
95% Self-Service Completion and 20% Higher Conversion: Measuring the Real ROI of AI Agents in Automotive Finance Dealerships
English Version First|中文完整版见后
How long does it take a customer to complete an auto-financing application at a dealership?
Under a traditional process, the answer may be more than 90 minutes.
A sales consultant often needs to stay with the customer throughout the entire journey: explaining financing products, completing forms, copying identification documents, waiting for preliminary approval, setting up repayment accounts, and signing contracts.
During peak hours, customers wait in line, sales consultants become overwhelmed, and a single adviser may be able to serve only around eight groups per day.
This is not a problem unique to one dealership.
For more than a decade, the automotive-finance industry has struggled with four persistent obstacles:
Lengthy service processes
Complex credit decisions
Customer concerns about personal-data privacy
Strict security and compliance requirements
On July 17, 2026, at WAIC 2026 in Shanghai, Yixin Group introduced what it described as China’s first integrated AI terminal for automotive finance.
It is not simply a chatbot placed inside a kiosk.
The terminal combines an agentic large-model system, multi-agent collaboration, digital-human guidance, risk controls, and end-to-end process automation in one physical service point.
Its purpose is not merely to answer financing questions.
It is designed to complete the transaction.
One Terminal, Multiple Agents Working in Parallel
After a customer selects a vehicle and submits the required information through the terminal, the agentic system begins coordinating multiple specialized agents in the background.
Different agents can work on different parts of the order simultaneously:
One processes vehicle-registration or document recognition
One evaluates the vehicle and transaction information
One conducts credit and risk analysis
One handles asset management, quality inspection, and process verification
In a traditional workflow, many of these steps are completed sequentially.
One department waits for another. A sales consultant waits for the customer. The customer waits for a risk decision. Any missing field can interrupt the entire process.
A multi-agent system changes the structure.
Tasks that do not depend on one another can run in parallel, while the orchestration layer monitors progress, shares necessary information, and determines which step should follow next.
According to the initial operating data cited in the original material, the intelligent risk-control system can produce a preliminary decision in approximately eight seconds.
The complete customer journey can be shortened from more than 90 minutes to roughly 15–20 minutes.
Depending on the original process and the specific transaction, this represents an improvement of approximately 4.5 to 6 times.
The reduction is not achieved by making one form load faster.
It comes from redesigning the entire workflow around parallel agents.
Two Numbers That Define the Business Case: 95% and 20%
As of June 30, 2026, the terminal’s early operating data reportedly produced two particularly important results:
A self-service completion rate above 95%
A business conversion-rate improvement above 20%
These figures matter because they reflect two different forms of ROI.
The 95% figure measures operational independence.
The 20% figure measures commercial performance.
Together, they show how an AI agent can influence both cost efficiency and revenue generation.
What Does a 95% Self-Service Completion Rate Mean?
A self-service completion rate above 95% means that, among every 20 customers, approximately 19 may be able to complete the standard process without requiring a sales consultant to accompany them through every step.
With digital-human guidance, customers can independently complete tasks such as:
Preliminary financing assessment
Identity and information collection
Document submission
Repayment-card binding
Contract review
Electronic signing
Progress tracking
This changes the role of the dealership adviser.
The consultant no longer needs to act as a full-time process operator for every transaction.
Instead, the adviser can focus on:
Explaining complex financing structures
Handling exceptional applications
Resolving incomplete or conflicting information
Supporting customers who need additional assistance
Managing higher-value customer relationships
In other words, the consultant moves from being an “operator” to being an “exception manager” and financial-service adviser.
Only the smaller proportion of complex, incomplete, or high-risk cases needs deeper human intervention.
This is where the operating leverage begins.
The system does not eliminate the employee.
It allows the employee to serve more customers without repeating the same administrative steps for every application.
Privacy Becomes Part of the Customer Experience
Self-service is not valuable only because it saves labor.
It also addresses one of the most sensitive issues in automotive finance: personal-document privacy.
In a traditional dealership process, a customer may need to hand identity documents, bank-card details, or other sensitive materials to a sales consultant for copying and submission.
Even when the process is compliant, many customers remain uncomfortable with third parties handling their personal information.
The integrated AI terminal uses a self-service upload model.
