Anta Teams Up with Volcano Engine: How Agents Are Running on the Front Lines of Global E-Commerce

When AI shifts from "answering questions" to "getting things done," Anta offers a real-world sample of enterprise-grade deployment through three Agents.

Agents are becoming "digital colleagues" in Anta's global business operations. As one of the world's top three sporting goods groups, Anta Group operates multiple brands across multiple markets, backed by massive volumes of products, content, orders, and after-sales processes. Today, these Agents have entered specific business areas including content production, overseas ordering, and cross-border returns, and are gradually being replicated across more brands and markets.

What's interesting is that these Agents are not lab experiments — they emerged from real pain points on the front lines. Through the Volcano Cup Agent Innovation Competition, Anta had business teams raise problems, FDEs assist with process decomposition and development/debugging, and then leveraged Volcano Engine's model and agent capabilities to put the Agents back into actual work for validation.

How Anta Determines Whether an Agent Has Truly Entered Production: Three Metrics

Wu Yongliang, Director of Global IT Delivery and Support at Anta, offers a pragmatic evaluation framework: To gauge Agent progress, truly look at three metrics — first, how many business teams can't function without Agents daily; that's the people dimension. Second, how many business processes now include Agents by default; that's the process dimension. Third, how many Agents have started replicating across countries; that's the regional dimension.

This statement captures the core challenge of enterprise AI deployment: it's not that models aren't smart enough — it's whether Agents can take over a complete workflow. Wu further noted: "Judging whether an Agent has truly entered production isn't about how cleverly it answers — it's about whether it can replace a specific step in a complete process. In the past, AI answered a question. In the future, AI completes a task. The endgame of Agent competition isn't chat — it's workflow."

From "answering questions" to "completing tasks" — this shift seems simple but is actually the key watershed where enterprise AI moves from being a tool to becoming productivity.

FDE (Forward Deployed Engineer) is a common role in enterprise AI deployment, first practiced by companies like Palantir. Unlike traditional R&D engineers, FDEs typically embed themselves on the front lines of client business, working with business staff to map processes and define problems, then translating business language into technical solutions and participating in development, debugging, and iterative launches. In Anta's practice, business teams are responsible for raising real pain points, FDEs decompose vague requirements into executable workflows and rule logic, and then implement them using large models and agent platforms. This "business + engineering" pairing approach significantly reduces requirement transfer loss, ensuring Agents fit real scenarios from the design stage — rather than "building a hammer first and looking for nails" in a lab.

Agent is fundamentally different from traditional conversational AI. Chatbots primarily handle "input question — output answer" single-turn or multi-turn interactions, while Agents emphasize autonomous planning, tool invocation, action execution, and result delivery — typically with a large model as the "brain," coordinating with memory, planning, and tool-calling modules. The recurring phrase "from answering questions to completing tasks" refers precisely to this capability leap: Agents need to understand the context of a complete workflow, invoke different systems and interfaces according to rules, and ultimately produce outputs that business teams can use directly — such as a video, a document, or a purchasing decision.

Three Agent Cases: Breaking Through in Content, Ordering, and Returns

Product Video Agent: New Hires Can Produce Independently in Half a Day

Southeast Asian e-commerce sees rapid product launches, and promotional content needs to keep pace. Previously, short-video operators had to find trends, break down videos, generate prompts, and publish to different sites — multiple steps dependent on manual, repetitive work.

The AI product video Agent connected this entire chain: ArkClaw collects trending and benchmark videos, completes video decomposition and prompt generation, Seedance 2.0 generates product clips, and then multi-site publishing is handled. In terms of results, intern new hires can independently produce qualified product videos in half a day with the Agent's help, and can produce about 15 more videos per day at the same Token consumption — allowing operators to spend more time on topic selection and creativity.

Li Han, short-video operator at Anta's Southeast Asia e-commerce department, admitted: "Southeast Asian sites launch so many new products daily that manual video editing simply can't keep up with the stocking pace. Using AI to batch-produce videos is the key direction for boosting store sales."

Overseas Ordering Agent: Purchasing Decisions Cut from Half a Day to 10 Minutes

Anta's International Business Division deals with four different templates for every order. The underlying data overlaps by about 90%, but item numbers, categories, translations, formulas, and formats still require manual processing item by item.

The product planning team used ArkClaw to encapsulate the four templates into a Skill, automatically completing item number conversion, category filtering, format alignment, and other standardized outputs. Currently, the overseas order automation Agent has shortened the complete purchasing decision from half a day to 10–20 minutes, reduced manual entry by about 90%, and achieved data accuracy exceeding 99%. This Agent now covers both footwear and apparel product lines, with plans to apply it to product line planning, competitor price band analysis, and other scenarios.

Wang Jieyi, product planner at Anta's International Business Division, confirmed the feasibility of this direction: "The four templates share about 90% of underlying data, but every time we had to manually move, translate, and align formulas and formats item by item. There's a lot of repetitive Excel work in product planning, and technical solutions can be reused."

