
From 2–4 Weeks to 2 Hours: How Financial Agents Are Rebuilding Enterprise Credit Due Diligence
English Version First|中文完整版见后
A traditional enterprise credit due diligence report can take anywhere from two to four weeks to complete.
Before any meaningful analysis begins, a relationship manager may need to log in to five or six separate systems—including corporate credit databases, tax platforms, ERP systems, judicial databases, and internal banking tools. Simply collecting and organizing the source material can take hours.
In some traditional workflows, a single report contains more than 990 fields that must be checked or entered manually.
Now imagine handing that entire process to a financial agent.
Within two hours, the first complete report is ready for review.
This is the efficiency inflection point represented by projects such as DojoAgents, which surpassed 1,000 GitHub stars just over a week after becoming open source in July 2026.
Its significance is not that it can write a more fluent report.
The real change is that it can begin to automate the research process behind the report:
Collecting data, calling tools, cross-checking evidence, identifying risks, and assembling a structured conclusion.
The shift from “two to four weeks” to “two hours” is not simply the result of a smarter model.
It happens because the agent has learned how to carry out the work itself.
Why Does Enterprise Due Diligence Take So Long?
Traditional enterprise due diligence has three persistent problems.
1. Information Is Fragmented Across Multiple Systems
The information required for a credit investigation rarely exists in one place.
A company’s operating data may be stored in its business system. Financial statements may arrive as Excel files from the customer. Litigation records may need to be searched through judicial databases. Tax information, shareholder structures, public opinion, industry data, and internal transaction history may all come from different platforms.
A relationship manager often needs to switch repeatedly among six to ten systems to:
Search for the company
Download records
Copy data
Reconcile company names
Standardize formats
Compare conflicting figures
Paste information into a report template
The problem is not a lack of data.
It is that the data is disconnected.
Before the analyst can begin making a judgment, a large portion of the working day has already been consumed by moving information from one system to another.
2. Manual Data Entry Consumes Most of the Workflow
A traditional corporate due diligence report may contain hundreds of required fields.
According to the workflow described in the original material, more than 990 fields may need to be reviewed or entered manually.
A typical allocation of time might look like this:
Three hours locating and collecting data
Two hours completing forms
Three hours analyzing risks and drafting the report
One business day can disappear before the report is ready for internal review.
For more complex customers, the entire process may extend across five to ten working days—or longer if the institution needs additional documents, corrections, or approval from multiple departments.
The most expensive part is often not professional judgment.
It is repetitive operational work.
3. Report Quality Depends Too Heavily on Individual Experience
Traditional due diligence quality can vary significantly between employees.
Two relationship managers investigating the same company may focus on different indicators, use different evidence, and reach different conclusions.
An experienced analyst may recognize hidden risks in:
Related-party transactions
Abnormal cash flows
Shareholder changes
Legal disputes
Industry-cycle pressure
Customer concentration
Supply-chain dependency
A less experienced employee may simply complete the required fields without identifying the relationships behind them.
Without a standardized research process, due diligence becomes difficult to audit, reproduce, and scale.
The institution may have a standard template, but not necessarily a standard reasoning process.
Agent Loop: Automating the Research Process Itself
The core of DojoAgents is not a “smarter chatbot.”
It is an Agent Loop.
Traditional AI usually works in a single-turn pattern:
The user asks a question, and the model generates an answer.
A chatbot mainly considers:
How should I respond?
A financial agent must also determine:
What information do I still need, which tool should I use, and how can I verify the result?
The operating logic of an Agent Loop can be summarized as:
Think → Act → Observe → Think Again
After receiving a due diligence assignment, the agent first evaluates what evidence is required.
It may need to collect:
Corporate registration information
Shareholder and beneficial-owner data
Financial statements
Tax records
Judicial and enforcement records
Public opinion and news
Industry-chain relationships
Customer and supplier concentration
Historical transaction data
It then calls the appropriate tools, observes the results, identifies missing or conflicting information, and determines what should be checked next.