Customers can scan, upload, and confirm sensitive documents directly through the device, reducing unnecessary human handling.
For the customer, the benefit is straightforward:
Documents remain under personal control
The process is easier to understand
Application progress is more transparent
Fewer people directly access sensitive information
The terminal also supports QR-code synchronization with a mobile device, allowing customers to continue or monitor the process from their phones.
A customer who arrives outside normal peak hours can begin the application immediately rather than waiting for a specific adviser to become available.
The value is not simply “faster service.”
It is greater control over the transaction.
What Does a 20% Conversion Improvement Mean?
A conversion-rate increase above 20% should not be interpreted as a guaranteed result for every dealership.
Actual performance will depend on customer traffic, vehicle brands, financing products, regional conditions, staff adoption, and the original conversion baseline.
But the commercial logic is clear.
When other conditions remain broadly similar, a shorter and more convenient process reduces the number of customers who abandon the financing application because of:
Long waiting times
Complicated forms
Repeated document submission
Unclear progress
Privacy concerns
Delayed preliminary decisions
Every customer who leaves midway represents a lost financing opportunity and, in some cases, a lost vehicle sale.
If the terminal allows more customers to complete the application during the same visit, the dealership can convert more existing foot traffic without opening another store or adding the same proportion of staff.
The improvement is therefore not merely a reduction in service time.
It can directly affect revenue.
From Eight Customer Groups to Fifteen per Adviser
According to the operating figures provided in the original material, one adviser’s daily service capacity increased from approximately eight customer groups to fifteen, while customer waiting time fell by around 70%.
This is a useful way to understand the operational ROI.
Suppose a dealership previously relied on five advisers, each able to serve eight groups per day.
Theoretical daily capacity would be:
5 advisers × 8 groups = 40 customer groups
If the effective capacity rises to fifteen groups per adviser, the same team could support:
5 advisers × 15 groups = 75 customer groups
The exact number of completed deals would still depend on customer demand and approval rates. However, the dealership would have substantially more service capacity without proportionally increasing headcount.
The system creates value in three ways:
More customers can be served
Existing employees spend less time on repetitive operations
Customers are less likely to abandon the process while waiting
That is why the ROI of a financial agent cannot be measured only through labor savings.
It also needs to include:
Higher service capacity
Shorter customer waiting time
Reduced abandonment
Better privacy experience
More consistent compliance execution
Potentially higher financing conversion
125 Million Calls: What Scale Really Looks Like
By the end of May 2026, Yixin’s AI platform had reportedly processed more than 125 million calls.
Approximately 65% of business processes were said to be delivered autonomously by agents, while overall operating efficiency improved by more than 100%.
These figures reveal an important difference between a demonstration and a production system.
A demonstration proves that an agent can complete one task.
A production system must complete the task:
Repeatedly
At high concurrency
Across different customer types
Under compliance constraints
With stable latency
With human escalation when needed
While maintaining predictable cost
At this scale, the agent is no longer an experimental interface.
It becomes part of the company’s operating infrastructure.
The Harness Behind the Agents: Governance Is Not Optional
High efficiency is meaningless if the system cannot meet financial security and compliance requirements.
Financial agents may produce incorrect information, misunderstand a customer’s situation, or generate commitments that the institution is not authorized to make.
A production-grade system therefore needs more than a capable model.
It needs a governance layer.
Yixin’s self-developed Harness governance system is designed to allow agents and human employees to move within the same order workflow.
When the system detects abnormal model behavior, a possible hallucination, an unauthorized statement, or a high-risk situation, it can trigger control measures and transfer the task to a human process.
This creates several layers of protection:
Real-time monitoring
Rule-based restrictions
Model-output review
Rapid circuit breaking
Human escalation
Order-level traceability
Audit records
The purpose is not to assume that the model will never make a mistake.
It is to make sure a mistake does not automatically become a financial commitment.
As Yixin Chief Technology Officer Jia Zhifeng emphasized at WAIC, security and compliance in financial large-model applications are not optional enhancements.
They are the foundation of the business.
Human and Agent Must Be Able to Share the Same Order
One of the most important design requirements in financial services is smooth human-agent handoff.
A customer should not be forced to restart the entire application simply because the system encounters an exceptional case.
