瓴羊退换货 Agent 效率+60%背后:企业级 Agent 供应商正在重做「Token交易市场」规则

Before deploying Lingyang Quick Service, processing a return or exchange took an average of 15 minutes. During peak sales periods, the company needed approximately 600 customer-service agents online at the same time.

Returns Cut From 15 Minutes to 90 Seconds—Who Pays the Token Bill?

A leading beauty e-commerce company recorded a striking set of results during China’s Double 11 shopping festival.

Before deploying Lingyang Quick Service, processing a return or exchange took an average of 15 minutes. During peak sales periods, the company needed approximately 600 customer-service agents online at the same time.

After deployment, the entire process could be completed automatically.

The agent checked the order, verified inventory, created the exchange request, and sent the customer a pickup code—all within 90 seconds.

The customer-service team was reduced from 600 people to 120, with the remaining staff focusing primarily on complex complaints and exceptional cases. The rate of manual intervention fell from 35% to 8%.

According to Lingyang, its return-and-exchange agent can automatically handle more than 60% of inquiries previously processed by human agents, improving overall handling efficiency by more than 60%.

The visible result is a dramatic reduction in labor and response time.

The less visible part of the equation is the enormous amount of token consumption behind every automated decision.

The secret behind reducing a 600-person team to 120 is not simply better customer-service software.

It is a fundamental shift from human labor to token-powered execution.

Efficiency Increased by 60%—But How Much Did Token Consumption Grow?

Lingyang Quick Service is built on Alibaba Cloud’s Qwen large-language-model ecosystem and is designed as a complete agentic loop covering perception, decision-making, and execution.

In a return-and-exchange workflow, the agent does far more than generate a polite response.

It may need to:

  • Retrieve order information from the transaction system

  • Check inventory across warehouses

  • Determine whether the request meets return or exchange policies

  • Call a logistics API to create a pickup order

  • Generate and deliver a collection code

  • Identify the reason for the return

  • Recommend a replacement product

  • Initiate a refund or exchange workflow

  • Verify whether every action has been completed successfully

Each step may involve a separate model inference, tool call, verification process, or retry.

Each step consumes tokens.

Traditional customer-service automation focused mainly on answering frequently asked questions. Agentic customer service does something fundamentally different: it not only understands what the user wants, but also enters enterprise systems and completes the task.

That difference creates the productivity gain.

It also creates a new cost structure.

Over the past two years, AI agents have begun operating continuously across enterprise back-end systems. They handle coding, customer service, document review, risk analysis, data processing, marketing, and internal operations 24 hours a day.

This is one reason they are increasingly described as “token black holes.”

When Lingyang’s return-and-exchange agent takes over more than 60% of inquiries previously handled by humans, it also takes over the reasoning, judgment, system access, and operational steps behind those inquiries.

Human labor costs fall, but token costs appear in their place.

Enterprise AI token consumption is reportedly growing at an annual rate of approximately 580%, far faster than the decline in unit compute costs.

A manufacturing company’s AI quality-inspection system reportedly contributed to a 12% decline in quarterly profit after token expenses exceeded expectations. A large commercial bank in eastern China reportedly spent tens of millions of yuan in a single year on tokens for intelligent customer service, risk auditing, and related applications.

Whether every company experiences the same outcome depends on its models, workflows, and procurement structure. But the economic direction is clear.

Lingyang may have reduced a 15-minute process to 90 seconds and removed the need for 480 concurrent service agents.

The efficiency improvement is real.

The token bill is real as well.

AI Agent Providers Are Rewriting the Rules of Token Distribution

As agents push token demand to a new scale, traditional API supply models are beginning to show their limitations.

A new distribution layer has already emerged in the market.

Some platforms purchase API quotas from model providers in bulk and then resell access to developers and enterprise customers. The model company remains the original token supplier, the distribution platform acts as an intermediary, and the end user pays for the actual model calls.

The business can scale quickly. A relatively small team may generate substantial monthly transaction volume by aggregating models, managing accounts, and distributing API capacity.

However, much of this market still operates through static, one-directional pricing.

The supplier publishes a price. The distributor adds a margin. The developer pays the final amount.