Return Document Agent: 1,000+ Auto-Created Orders Per Month

There's a typical process gap in cross-border returns: after an e-commerce platform order is returned, the platform information can't directly flow back into Anta's system. Customer service needs to move the information to ERP to create a return order, then follow up on each order's status individually. As return volumes increase, this creates massive cross-system repetitive work.

The return document automation Agent filled this gap — automatically pulling return data from e-commerce platforms, generating documents according to ERP templates, and periodically summarizing return statuses with alerts. Currently, this Agent has entered enterprise production systems and automatically creates over 1,000 return orders per month in Anta brand's TikTok cross-border business. Manual order creation and per-order follow-up have been significantly reduced, and warehouse inbound efficiency has improved accordingly. Next, this workflow is planned to be opened up to FILA and other brands' TikTok business teams.

ArkClaw and Seedance 2.0 are the key capability components behind this series of Agents. ArkClaw is an agent development and runtime framework provided by Volcano Engine, emphasizing task flow orchestration through natural language or visual methods, supporting tool invocation, multi-step execution, and process encapsulation. The case's "encapsulating four templates into a Skill" is a typical use case. Seedance 2.0 is a video generation model used to generate product clips based on prompts, handling the "asset generation" role in content production.

Skill in an agent system typically refers to the smallest reusable capability unit, encapsulating a fixed set of rules or tool invocations for Agents to repeatedly call across different tasks. Anta's overseas ordering scenario turns the conversion rules of four templates into a Skill — essentially making implicit "manual experience" explicit and standardized so the Agent can execute reliably. Token is the basic unit of measurement for large model text processing, used both to measure invocation costs and often to refer to "how much output can be produced under the same compute budget." Understanding these concepts helps clarify whether the "efficiency gains" in these cases come from model capabilities, process orchestration, or the combination of both.

From Single-Brand Efficiency to Cross-Brand Replication: The Scaling Path for Agents

The three Agents correspond to content production, overseas ordering, and cross-border after-sales respectively, but their deployment paths are highly consistent: Start from front-line problems, clearly decompose processes and rules, build an Agent, then validate results back in real business operations.

For Anta Group, the current focus is on stably integrating validated Agents into daily workflows, establishing a unified platform and governance capabilities, and gradually forming cross-brand, cross-regional capability replication.

The value of this path lies in replicability. Once an Agent runs stably in one brand's process, it has the opportunity to be replicated to other brands, other markets, and other business teams. The Agent then transforms from a competition entry into a sustainable business tool.

From a broader perspective, Anta's practice reveals a trend: competition in enterprise AI deployment is shifting from a contest of model capabilities to a contest of workflow integration. Whoever embeds Agents into core business processes faster will reap efficiency dividends sooner. And for multi-brand, cross-market global enterprises, the ability to replicate Agents across regions may be more strategically significant than single-point efficiency gains.

Agent scaling for multi-brand, cross-market enterprises — the difficulty often isn't in developing a single Agent, but in governance and replication. Anta mentions establishing "a unified platform and governance capabilities," which typically includes permission management, data isolation, model invocation standards, performance monitoring, and version iteration mechanisms. When an Agent needs to reuse the same process across different brands and countries, it must maintain consistent core logic while adapting to differences in local language, tax systems, and platform rules. Therefore, decomposing processes into configurable Skills and templates offers greater cost advantages than rewriting from scratch.

Cross-border replication capability as one of the evaluation metrics corresponds to the "marginal cost" logic of enterprise AI: the first Agent in a business typically requires large investment and long cycles, but once it's consolidated into standardized capability, the incremental cost of replicating to a second brand or market drops significantly. For global enterprises, this means Agents not only improve single-point efficiency but also rapidly transfer successful experience from one market to others, forming a compounding organizational capability — which is why its strategic value exceeds single-point efficiency gains.

One-line Summary

Anta embeds AI into content, ordering, and returns workflows through three Agents, revealing enterprise AI's shift from answering to workflow integration.

Paragraph Summary

Through the Volcano Cup Agent Innovation Competition, Anta had business teams raise real pain points and FDEs assist with process decomposition and development, deploying Agents across three major scenarios: content production, overseas ordering, and cross-border returns. The product video Agent enables intern new hires to independently produce qualified videos in half a day; the overseas ordering Agent cuts purchasing decisions from half a day to 10–20 minutes, reduces manual entry by about 90%, and achieves over 99% accuracy; the return document Agent has been auto-creating over 1,000 orders per month in TikTok cross-border business and is planned to be opened to FILA and other brands. Anta evaluates whether Agents have truly entered production across three dimensions — business dependency, default process participation, and cross-border replication capability — noting that enterprise AI competition is shifting from model capability to workflow integration, with cross-regional replication being more strategically significant for global enterprises.