For example, human intervention may be required when:
Document recognition confidence is too low
Financial information is incomplete
The customer disputes a risk result
The transaction structure is unusual
Compliance rules require additional verification
The customer does not understand a contract clause
A model-generated statement requires review
A mature agent system needs to preserve all prior context and transfer the order to an employee without breaking the process.
This is why the ability to switch between an agent and a human within the same order is as important as the model’s reasoning capability.
Automation handles the standard path.
Humans handle exceptions, responsibility, and final judgment.
A Practical ROI Framework for Dealerships
The ROI of an automotive-finance terminal can be evaluated through four categories.
1. Time Savings
The customer journey can fall from more than 90 minutes to approximately 15–20 minutes.
That means:
Shorter queues
Faster vehicle delivery
Higher customer satisfaction
More transactions completed during peak periods
2. Labor Leverage
A larger proportion of standard applications can be completed without full-time adviser accompaniment.
Consultants can focus on exceptions, sales support, and customer relationships.
The dealership does not necessarily need to add employees at the same rate as transaction volume grows.
3. Conversion Improvement
A shorter, clearer, and more private process can reduce abandonment and help more customers complete financing applications.
Under comparable operating conditions, this can translate into more completed transactions from the same customer traffic.
4. Risk and Compliance Consistency
A standardized agent workflow can apply the same document checks, process rules, and escalation conditions across more transactions.
This reduces reliance on individual employees remembering every step correctly.
The value is not only greater speed.
It is more consistent execution.
The Next Expansion: Dealerships, Used-Car Markets, and County-Level Outlets
Yixin plans to gradually extend the integrated AI terminal into more scenarios, including:
New-car dealerships
Used-car markets
County-level service outlets
Regional automotive-finance channels
Other offline customer-service points
This expansion changes the economic problem.
A pilot project may run on a limited number of terminals with controlled traffic.
A nationwide deployment may need to support:
Thousands of devices
High daily transaction volume
Continuous model calls
Real-time risk decisions
Digital-human conversations
Document analysis
Contract processing
Compliance monitoring
At that point, model capability is only one part of the equation.
The cost of operating the agents becomes a major infrastructure decision.
Compute Cost: The Final Barrier to Scaling
The reported results—95% self-service completion, more than 20% conversion improvement, and eight-second preliminary decisions—are supported by a platform that has already processed more than 125 million calls.
Every call consumes compute resources.
A single order may trigger multiple agents to perform:
Document recognition
Customer-intent analysis
Risk reasoning
Vehicle evaluation
Contract interpretation
Quality inspection
Compliance checks
Digital-human interaction
When the system moves from a small pilot to thousands of terminals, Token usage expands rapidly.
The key question changes from:
Can the agent complete the transaction?
to:
Can the enterprise afford to operate the agent across its entire network?
This is where Unexhub enters the infrastructure layer.
Unexhub: Making Token Capacity More Market-Oriented
Unexhub is building a Token aggregation and distribution network designed to connect model providers, compute suppliers, enterprise customers, and developers.
Its objective is to transform Token capacity from a resource sold separately by individual providers into capacity that can be aggregated, routed, priced, distributed, and settled through a compute trading center.
The model can be understood through three mechanisms.
1. Token Aggregation and Distribution
AI capacity is distributed across:
Major model providers
Regional computing centers
Specialized AI companies
Small and medium-sized suppliers
Independent development teams
Individual model providers
Many smaller suppliers have high-quality models or underused inference capacity but lack:
Stable customer traffic
Global distribution
Intelligent routing
Unified billing
Market visibility
Pricing power
Unexhub brings these resources into a unified supply pool and distributes them through standardized APIs.
For enterprises, this creates more model and endpoint choices.
For suppliers, it creates an additional commercialization channel.
2. Faster Settlement and Lower Supplier Pressure
Traditional enterprise payment cycles can place substantial pressure on smaller compute suppliers.
Unexhub aims to calculate model usage and supplier revenue based more closely on actual API consumption, shortening the period between providing capacity and receiving payment.
Suppliers can obtain clearer visibility into:
Token usage
Revenue by model
Customer demand
Endpoint performance
Settlement progress
Qualified early suppliers may also receive promotional commission rates as low as 0%.