The loop continues until the available evidence is sufficient to support a structured report—or until the agent identifies items that must be escalated for human review.
This is fundamentally different from asking a model to write a report from one large prompt.
The agent does not simply generate a conclusion.
It follows a process.
From One Prompt to a Repeatable Due Diligence Workflow
A prompt-based system usually attempts to put everything into one instruction:
Which data to examine
Which risks to identify
Which report format to follow
Which industry factors to consider
Which conclusions to produce
That approach may work for a demonstration, but it is difficult to rely on in a high-stakes credit workflow.
Enterprise due diligence is not a one-step writing task.
It is a sequence of dependent actions.
A financial agent can divide the assignment into stages:
Identify the company and confirm entity information
Retrieve registration and ownership records
Collect financial and tax data
Search litigation, enforcement, and public-opinion records
Analyze industry and supply-chain relationships
Compare figures across different sources
Identify contradictions and missing evidence
Generate risk findings
Match the appropriate report template
Produce a reviewable draft with evidence references
The value of the Agent Loop is that each stage can influence the next.
If the agent discovers a major shareholder change, it can investigate the new shareholder.
If financial data conflicts with tax records, it can flag the discrepancy.
If a lawsuit involves a critical supplier, it can expand the analysis to supply-chain risk.
The system is not simply filling in blanks.
It is following clues.
A Comparable Banking Case: From More Than 990 Fields to Around 90
A banking case referenced in the original material illustrates what this kind of workflow transformation can look like.
In the pre-loan investigation process, the number of fields requiring manual completion was reduced from more than 990 to slightly over 90.
The human processing time for a single report fell from approximately two to three days to around two hours.
After the intelligent investigation report system went into operation, average daily agent calls reportedly exceeded 3,000.
These figures should not be interpreted as a guaranteed result for every bank or every due diligence scenario. The improvement depends on data availability, system integration, report complexity, and internal approval requirements.
But the direction is clear:
When machines handle data collection, field population, initial cross-checking, and report structuring, human employees can focus on exceptions and judgment.
How Can More Than 900 Manual Fields Be Removed?
The reduction does not come from deleting important information.
It comes from changing who collects and enters it.
Automatic Collection of Key Enterprise Information
The system can retrieve and organize information such as:
Main business activities
Shareholder structure
Financial indicators
Tax information
Upstream and downstream relationships
Judicial risks
Public-opinion signals
Industry classification
The agent can then cross-check related fields and identify inconsistencies.
The relationship manager no longer needs to log in to five or six systems and repeatedly copy the same information into a template.
The system performs the first round of retrieval and reconciliation.
Automatic Matching of the Due Diligence Framework
Different enterprises require different investigation priorities.
A manufacturing company should not be evaluated in exactly the same way as a software company. A small private enterprise may require a different risk framework from a listed company or state-owned enterprise.
The agent can select or adapt the investigation structure based on:
Customer type
Company size
Industry characteristics
Credit purpose
Financing structure
Collateral type
Existing risk signals
Instead of forcing every customer into the same rigid template, the system can determine which sections deserve deeper investigation.
This does not remove institutional standards.
It applies those standards more precisely.
Faster Draft Generation
Once corporate, financial, judicial, and public-opinion data have been connected, the system can automatically generate a first draft of the enterprise analysis.
The draft can include:
Company overview
Ownership and governance
Operating performance
Financial condition
Industry position
Litigation and compliance risks
Supply-chain relationships
Credit-risk observations
Missing evidence
Items requiring manual confirmation
The report is not necessarily ready for final approval.
But it is ready for professional review.
That distinction is important.
The objective is not to eliminate the credit officer.
It is to ensure that the officer begins with a structured evidence package rather than an empty template.
The Real Transformation Is Not Speed, but Systemization
Producing a report in two hours is an attractive result.
But speed is only the visible outcome.