When supply increases, prices may remain unchanged. When demand surges, developers may face rate limits or sudden markups. Idle capacity and peak-hour capacity are often treated as though they have exactly the same value.

That is distribution, but it is not yet an efficient market.

China’s average daily token consumption reportedly increased from approximately 100 billion tokens in early 2024 to more than 140 trillion by March 2026.

Tokens have moved from a resource that providers once struggled to monetize into capacity that can become scarce during periods of heavy demand.

Yet the circulation mechanism remains largely one-way:

Suppliers set the price, and developers accept it.

As agentic workloads continue to grow, that structure will become increasingly difficult to sustain.

The End State of the Token Market: From Supplier Pricing to Market Pricing

The deeper transformation is not simply about selling more API access.

It is about rewriting the rules by which tokens are priced, distributed, and settled.

Tokens are becoming a value anchor for the intelligent economy, much as oil is measured by the barrel and electricity by the kilowatt-hour.

At GTC 2026, Jensen Huang introduced the concept of the “token factory.” In this model, data centers are no longer merely warehouses that store information. They are factories that convert electricity, chips, data, and models into intelligent output.

But producing tokens is only the first half of the market.

The larger question is what happens after production.

How should available token capacity circulate? How should idle supply be discovered? How should buyers compare prices? How should suppliers receive payment? How should prices respond when demand changes?

This is the market layer Unexhub is designed to address.

Unexhub is building a compute marketplace for aggregating and distributing token supply, with the goal of shifting the market from supplier-defined prices toward prices discovered through real supply and demand.

Its model combines several mechanisms.

Real-Time Revenue Settlement

Traditional distribution channels often require smaller suppliers to wait through lengthy billing and payment cycles.

Unexhub links revenue more directly to actual API usage. Once a supplier’s model or inference capacity is called, the corresponding revenue share can be calculated and settled in real time.

For small and medium-sized suppliers, faster settlement can significantly reduce working-capital pressure.

Instead of providing capacity today and waiting weeks or months to recover revenue, suppliers can convert usage into cash flow more quickly.

The principle is straightforward:

The more useful capacity a supplier contributes, the more rapidly that capacity can be translated into income.

Early-Supplier Commission Incentives

To encourage more high-quality model and compute providers to join the marketplace, eligible early suppliers may receive commission incentives, with selected participants paying as little as 0% platform commission during the applicable promotional period.

The value is not limited to saving a service fee.

Lower early-stage commission costs allow suppliers to retain more of their API revenue while building call volume, customer relationships, and market visibility.

This is particularly useful for independent model developers and smaller AI companies that may have strong technical capabilities but limited distribution resources.

Dynamic API Bidding

The most important mechanism is dynamic API bidding.

Instead of assigning one permanently fixed price to every call, available API capacity can compete according to real-time market conditions.

During off-peak periods, developers can gain access to lower-priced capacity. During periods of high demand, suppliers that release additional capacity may earn higher returns.

The pricing logic therefore moves from:

“The supplier publishes a price, and the developer pays it.”

to:

“The market matches the task with the most suitable available supply.”

This transforms tokens from products with fixed price tags into assets that can be routed, traded, and priced in real time.

What This Could Mean for High-Volume Agents

A return-and-exchange agent may process a huge number of decisions every day.

Each customer request can trigger order retrieval, inventory checks, policy analysis, logistics calls, refund decisions, recommendations, and result verification.

At scale, these workflows generate enormous token demand.

Under a traditional procurement model, the enterprise purchases all tokens from one or several suppliers at fixed prices. It pays approximately the same rate whether the request is urgent or flexible, whether infrastructure is congested or idle, and whether another supplier has unused capacity.

A dynamic marketplace creates more room for cost optimization.

Non-urgent tasks could be routed toward lower-priced off-peak capacity. Peak-hour workloads could receive priority access through higher service tiers. If one model channel becomes expensive or unavailable, requests could be redirected toward another qualified supplier.

For high-volume customer-service agents, even a small reduction in the average cost per task can produce significant savings when multiplied across millions of workflow steps.

The advantage is not simply “finding a cheaper model.”

It is matching each task with the right model, at the right price, through the right supplier, at the right time.

From Internal Token Control to External Market Access

When Lingyang’s customer-service case is viewed alongside the emergence of token trading, the larger picture becomes clearer.