This benefit is not only a short-term saving.
It gives suppliers a lower-cost period in which to test demand, build initial traffic, validate service quality, and establish a distribution channel before moving into a standard commission structure.
3. API Bidding and Dynamic Pricing
Traditional API prices are usually determined by individual model providers.
Unexhub proposes a market-oriented mechanism in which effective API costs can respond to:
Available capacity
Real-time demand
Model performance
Response latency
Endpoint reliability
Peak and off-peak periods
Competition among suppliers
When capacity is abundant, enterprises may gain access to more economical routes.
When demand rises, suppliers capable of delivering stable service may receive stronger returns.
Token pricing therefore becomes a result of market matching rather than only a fixed number on a provider’s official website.
From One Model for Everything to Task-Based Routing
An automotive-finance order does not need the most expensive model for every step.
Different tasks may require different capabilities:
Document extraction may use a fast, lower-cost model
Customer guidance may prioritize latency and natural language
Complex risk analysis may require a stronger reasoning model
Compliance checks may use a specialized or local model
Final contract review may require a more accurate model
Sensitive data may need to remain in a private environment
Unexhub’s multi-model aggregation and millisecond-level routing can allow enterprises to assign different tasks to different model routes instead of forcing the entire process through one premium endpoint.
This creates two layers of savings:
Use a model that matches the task
Source that model through a more competitive route
Depending on model availability, supplier competition, route conditions, and promotional policies, selected services may be available at prices as low as approximately 30% of comparable market reference rates.
For one financing application, the difference may seem small.
Across millions of calls, it can become a major operating-cost advantage.
Pay Only for Actual Usage
For enterprises expanding from pilot projects to national deployment, fixed packages can create two problems.
If the package is too small, the system reaches its limit during peak demand.
If it is too large, the enterprise pays for capacity that may remain unused.
Unexhub’s pay-as-you-use Token model gives enterprises greater flexibility.
They can:
Pay according to actual consumption
Route different tasks to different models
Adjust capacity as transaction volume changes
Avoid maintaining multiple provider accounts
View usage through a centralized billing structure
Reduce the burden of integrating separate APIs
The value is not simply access to a lower advertised price.
It is a cost structure that can expand with the business.
When Automotive Finance Agents Meet a Compute Trading Center
Yixin’s integrated AI terminal and Unexhub operate at different layers of the same transformation.
The automotive-finance terminal rebuilds the application workflow.
It turns document collection, preliminary risk assessment, repayment setup, contract signing, and quality control into a coordinated multi-agent process.
Unexhub attempts to rebuild the infrastructure cost structure.
It aggregates model supply, routes tasks between endpoints, introduces market competition, and accelerates supplier settlement.
One answers:
How can automotive financing be completed faster and with less manual involvement?
The other answers:
How can the enterprise operate these agents at large scale without allowing model costs to grow out of control?
When thousands of terminals are processing transactions simultaneously, model routing and procurement become just as important as the customer interface.
The industry has already demonstrated that financial agents can work.
The next stage is proving that they can work economically across an entire network.
From “Can It Be Built?” to “Can It Be Scaled?”
Yixin has also indicated plans to open-source its AI infrastructure within the year.
If this capability expands from one company to a broader ecosystem of dealerships, used-car markets, county-level outlets, and automotive-finance providers, compute cost will become an industry-wide issue.
The question will no longer be:
Can we build an automotive-finance agent?
It will become:
Can we operate thousands of agents continuously, securely, and profitably?
Dojo-style agent frameworks solve workflow automation.
Platforms such as Unexhub attempt to solve model access, routing, and cost sustainability.
The business case is therefore not built on one impressive number.
It is built on a complete operating loop:
Customers complete more steps independently. Advisers serve more people. Transactions finish faster. Risk rules remain enforceable. Model costs remain manageable as volume grows.
That is the real ROI of an automotive-finance agent.
Not simply replacing a person with a machine, but rebuilding the service process so that the same dealership, the same staff, and the same customer traffic can create more completed business.
业务转化率提升 20%、自助办理达成率 95%:汽车金融门店智能体的真实 ROI 测算
中文完整版
在一家 4S 店办理一笔汽车分期贷款,客户通常要花多长时间?