The deeper transformation is that the entire chain becomes connected:
Information collection → evidence verification → risk penetration → analytical judgment → report generation
Traditional systems may accelerate one stage while leaving the rest dependent on manual work.
An agent-driven system attempts to connect the whole workflow.
The institution gains a more consistent process for:
Gathering evidence
Checking data
Recording sources
Identifying contradictions
Escalating exceptions
Producing standardized output
This makes the process easier to review and audit.
The most important benefit may not be that the fastest report becomes faster.
It may be that the average report becomes more consistent.
Large Models, Knowledge Graphs, and End-to-End Credit Agents
Some intelligent due diligence systems combine large language models with knowledge graphs.
The language model interprets documents, extracts information, and produces explanations. The knowledge graph connects companies, shareholders, executives, litigation records, customers, suppliers, and related entities.
Together, they can support deeper risk analysis.
For example, the system may discover that:
A customer and supplier share the same controlling person
A legal dispute involves a key upstream company
Several related companies use the same registered address
A sudden shareholder change occurred before a financing application
Financial growth is inconsistent with tax or transaction data
The value of a knowledge graph is not merely visualizing relationships.
It allows the agent to follow those relationships and decide where additional investigation is required.
In suitable scenarios, models such as DeepSeek-R1 can serve as part of a “credit super-agent,” helping coordinate data retrieval, reasoning, verification, and report drafting.
However, final credit decisions still require institutional governance, risk rules, human review, and clear responsibility boundaries.
Automation can accelerate the process.
It does not remove accountability.
Open-Source Frameworks Lower the Entry Barrier
Large banks have the resources to build proprietary intelligent due diligence platforms.
Smaller institutions and enterprise teams often do not.
Open-source frameworks such as DojoAgents can reduce the technical barrier to testing agent-based research workflows.
The project supports Python 3.11 and above. Depending on the environment and model configuration, a local financial research agent can reportedly be launched within approximately 10 to 15 minutes.
It can connect to multiple API-model ecosystems, including:
OpenAI
Anthropic
Gemini
DeepSeek
GLM
It also supports locally deployed models through frameworks such as:
Ollama
vLLM
This flexibility allows institutions to route different tasks to different models.
For example:
Complex risk reasoning can use a more capable model
Repetitive field extraction can use a lower-cost model
Sensitive customer data can remain within a local deployment
High-volume classification can use a faster endpoint
Final report review can use a higher-accuracy model
The model is no longer a fixed component.
It becomes a configurable resource within the workflow.
Compute Cost: The Final Barrier to Large-Scale Deployment
An Agent Loop can compress a due diligence process from weeks to hours.
But every loop consumes Tokens.
A single investigation may involve repeated model calls for:
Document parsing
Entity identification
Data extraction
Cross-source verification
Industry analysis
Relationship mapping
Risk classification
Report generation
Self-review
When one employee runs several reports, the cost may remain manageable.
When an institution runs thousands of calls every day, model spending becomes a serious engineering and financial constraint.
The key question changes from:
Can the agent complete the task?
to:
Can the institution afford to run the agent at scale?
This is where Unexhub enters the infrastructure layer.
Unexhub: Turning Token Capacity into a Market-Oriented Resource
Unexhub is building a Token aggregation and distribution network designed to connect model providers, compute suppliers, enterprise users, and developers.
Its goal is to transform Token capacity from a resource sold separately by individual vendors 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
Model capacity is spread across:
Major AI companies
Regional computing centers
Specialized model providers
Small and medium-sized AI enterprises
Independent development teams
Individual suppliers
Many smaller providers have valuable models or unused inference capacity but lack customer traffic, billing infrastructure, global distribution, and bargaining power.
Unexhub aggregates these resources into a unified supply pool and makes them available to enterprises and developers through standardized interfaces.
This gives users more model and endpoint choices, while providing suppliers with an additional commercialization channel.
2. Faster Settlement for Suppliers
Traditional enterprise payment cycles can create financial pressure for smaller model and compute suppliers.
