rate, supplier response rate, and transaction matching efficiency. In the past, platforms might have relied on search ranking, category navigation, supplier certification, and manual sourcing services to reduce buyers’ decision costs; now, AI can move demand understanding, supplier matching, qualification judgment, and solution comparison forward into the system, allowing the platform to move from “showing more options” toward “helping buyers form credible choices”. Similarly, SourcingAI also requires multimodal demand understanding, supplier data integration, intelligent matching, risk identification, and operational evaluation behind the scenes, but these capabilities do not need to be abstracted into another separate budget category; they are already embedded in the buyer procurement decision-making capability budget.

Therefore, the directions that are truly beginning to become budgetized in this issue can be narrowed down to two categories: one is seller operating capability budget, serving merchant success and seller operating efficiency; the other is buyer procurement decision-making capability budget, serving procurement efficiency and transaction matching conversion. Together, they show that AI investment by cross-border B2B platforms is moving from “making it easier for buyers and sellers to meet” toward “making buyers and sellers more capable of completing transactions”. What budget purchases is not an isolated AI feature, but a system capability that enables sellers to operate, buyers to procure, and the platform transaction chain to run more smoothly.

The first core friction is the continuous calibration of transaction context. After AI deployment on cross-border B2B platforms truly enters deep waters, the core bottleneck lies in whether these agents can continuously stay close to real transaction scenarios. Cross-border B2B transactions are inherently highly dynamic: product status changes, inventory and lead times change, supplier capacity changes, overseas market rules change, and buyer preferences and seller operating goals also change. Even if agents already have context-building capabilities, they still need to continuously absorb these changes; otherwise, the problem may arise that “the system can execute tasks, but the execution direction is not sufficiently aligned with real business”.

This is especially obvious on the seller side. For example, if a seller wants to improve a product’s performance in the European market, the system needs to judge whether the problem comes from incomplete product information, keyword mismatch, missing certification materials, uncompetitive pricing, or insufficient inventory and lead time; and these judgments depend on product status, market feedback, inquiry performance, and regional rules, all of which may keep changing. The buyer side follows the same logic. Supplier qualifications, capacity, quotations, response speed, and delivery stability will change over time; if AI recommendations cannot absorb these changes in time, they may easily package outdated information as seemingly reasonable procurement advice.

This kind of friction will not automatically disappear once agents go live. The deeper agents enter operating and procurement processes, the higher the requirements for contextual accuracy, timeliness, and category adaptability. In early AI search or copywriting tools, incomplete context may only affect answer quality; but when AI begins to decompose operating tasks, recommend suppliers, compare quotations, flag compliance risks, or push workflows forward, contextual deviation will directly affect transaction judgment. Once the problem becomes systematized, it will turn into continuous costs in data governance, category knowledge maintenance, regional rule updates, feedback evaluation, and human calibration. In other words, transaction context is not a one-off data project, but a long-running business calibration mechanism.

The second core friction is that platform responsibility boundaries become more complex after AI intervenes in transaction judgment. In the past, cross-border B2B platforms mainly played the role of connection, display, search, inquiry, and transaction trust infrastructure. How sellers operate and how buyers judge were mainly borne by the two sides of the transaction themselves; the platform provided information entry points and basic safeguards. But when seller operating agents begin to advise merchants on how to optimize products, follow up with customers, handle inventory, and prepare compliance documents, and buyer procurement agents begin to help buyers screen suppliers, compare quotations, judge qualifications, and form procurement solutions, the platform is no longer merely presenting information, but is participating more deeply in the judgment and action process within transactions.

This will bring new governance pressure. Which content is only advice, which actions can be advanced automatically by the system, which must be confirmed by the seller or buyer, which risks need to be prominently flagged, and which judgments must retain human review all need to be redefined. For example, if AI recommends that a certain supplier better matches a buyer’s needs and a delivery problem subsequently arises, the platform needs to clarify whether this is an algorithmic recommendation, a platform endorsement, or an independent buyer decision; when AI helps sellers generate compliance materials or marketing expressions, it also needs to clarify which content can be generated automatically and which content requires the merchant to confirm its truthfulness and applicability. As AI gets closer to operating and procurement decisions, these boundaries will become increasingly important. What platforms need is not simply an added disclaimer, but the embedding of permission control, human confirmation, risk alerts, audit trails, and model evaluation into agent workflows.

This is also where NextAI+ can play a role. For seller operating agents, NextAI+ can help platforms determine which operating processes are most suitable for prioritized agentification. There are many seller-side operating tasks that AI can enter, including product publishing, title optimization, inquiry response, customer segmentation, inventory reminders, logistics notifications, compliance material preparation, and marketing content generation, but not all of them are suitable as first-batch deployment scenarios. The processes most suitable for early prioritization usually need to meet three conditions at the same time: first, high frequency, meaning they can cover enough sellers’ daily operating actions; second, controllable risk, meaning that even if AI output is imperfect, risk can be reduced through human confirmation or rule checks; third, measurable value, meaning their effect can be verified through inquiry response speed, product information completeness, content generation efficiency, conversion rate, or operating labor savings. NextAI+ can help platforms rank these operating processes by value, risk, data readiness, and implementation difficulty, and design automation boundaries and value verification methods, avoiding the choice of complex scenarios that look ambitious but are difficult to validate in the short term.

For buyer procurement agents, NextAI+ can help platforms sort out the procurement decision-making chain, and break demand understanding, supplier matching, qualification verification, quotation comparison, and risk alerts into testable and evaluable task modules. The core is not to let AI make final procurement judgments on behalf of buyers from the very beginning, but to first identify which links are suitable for AI-assisted screening, ranking, prompting, and inquiry drafting, and which links must retain human confirmation. In this way, platforms can validate the real value of procurement agents through supplier matching quality, inquiry conversion rate, procurement cycle, and buyer adoption rate while keeping risks under control.

Therefore, the role of NextAI+ is to help platforms determine which seller-side operating processes should be prioritized for agentification, and which buyer-side procurement steps are suitable for AI to enter first. It turns complex cross-border trade operations into a set of deployment decisions that can be prioritized, validated, and scaled.