按照传统流程,答案可能是 90 分钟以上。
从产品介绍、填写表格、复印证件,到等待预审、绑定还款卡、确认合同和完成签约,销售顾问往往需要全程陪同。
到了门店高峰期,客户排队等待,销售顾问忙于重复操作,一名顾问一天可能只能接待 8 组左右的客户。
这并不是某一家门店特有的问题。
过去十多年,汽车金融行业始终受到四类问题制约:
业务流程冗长
授信决策复杂
客户对隐私信息存在顾虑
安全与合规要求严格
2026 年 7 月 17 日,在 WAIC 2026 上海会场,易鑫集团发布了其所称的国内首台汽车金融 AI 智能一体机。
它并不是把一个聊天机器人装进一台机器里。
这套实体终端将 Agentic 大模型、多智能体协同、AI 数字人、安全治理与端到端自动化流程整合在一起。
它的目标也不只是回答客户提出的金融问题。
而是把业务真正办完。
一台机器,多个智能体同时开工
客户在一体机上选定车辆、提交所需资料后,后台的 Agentic 系统会开始调度多个专业智能体。
不同智能体可以同时承担不同任务:
一个负责证件及车辆资料识别
一个负责车辆与交易信息评估
一个负责征信和风险分析
一个负责资产管理、质检及流程核验
传统流程中,许多步骤按照先后顺序串行推进。
一个部门等待另一个部门,销售顾问等待客户补资料,客户又要等待审批结果。任何一个字段缺失,都可能让整个流程停下来。
多智能体协作改变了这种结构。
没有强依赖关系的任务可以并行运行,调度层负责共享必要信息、监控任务进度,并自动判断下一步应该执行什么。
按照原文引用的初期运营数据,智能风控系统大约可以在 8 秒内生成预审结果。
完整业务流程则有机会从 90 分钟以上,缩短至约 15—20 分钟。
根据原有流程和具体业务情况计算,整体效率约提升 4.5—6 倍。
这并不是把某一张表格加载得更快。
而是将整条业务流程改造成由多个智能体并行推进的新结构。
两个决定商业价值的数字:95% 与 20%
截至 2026 年 6 月 30 日,这台设备基于初期运营数据交出了两项值得关注的结果:
自助办理达成率超过 95%
业务转化率提升超过 20%
这两个数字分别代表两种不同的 ROI。
95% 衡量的是流程的自助完成能力。
20% 衡量的是商业转化效果。
前者影响运营成本和人员效率,后者则直接关系到成交与收入。
95% 的自助办理率意味着什么?
自助办理达成率超过 95%,意味着每 20 位客户中,大约有 19 位能够在不需要销售顾问全程陪同的情况下,独立完成标准业务流程。
在 AI 数字人的拟人化引导下,客户可以自行完成:
分期预审
身份和信息采集
证件材料上传
还款银行卡绑定
合同查看
电子签约
办理进度查询
这会直接改变门店顾问的角色。
销售顾问不再需要为每一位客户充当全流程“操作员”,而可以把更多精力投入:
解释复杂金融方案
处理异常申请
解决材料缺失或信息冲突
协助不熟悉设备的客户
经营更高价值的客户关系
换句话说,顾问开始从“流程操作员”转变为“异常处理员”和金融服务顾问。
只有那部分复杂、资料不完整或风险较高的案例,需要更深入的人工介入。
这正是智能体带来运营杠杆的地方。
系统不是简单取消员工。
而是让员工不再为每一笔申请重复执行相同的机械步骤。
隐私体验,也会影响业务转化
自助办理的价值并不只是节省人工。
它还解决了汽车金融中最敏感的问题之一:客户证件与个人信息的隐私。
在传统办理过程中,客户可能需要把身份证件、银行卡信息等敏感材料交给销售顾问复印、上传。
即便操作符合规定,许多客户依然会对第三方人员直接接触个人信息感到不安。
AI 一体机采用终端自助上传模式。
客户可以自行扫描、上传和确认敏感材料,减少不必要的人工经手。
对客户而言,这意味着:
证件始终掌握在自己手中
每一步操作更加清晰
办理进度更加透明
接触敏感信息的人员更少
终端还支持扫码同步至手机端,客户可以通过移动设备继续办理或查看全链路进度。
即使销售顾问正在接待其他客户,新的申请人也可以先开始操作,不必等待某一名员工空闲。
它带来的不只是“速度更快”。
而是客户对整个交易拥有更强的控制感。
20% 的转化率提升意味着什么?