Unexhub aims to calculate usage and supplier revenue based more closely on actual API consumption, shortening the period between providing capacity and receiving payment.
Suppliers can gain clearer visibility into:
Token usage
Revenue by model
Customer demand
Route performance
Settlement progress
Qualified early suppliers may also receive promotional commission rates as low as 0%.
The value is not simply a temporary reduction in platform fees.
It creates a lower-cost market-validation period, allowing suppliers to test demand, accumulate initial usage, and evaluate the competitiveness of their models before taking on standard channel costs.
3. API Bidding and Dynamic Pricing
Under traditional API pricing, one provider publishes a fixed rate.
Unexhub proposes a more market-oriented mechanism in which effective API prices can respond to:
Available capacity
Real-time demand
Model performance
Response speed
Endpoint stability
Peak and off-peak periods
Competition among suppliers
When capacity is abundant, developers may access more economical routes.
When demand rises, suppliers offering stable and high-quality capacity may receive stronger returns.
Token pricing therefore becomes more than a static number on a vendor’s official price page.
It becomes a market-matching result.
When DojoAgents Meets a Compute Trading Center
DojoAgents and Unexhub solve different parts of the same scaling problem.
DojoAgents rebuilds the due diligence workflow at the application layer.
It transforms data collection, cross-checking, risk analysis, and report generation into a repeatable Agent Loop.
Unexhub attempts to rebuild the cost structure at the infrastructure layer.
It aggregates supply, introduces dynamic routing and market competition, and accelerates supplier settlement.
One answers:
How can enterprise due diligence be automated?
The other answers:
How can continuous automated due diligence remain affordable?
When the DojoAgents Agent Loop is repeatedly collecting data, checking evidence, mapping risks, and generating reports, Unexhub can route these model requests toward more cost-effective capacity.
Developers can pay according to actual Token consumption rather than purchasing an unnecessarily large package in advance.
Depending on the selected model, endpoint availability, supplier competition, and promotional conditions, some routes may be available at prices as low as approximately 30% of comparable market reference rates.
For one report, the difference may appear limited.
For an institution processing thousands of investigations, risk checks, and document workflows, lower unit costs can accumulate into meaningful monthly and annual savings.
Unexhub’s Value for Developers and Suppliers
Unexhub is designed to aggregate access to mainstream domestic and international AI ecosystems, including services associated with:
Gemini
ChatGPT
Claude
DeepSeek
GLM
Independent and specialized models
Through one unified entry point, developers can reduce the need to maintain separate:
Model accounts
API formats
Authentication systems
Billing dashboards
Supplier agreements
Routing rules
Its planned capabilities include:
Multi-model API aggregation
Pay-as-you-use Token billing
Millisecond-level intelligent routing
Dynamic endpoint selection
API bidding
Centralized usage and budget management
Faster supplier revenue settlement
Enterprise integration and customization
For developers, the benefit is not only a lower advertised price.
They can match different parts of the due diligence workflow with different models and pay only for actual usage.
For suppliers, standardized access to a distribution channel can reduce the cost of acquiring customers, building billing systems, and collecting revenue.
From “A Smarter Model” to “An Agent That Does the Work”
Reducing enterprise due diligence from two to four weeks to two hours is not simply a story about model intelligence.
It is a story about workflow design.
The model is no longer waiting for someone to paste all the information into a prompt.
The agent can:
Decide what information is needed
Retrieve it through tools
Compare multiple sources
Identify inconsistencies
Follow risk clues
Select a reporting framework
Produce a reviewable draft
The institution’s employees are not removed from the process.
Their role changes.
They spend less time copying data and more time reviewing exceptions, challenging assumptions, and making accountable credit decisions.
DojoAgents makes the work more automated.
Unexhub aims to make the Token cost behind that work more sustainable.
That is the real transition:
From a model that can write a report to an agent that can perform the investigation—and from an agent that can run to an agent an institution can afford to run at scale.