Inside the enterprise, a token pool addresses the question of control:

  • How much can each department spend?

  • Which agent receives priority?

  • How should usage be monitored?

  • Which business unit should bear the cost?

  • When should abnormal consumption trigger an alert?

Outside the enterprise, a token marketplace addresses the question of procurement:

  • Where should tokens be purchased?

  • Which supplier offers the best combination of price and quality?

  • When should flexible tasks be moved to lower-cost capacity?

  • How can supply be maintained during peak demand?

The two systems will eventually need to connect.

An enterprise token pool should not remain only an internal quota ledger. It should become an intelligent purchasing interface linked to the external token market.

When internal capacity runs low, the system could purchase additional tokens according to real-time prices, approved model lists, latency requirements, and preset budget rules.

When reserved capacity is idle, the enterprise or supplier could potentially release part of it back into the market.

This creates a complete optimization loop:

Internal budgeting controls demand, while external market access optimizes supply.

Unexhub: Connecting Model Supply With Real-Time Demand

Unexhub is building this connection layer by aggregating AI supply across different scales.

The platform includes resources connected to major model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, while also providing distribution opportunities for models and inference capacity offered by independent developers and small and medium-sized AI companies.

Instead of forcing developers to maintain separate accounts, API formats, billing systems, and monitoring tools for every supplier, Unexhub provides a unified model-access and distribution environment.

Its intelligent routing mechanism evaluates factors such as:

  • Real-time API prices

  • Model availability

  • Response latency

  • Task requirements

  • Supplier capacity

  • Service stability

Requests can then be directed toward a suitable channel within milliseconds.

For developers and enterprise users, this means they can pay according to actual token consumption rather than committing to large quantities of unused capacity.

Depending on the model, task, and available supply, selected API routes may cost as little as 30% of prevailing market prices. Flexible workloads can also benefit from lower-cost off-peak resources without requiring developers to rebuild their applications around multiple providers.

For suppliers, access is designed to be straightforward.

A model provider or compute supplier can connect through a unified channel, reach additional enterprise and developer demand, and earn revenue according to actual API calls.

Real-time revenue settlement helps shorten payment cycles, while early-stage commission incentives allow eligible suppliers to retain a larger share of their earnings.

This creates value on both sides of the marketplace:

Developers gain more choices, lower potential costs, and millisecond-level routing.

Suppliers gain additional distribution, higher capacity utilization, and faster cash recovery.

The Real Question Behind the 90-Second Return

Lingyang reduced the return-and-exchange process from 15 minutes to 90 seconds and allowed 480 customer-service positions to be redirected away from repetitive work.

That is a meaningful operational breakthrough.

But the factor that will ultimately determine whether this model can scale is not only how fast the agent works.

It is also where the required tokens come from, how much they cost, whether supply can be guaranteed, and whether capacity can move freely between providers.

If an enterprise can only purchase tokens from a single supplier at a fixed price, rising token demand may eventually absorb a growing share of the labor savings created by automation.

If its token pool can connect to a dynamic market, the enterprise gains more options.

It can route flexible tasks toward lower-cost periods, secure priority capacity during demand spikes, switch suppliers when channels become congested, and potentially monetize unused quotas or compute resources.

That is the difference between simply deploying an agent and building a sustainable agent economy.

The era of unmanaged token consumption is ending.

The era of token liquidity is beginning.

The move from 15 minutes to 90 seconds proves the value of AI agents.

It also exposes the next bottleneck: the market infrastructure required to supply, route, price, and settle the tokens that keep those agents running.

When tokens can circulate freely, prices can respond to supply and demand, and suppliers can connect directly with developers and enterprises, the next stage of the compute economy will truly begin.


退换货从 15 分钟缩短至 90 秒,Token 账单该由谁来结?