业务转化率提升超过 20%,不应被理解为适用于所有门店的固定结果。
实际效果仍会受到客流量、车型品牌、金融产品、区域市场、门店原有转化率和员工使用程度等因素影响。
但它背后的商业逻辑十分清楚。
办理流程变短、操作更简单、隐私感更强,可以减少客户因为以下原因中途放弃:
等待时间过长
表格填写过于复杂
证件反复提交
办理进度不清晰
对个人信息安全存在担忧
预审结果迟迟无法获得
每一位在流程中途离开的客户,都可能意味着一笔分期业务流失,甚至可能影响最终车辆成交。
如果一体机能够让更多客户在同一次到店过程中完成申请,门店就可以在不新增门店、也不同比例扩张员工数量的情况下,把现有客流转化为更多业务。
因此,20% 的提升不只是“节省了一点时间”。
它有机会直接作用于营收。
单个顾问从每天 8 组提升至 15 组
根据原文提供的运营口径,单个顾问的日服务量由约 8 组提升至 15 组,客户等待时间则减少约 70%。
这个数字可以更直观地解释门店的运营 ROI。
假设一家门店原本有 5 名顾问,每人每天服务 8 组客户,理论日接待能力为:
5 人 × 8 组 = 40 组客户
如果单人有效服务能力提升至 15 组,同一团队理论上可支持:
5 人 × 15 组 = 75 组客户
最终成交数量仍然取决于实际客流、审批通过率与客户意愿,但门店能够在不同比例增加员工的情况下,获得更大的服务容量。
系统通过三种方式创造价值:
同一时间可以服务更多客户
顾问用于重复操作的时间明显减少
客户因等待过久而放弃的概率下降
因此,衡量金融智能体的 ROI,不能只看节省了多少人工。
还要同时计算:
服务容量是否增加
客户等待时间是否下降
中途流失是否减少
隐私体验是否改善
合规执行是否更加稳定
金融业务转化是否提高
1.25 亿次调用,才是真正的规模化考验
截至 2026 年 5 月底,易鑫 AI 平台累计调用量据称已经超过 1.25 亿次。
其中约 65% 的业务由智能体自主交付,整体运营效率提升超过 100%。
这些数据揭示了“演示系统”与“生产系统”之间的区别。
演示只需要证明智能体能完成一次任务。
生产系统则需要保证它能够:
不断重复完成任务
支撑高并发请求
处理不同类型的客户
在合规边界内运行
保持稳定响应速度
在必要时及时转人工
同时控制长期成本
达到这一规模后,智能体就不再只是一个实验性入口。
它已经成为企业运营基础设施的一部分。
智能体背后的“缰绳”:安全治理不能缺席
如果金融智能体无法通过安全与合规要求,再高的效率也没有实际意义。
模型可能理解错客户信息、生成不准确的内容,甚至给出机构无权作出的承诺。
因此,真正能够上线的系统不能只依赖模型能力。
它还需要一套治理层。
易鑫自研的 Harness 治理体系,支持智能体与真人在同一订单流程中切换。
当系统识别到模型异常、潜在幻觉、违规承诺或高风险情况时,可以触发治理策略,并将任务转入人工处理链路。
这套机制可能包含:
实时监控
规则约束
输出审核
快速熔断
人工升级
订单级追踪
审计记录
设计目标并不是假设模型永远不会出错。
而是确保模型的错误不会自动变成一项金融承诺。
正如易鑫首席科技官贾志峰在 WAIC 上所强调的,在金融领域,大模型的安全与合规不是锦上添花,而是业务底线。
真人与智能体,必须能够共享同一笔订单
金融服务中一个非常重要的设计要求,是智能体与人工之间可以无缝交接。
系统遇到复杂情况后,不能让客户从头重新填写整份申请。
例如,以下场景可能需要人工介入:
证件识别置信度过低
财务或身份信息不完整
客户对风险结果存在异议
交易结构较为特殊
合规规则要求追加核验
客户无法理解合同条款
模型生成的内容需要复核
成熟的智能体系统必须保留前序流程中的全部上下文,并在不中断订单的情况下,将任务交给员工继续处理。
因此,同一订单中支持“智能体—真人”实时切换,与模型推理能力同样重要。
标准流程交给自动化。
异常处理、责任承担与最终判断仍然由人完成。
门店可以怎样测算真实 ROI?