从 2—4 周到 2 小时:金融智能体如何重构企业信用尽调流程
中文完整版
按照传统流程,一份企业信用尽职调查报告通常需要 2—4 周才能完成。
在真正开始分析之前,客户经理往往要先登录征信、税务、ERP、司法查询和银行内部系统等五六个平台。仅仅是收集、核对和整理原始材料,就可能花费数小时。
在部分传统工作流中,单份尽调报告需要人工核对或填写的字段超过 990 个。
现在,把这套流程交给一个金融智能体。
两个小时后,第一版完整报告已经可以进入人工复核。
这正是 DojoAgents 所代表的效率拐点。2026 年 7 月,DojoAgents 在 GitHub 开源仅一周多,便突破了 1000 Star。
它真正解决的,不只是“把报告写得更快”。
而是把报告背后的研究过程逐步自动化:
采集数据、调用工具、交叉验证、识别风险,再把证据整理成一份结构化结论。
从“2—4 周”压缩到“2 小时”,并不只是因为模型变得更聪明。
更重要的是,智能体开始学会自己推进工作。
企业尽调为什么这么慢?
传统企业信用尽调长期面临三个绕不开的问题。
一、信息分散在多个系统中
完成一份企业尽调所需的信息,几乎不可能集中在同一个平台。
企业经营数据可能存放在业务系统中,财务报表通常由客户以 Excel 形式提供,司法诉讼和执行记录分散在公开数据库,税务、股权结构、舆情信息、行业数据和内部交易记录又分别来自不同系统。
客户经理需要在 6—10 个平台之间反复切换:
搜索企业主体
下载数据
复制信息
核对企业名称
统一字段格式
比较冲突数据
将结果粘贴到尽调模板
问题并不是没有数据。
而是数据彼此割裂。
真正开始进行风险判断之前,客户经理的大量时间已经消耗在“把数据搬到一起”这件事上。
二、手工填报占据了大部分时间
传统企业尽调报告通常包含大量必填字段。
按照原文描述的业务口径,单份报告可能需要人工核对或填写 990 多个字段。
典型时间分配可能是:
查找和整理数据 3 小时
填写表格 2 小时
分析风险、撰写报告 3 小时
一个完整工作日很快就被消耗掉。
如果客户结构复杂,或需要补充材料、修正数据、等待多个部门确认,整个流程还可能延长至 5—10 个工作日,甚至更久。
真正昂贵的环节,往往不是专业判断。
而是大量重复性的操作工作。
三、尽调质量过度依赖个人经验
传统尽调质量在很大程度上依赖客户经理和审查人员的个人经验。
面对同一家企业,不同人员可能关注不同指标、采用不同证据,最终得出完全不同的结论。
经验丰富的人员可能快速发现:
关联交易
异常现金流
股东频繁变更
重大司法纠纷
行业周期压力
客户集中度风险
对单一供应商的高度依赖
经验不足的人员则可能只完成表面字段,却没有识别这些信息之间的关系。
机构可能拥有标准化模板,却未必拥有标准化的研究过程。
这使尽调结论难以审计、复现和规模化复制。
Agent Loop:把“研究”这个过程本身自动化
DojoAgents 的核心,并不是一个“更聪明的聊天机器人”。
而是一套 Agent Loop(智能体循环)。
传统 AI 通常采用一次性问答模式:
用户提出问题,模型生成答案。
聊天机器人主要考虑:
这个问题应该怎么回答?
金融智能体还必须进一步判断:
当前还缺少什么信息、应该调用哪个工具,以及怎样验证结果?