某头部美妆电商在“双11”期间公布的一组数据,揭示了智能体给企业经营带来的巨大变化。

部署瓴羊速服之前,完成一次退换货平均需要约 15 分钟。大促高峰期,企业需要安排约 600 名客服同时在线,才能承接集中涌入的咨询。

部署智能体之后,整个流程可以自动完成。

从查询订单、核对库存,到生成换货工单、推送上门取件码,全程只需要约 90 秒。

企业的在线客服规模也从 600 人缩减至 120 人,留下的人员主要负责复杂客诉和特殊情况。人工干预率从 35% 降低至 8%。

根据瓴羊公布的数据,其“退换货智能体”能够自动处理超过 60% 原本需要人工承接的咨询,整体处理效率提升 60% 以上。

用户能够直接看到的,是响应速度变快、人工成本下降。

但在这些数据背后,还有另一笔不那么显眼的账:

每一次自动判断和系统操作,都在持续消耗 Token。

从 600 人减少到 120 人的秘密,并不只是上线了一套更聪明的客服软件。

它本质上是将大量人力操作,转移给由 Token 驱动的智能体执行。

效率提升 60%,Token 消耗增长了多少?

瓴羊速服基于阿里云通义千问大模型体系打造,定位并不是简单的智能问答,而是形成“感知—决策—执行”的完整智能体闭环。

在退换货场景中,智能体需要完成的工作远不止回复一句话。

它可能需要:

  • 从交易系统中查询订单;

  • 核对不同仓库的库存状态;

  • 判断用户是否符合退换货规则;

  • 调用物流接口创建取件工单;

  • 生成并推送取件码;

  • 识别用户申请退货的原因;

  • 推荐可替换的商品;

  • 发起退款或换货流程;

  • 检查每一个操作是否执行成功。

其中的每一个步骤,都可能对应一次模型推理、工具调用、结果验证或失败重试。

每一步都需要消耗 Token。

传统客服自动化主要解决的是“重复问题自动回复”。

智能体客服则完全不同:它不仅需要理解用户想要什么,还要进入企业的订单、库存、物流和售后系统,把任务真正执行完毕。

这正是效率提升的来源。

但与此同时,它也带来了一套全新的成本结构。

过去两年,AI 智能体开始在企业后台 7×24 小时持续运行,覆盖编程、客服、合同审核、风险分析、数据处理、营销和内部运营等大量场景。

这也是智能体被称为“Token 黑洞”的原因。

当瓴羊退换货智能体接管超过 60% 的人工咨询时,原本由员工完成的推理、判断、系统查询和业务操作,也随之转移给了模型。

人工成本降低了,Token 成本则开始出现。

企业级 AI 应用的 Token 消耗量,据称正以每年约 580% 的速度增长,远远超过单位算力成本的下降速度。

某制造企业的 AI 质检系统曾因 Token 成本超过预算,导致季度利润下降 12%;华东地区某大型商业银行,一年仅在智能客服、风险审计等场景上的 Token 支出就达到数千万元。

不同企业的模型、工作流和采购体系存在差异,并不一定都会面临完全相同的结果。

但方向已经非常明确。

瓴羊将退换货流程从 15 分钟压缩到 90 秒,减少了 480 个需要同时在线的人工岗位。

效率提升是真实的。

Token 账单同样是真实的。

智能体正在重写 Token 的分发规则

当智能体将 Token 需求推向新的数量级,传统 API 供应方式的局限性也开始显现。

市场中已经出现了一层新的 Token 分发体系。

部分平台会批量采购模型厂商的 API 配额,再将模型调用能力分发给开发者和企业用户。模型厂商是 Token 的源头供应商,分发平台承担渠道和账户管理,最终用户则为实际调用付费。

这种模式可以快速扩大交易规模。即使团队规模不大,只要能够聚合模型、管理账号并向下游分发 API,也可能实现较高的月度流水。

但目前大量 Token 分发业务,仍然采用静态、单向的定价方式。

供应商先制定价格,分发渠道增加一定服务费用,开发者按照最终报价采购。

市场供给增加时,价格未必下降;需求突然上涨时,开发者可能面临限流、排队或临时加价。闲置时段的算力与高峰时段的算力,往往被视为拥有完全相同的价值。

这可以称为分销,却还算不上一个高效市场。

国内 Token 日均调用量据称已经从 2024 年初的约千亿,增长至 2026 年 3 月的 140 万亿以上。

Token 已经从过去“打折也不一定能卖出去”的资源,变成高峰时期可能“加价也难以获得”的紧张产能。

然而,现有流通体系仍然主要依赖一条单向通道:

供应商定价,开发者买单。

随着智能体任务继续增长,这种结构将越来越难以支撑长期扩张。

Token 市场的终局:从供应商定价走向市场定价

真正的变化,并不只是销售更多 API 配额。

更深层的变革,是重新定义 Token 的定价、分发和结算规则。

Token 正在成为智能经济的重要价值锚点,就像原油按桶计价、电力按千瓦时结算一样。

黄仁勋在 2026 年 GTC 大会上提出了“Token 工厂”的概念。在这一体系中,数据中心不再只是保存文件的仓库,而是将电力、芯片、数据和模型转化为智能输出的生产工厂。

但生产 Token 只是市场的一半。

更重要的问题是,生产完成之后该怎么办。

闲置的 Token 产能如何流通?买家如何比较价格?供应商如何快速回款?当市场需求发生变化时,价格又该如何及时响应?

这正是 Unexhub 希望解决的市场环节。

Unexhub 正在搭建一个聚合与分发 Token 供给的算力交易中心,推动 Token 从“由供应商单方面定价”,逐渐转向“由真实供需共同发现价格”。

其核心机制主要体现在以下几个方面。

实时结算调用收益

传统分销渠道通常存在较长的对账和回款周期。

中小供应商今天提供算力或模型服务,可能需要等待数周甚至更长时间,才能收到相应收入。

Unexhub 将收益与实际 API 调用直接关联。当供应商提供的模型或推理资源产生调用后,对应的分成可以根据实际用量实时计算与结算。

对于中小型供应商而言,更快的结算速度能够有效缓解资金周转压力。

供应商不必长期垫付服务器、带宽和推理成本,再等待漫长账期回款,而是可以更快地将实际使用量转化为现金流。

其逻辑非常直接:

提供多少有效算力,就有机会更快回收相应收入。

早期供应商佣金激励

为了吸引更多优质模型和算力资源加入,符合条件的早期供应商可根据平台活动规则获得佣金优惠,部分参与者最低可享受 0 佣金政策。

这项权益的价值并不只是减少一笔平台费用。

更低的早期渠道成本,能够帮助供应商在积累调用量、用户资源和市场曝光的同时,保留更多 API 收入。

对于具备模型或推理能力,却缺少销售渠道的独立开发者和中小型 AI 企业而言,这种机制尤其具有现实价值。

API 动态竞价

Unexhub 更重要的机制,是 API 动态竞价。

平台不要求所有调用长期绑定同一个固定价格,而是允许可用的 API 资源根据实时供需参与竞争。

在需求较低的闲置时段,开发者可以匹配价格更低的可用算力;在需求高峰期,愿意释放更多产能的供应商,则有机会获得更高收益。

Token 的价格逻辑由过去的:

“供应商公布价格,开发者按照价格购买。”

转变为:

“市场根据任务需求和实时供给,撮合更加合适的资源。”

Token 不再是一种贴着固定价签的商品,而是在逐渐成为可以路由、交易和实时定价的数字资产。

这对高频智能体意味着什么?

一套退换货智能体,每天可能需要处理数量庞大的用户请求。

每一次请求都可能触发订单查询、库存核对、规则判断、物流调用、退款决策、商品推荐和结果验证。

当调用规模不断扩大,背后所对应的 Token 需求将十分可观。

在传统采购模式下,企业通常按照固定价格,从一家或少数几家供应商采购全部 Token。

任务是否紧急、算力是否处于闲置状态、其他供应商是否存在空余产能,往往都不会直接影响企业所支付的单价。

动态交易市场则提供了更多成本优化空间。

不紧急的批量任务,可以优先路由至价格更低的闲时算力;大促高峰期的核心售后请求,可以通过更高等级的服务获得优先资源。

如果某个模型渠道出现拥堵、涨价或暂时不可用,请求还可以自动切换到其他符合要求的供应商。

对于高频客服智能体而言,即使单次任务只降低一小部分平均成本,当这种节省被放大到数百万次工作流步骤时,也会形成非常明显的经营价值。

真正的优势,不只是“找到一个更便宜的模型”。

而是在正确的时间,通过正确的供应商,以更合适的价格,为每一个任务选择更匹配的模型。

从内部 Token 管控走向外部市场接入

将瓴羊的客服案例与 Token 交易市场放在一起,未来的企业 AI 成本体系便会变得更加清晰。

在企业内部,Token 池解决的是“怎么管”的问题:

  • 每个部门可以使用多少预算?