汽车金融 AI 一体机的 ROI,可以从四个维度进行评估。
一、时间收益
客户办理时间由 90 分钟以上缩短至约 15—20 分钟。
可能带来的结果包括:
门店排队时间缩短
提车速度加快
客户满意度提高
高峰期可完成更多业务
二、人效收益
大部分标准申请可以由客户自主完成,不再需要顾问全程陪同。
顾问能够将时间投入异常处理、销售支持和客户关系维护。
即使交易量增长,门店也不必按照相同比例增加员工。
三、转化收益
流程更短、操作更清晰、隐私体验更好,可以减少中途放弃,让更多客户完成金融申请。
在其他运营条件近似不变的情况下,这有机会将相同客流转化为更多成交。
四、风险与合规收益
标准化智能体流程可以在更多订单中统一执行相同的材料检查、流程规则与升级条件。
业务不再过度依赖某一名员工是否记得所有操作细节。
它带来的不只是速度。
还有执行一致性。
下一步:4S 店、二手车市场与县域网点
易鑫计划将这类 AI 一体机逐步拓展至更多场景,包括:
新车 4S 店
二手车交易市场
县域服务网点
区域汽车金融渠道
其他线下客户服务终端
场景扩张之后,企业面对的经济问题也会发生变化。
一个小范围试点,只需要少量设备和有限流量。
全国性部署则可能需要支撑:
数千台实体终端
大量每日交易
持续模型调用
实时风险判断
AI 数字人对话
文档识别和分析
合同处理
合规监控
到了这一阶段,模型能力只是其中一部分。
智能体长期运行所产生的算力成本,将成为一项关键基础设施决策。
算力成本:规模化的最后一道门槛
95% 自助办理率、20% 以上转化提升和 8 秒预审背后,是一个已经累计调用超过 1.25 亿次的平台。
每一次调用都会消耗算力。
一笔汽车金融订单可能需要多个智能体分别完成:
证件识别
客户意图理解
风险推理
车辆评估
合同解释
质量检查
合规核验
数字人交互
当系统从少量试点走向数千台终端,Token 用量会迅速扩大。
企业面对的问题开始从:
智能体能不能办完这笔业务?
转向:
企业能不能承担智能体覆盖整个服务网络后的长期成本?
这正是 Unexhub 从基础设施层切入的位置。
Unexhub:让 Token 成为可市场化配置的资源
Unexhub 正在建设一个连接模型供应方、算力供应方、企业客户与开发者的 Token 聚合和分发网络。
它希望把原本由不同模型厂商分别出售的 Token 能力,转化为可以通过算力交易中心统一聚合、路由、定价、分发与结算的资源。
这套模式可以拆解为三项核心机制。
第一层:Token 聚合与分发
AI 模型能力分散在:
大型模型供应商
区域算力中心
垂直 AI 企业
中小型算力供应商
独立开发团队
个人模型供应方
不少中小型供应方拥有优质模型或闲置推理容量,却缺少:
稳定用户流量
全球分发渠道
智能路由系统
统一计费能力
市场曝光
价格议价权
Unexhub 将这些资源聚合进统一供给池,再通过标准化 API 分发给企业与开发者。
企业可以获得更多模型和节点选择,供应方则增加了一条商业化渠道。
第二层:加快结算,降低供应方资金压力
传统企业账期可能给中小型模型和算力供应方带来较大资金压力。
Unexhub 希望按照真实 API 调用量,更及时地计算模型用量和供应方收益,缩短算力输出与资金回收之间的周期。
供应方可以更加清晰地查看:
Token 使用量
不同模型的收入
客户需求变化
节点运行表现
当前结算进度
符合条件的早期供应方,还可享受最低 0% 佣金的阶段性权益。
这项优惠的价值不只是短期节省费用。
它为供应方提供了一段成本更低的市场验证期,可以先测试真实需求、积累调用量、检验服务稳定性,再进入常规佣金结构。
第三层:API 竞价与动态定价
传统 API 价格通常由单一模型供应商制定。
Unexhub 希望引入更加市场化的机制,让实际调用价格根据以下因素动态变化:
当前可用算力
实时市场需求
模型性能
响应延迟
节点稳定性
高峰与低谷时段
不同供应方之间的竞争
当算力供给充足时,企业可以获得更具成本优势的线路。
需求进入高峰后,能够稳定提供优质服务的供应方,也可以获得更加合理的收益。
Token 价格由此不再只是模型厂商官网上的固定数字。
而逐渐成为市场供需匹配后的结果。
不同任务,不必全部使用同一款高价模型
一笔汽车金融订单中的所有任务,并不需要使用同一款最昂贵的模型。
不同环节需要的模型能力并不相同:
文档提取可以使用高速、低成本模型
客户引导更关注响应速度和自然对话
复杂风控分析需要更强的推理模型
合规检查可以使用专业模型或本地模型
最终合同复核需要更高准确率
敏感信息可能需要留在私有环境中处理
Unexhub 的多模型聚合与毫秒级智能路由,可以让企业根据任务类型分配不同模型线路,而不是让整套流程全部经过同一个高价接口。
这会形成两层成本优化:
让模型能力与具体任务相匹配
通过更具竞争力的线路采购这部分模型能力
根据模型供应、节点竞争、线路条件和阶段性权益的不同,部分服务最低可享受约为市场参考价格 3 折的成本方案。
对一笔金融申请而言,这种差异可能并不明显。
但放到数千万甚至上亿次调用中,就可能转化为明显的运营成本优势。
按实际用量付费,让成本跟随业务扩张
企业从试点走向全国性部署后,固定套餐可能带来两个问题。
套餐太小,高峰期容易触及额度。
套餐太大,企业又需要为没有实际使用的容量提前付费。
Unexhub 的按 Token 实际用量付费模式,能够给企业提供更灵活的成本结构。
企业可以:
按真实调用量付费
为不同任务选择不同模型
根据交易规模动态调整算力
减少维护多家模型账户
使用统一账单查看整体用量
降低分别接入多套 API 的工程负担
这项权益的价值并不只是看到更低的宣传价格。
而是让算力成本可以随着业务量同步伸缩。
当汽车金融智能体遇上算力交易中心
易鑫 AI 智能一体机与 Unexhub,处在同一场变革的不同层级。
汽车金融一体机重构的是应用层业务流程。
它把资料采集、风险预审、绑卡还款、合同签约和质量检查,改造成由多个智能体协同完成的流程。
Unexhub 尝试重构的是基础设施成本。
它聚合模型供给,在不同节点之间智能路由,引入市场竞价,并加快供应方结算。
一个回答:
汽车金融业务怎样才能办得更快、减少人工陪同?
另一个回答:
企业怎样才能在大规模部署智能体之后,依然控制住模型成本?
当数千台终端同时处理业务时,底层模型路由与算力采购的重要性,并不低于用户看到的交互界面。
行业已经开始证明金融智能体“能用”。
下一阶段要证明的,是它能够以可承受的成本覆盖整个服务网络。
从“能不能做”,走向“能不能规模化”
易鑫还计划在年内开源其 AI Infra。
如果这套能力从一家公司扩展到更多 4S 店、二手车市场、县域网点与汽车金融机构,算力成本就会成为整个行业必须共同面对的问题。
问题不再是:
能不能开发一套汽车金融智能体?
而是:
能不能让数千个智能体持续、安全、合规并且有利润地运行?
Agentic 系统解决的是流程自动化。
Unexhub 这类算力聚合与交易平台,尝试解决的是模型接入、智能路由和长期成本可持续性。
汽车金融智能体的商业价值,也不能只建立在某一个漂亮数字之上。
它需要形成一套完整的运营闭环:
客户可以自主完成更多步骤,顾问能够服务更多人,业务办理速度更快,风险规则仍然有效,模型成本也能随着规模扩大保持可控。
这才是门店金融智能体真正的 ROI。
它不是简单地用机器替代一个人。
而是重新设计整个服务流程,让同样的门店、同样的团队与同样的客户流量,创造出更多可完成的业务。