  • 哪些智能体拥有更高优先级?

  • Token 用量应该如何监控?

  • 成本应当计入哪个业务部门?

  • 什么情况下需要触发异常告警?

在企业外部,Token 交易市场解决的是“怎么买”的问题:

  • 应该从哪个供应商采购?

  • 哪条线路能够兼顾价格与服务质量?

  • 哪些任务可以转移到更低成本的算力上运行?

  • 大促高峰期如何保障稳定供给?

最终,这两套体系需要连接起来。

企业 Token 池不应只是一本封闭的内部配额账,而应成为一个连接外部 Token 市场的智能采购接口。

当内部额度不足时,系统可以根据实时市场价格、可信模型名单、延迟要求和预设预算,自动补充 Token。

当预留算力存在闲置时,企业或供应商也可以尝试将部分产能重新释放到市场中。

由此形成一套完整的成本优化闭环:

内部预算体系负责控制需求,外部交易市场负责优化供给。

Unexhub:连接分散供给与实时需求

Unexhub 正通过聚合不同规模的 AI 供应资源,建设这一连接层。

平台不仅覆盖 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型生态,也为独立开发者、专业团队和中小型 AI 企业提供模型及推理算力的分发机会。

开发者无需再为每一家供应商分别维护账户、API 格式、密钥、账单和监控工具,而是可以通过统一的模型接入与分发环境调用不同资源。

Unexhub 的智能路由机制,可以综合评估:

  • API 实时价格;

  • 模型可用状态;

  • 响应延迟;

  • 任务需求;

  • 供应商剩余产能;

  • 服务稳定性。

随后,平台可以在毫秒级时间内,将请求路由至更加匹配的调用渠道。

对于开发者和企业用户而言,这意味着可以根据实际 Token 消耗付费,而不必提前采购大量可能无法充分利用的固定算力。

根据具体模型、任务类型和实时供应情况,部分 API 线路最低可至常规市场价格的 3 折。弹性任务还可以优先匹配价格更低的闲时资源,同时无需开发者为不同供应商重复改造应用。

对于供应商而言,接入流程也更加直接。

模型或算力供应商可以通过统一渠道接入资源,触达更多企业和开发者需求,并按照实际 API 调用量获得收益。

实时结算机制有助于缩短账期,早期佣金优惠则能够让符合条件的供应商保留更大比例的调用收入。

最终,交易双方都能获得更加具体的价值:

开发者拥有更多模型选择、更低的潜在成本和毫秒级智能路由;

供应商获得新的分发渠道、更高的算力利用率以及更快的资金回收速度。

90 秒退换货背后的真正问题

瓴羊将退换货流程从 15 分钟缩短至 90 秒,并让 480 名客服人员从大量重复工作中得到释放。

这是一项具有现实价值的效率突破。

但最终决定这套模式能否持续扩大规模的,并不只是智能体运行得有多快。

还要看支撑智能体运行的 Token 从哪里来、以什么价格获得、能否保障供给,以及是否能够在不同供应商之间自由流动。

如果企业只能按照固定价格,从单一供应商采购 Token,那么不断增长的调用量,迟早会吞噬越来越多人力成本节省所带来的红利。

如果企业 Token 池能够接入动态交易市场,选择空间就会完全不同。

弹性任务可以转移到更低价的时段运行;大促高峰可以提前锁定优先算力;某一渠道出现拥堵时可以灵活切换;闲置配额和推理资源也有机会被重新利用。

这正是“上线一个智能体”与“建立可持续智能体经济”之间的区别。

粗放式 Token 管理的时代正在结束。

Token 流动的时代正在到来。

从 15 分钟缩短至 90 秒,验证了智能体能够为企业创造多大价值。

它同时也暴露出了下一个瓶颈:如何为这些智能体持续提供可计量、可路由、可定价、可结算的 Token 资源。

当 Token 可以自由流通,价格能够随供需变化,供应商可以与开发者和企业直接撮合时,算力经济的下一阶段才会真正开启。