Agent Loop 的运行逻辑可以概括为:
思考 → 行动 → 观察 → 继续思考
收到企业尽调任务后,智能体会先判断完成调查需要哪些证据。
例如:
工商登记信息
股东及实际控制人情况
财务报表
税务记录
司法诉讼与执行信息
新闻和舆情
产业链关系
客户与供应商集中度
历史交易数据
随后,智能体调用相应工具获取信息,观察返回结果,识别缺失项或矛盾数据,再决定下一步继续调查什么。
这个循环会持续推进,直到现有证据足以支持一份结构化报告,或系统识别出必须交由人工判断的异常事项。
这与把所有材料塞进一个长 Prompt,再让模型一次性生成报告,存在本质区别。
智能体不只是输出结论。
它会遵循一套调查流程。
从一个 Prompt,走向可重复执行的尽调流程
传统 Prompt 方案往往试图把所有要求写进一次指令:
需要查看哪些数据
应该识别哪些风险
使用什么分析框架
按什么格式输出
不同行业需要关注什么
这种方法可以用于演示,却难以支撑高风险的授信工作流。
企业尽调并不是一次写作任务,而是一系列相互依赖的动作。
金融智能体可以将任务拆分为多个阶段:
确认企业主体与基础信息
查询工商登记和股权结构
获取财务与税务数据
检索诉讼、执行和舆情记录
分析行业与上下游关系
对比不同数据源中的数值
识别矛盾、缺失和异常项
形成风险判断
匹配对应的尽调模板
输出附带证据依据的初稿
每一个阶段的结果,都可能改变下一步行动。
如果系统发现企业近期出现重大股东变更,就可以继续调查新股东背景。
如果财务数据与税务数据不一致,可以主动标记异常。
如果诉讼涉及关键供应商,还可以进一步分析供应链风险。
系统不再只是填写空格。
而是在沿着风险线索继续调查。
银行业参照案例:从 990 多个字段降至 90 多个
原文引用的银行业案例,很能说明这类工作流改造可能带来的效果。
在贷前调查环节,需要人工填写的字段从 990 多个减少到 90 多个。
单份报告的人工处理时间,也从约 2—3 天缩短至 2 小时左右。
智能调查报告上线后,智能体日均调用量据称超过 3000 次。
这些数据并不意味着所有银行、所有尽调项目都一定能够获得相同结果。实际效果仍取决于数据接入程度、系统集成情况、报告复杂度和内部审批要求。
但其趋势非常明确:
当机器接管数据采集、字段填写、初步核验和报告整理后,人可以把精力集中在异常情况与专业判断上。
900 多个人工字段,是怎样被省下来的?
字段减少并不是因为重要信息被删除了。
而是因为信息采集和录入的执行者发生了变化。
自动获取企业关键信息
系统可以主动检索并整理:
企业主营业务
股权结构
财务指标
税务信息
上下游关系
司法风险
舆情信号
行业分类
智能体还可以对相关字段进行交叉验证,识别不同数据源之间的不一致。
客户经理不再需要手动登录五六个系统,再把同一份信息重复复制到报告模板中。
系统先完成第一轮数据获取与核对。
自动匹配尽调框架
不同企业需要关注的尽调重点完全不同。
制造企业与软件公司的风险结构并不一样;小型民营企业与上市公司、国有企业也不能使用完全相同的审查逻辑。
智能体可以根据以下因素选择或调整尽调框架:
客户类型
企业规模
行业特征
授信用途
融资结构
担保方式
已识别的风险信号
它并不是取消机构原有的标准。
而是让标准更精准地应用于不同客户。
更快生成可复核的报告初稿
当工商、财务、司法和舆情数据被连接起来后,系统可以自动生成企业分析初稿。
内容可以包括:
企业概况
股权与治理结构
经营表现
财务状况
行业地位
司法与合规风险
供应链关系
授信风险观察
缺失材料
需要人工确认的事项
这份报告并不一定可以直接进入最终审批。
但它已经具备人工复核的基础。
这一点非常重要。
智能体的目标并不是取消客户经理或审查人员,而是让他们面对的不再是一张空白模板,而是一套已经完成初步整理的证据材料。
真正的变化不只是“快”,而是“系统化”
两个小时完成报告,是最直观的结果。
但更深层的改变,是整个链路被连接了起来:
信息采集 → 证据核验 → 风险穿透 → 智能分析 → 报告生成
传统系统可能只在某一个阶段提速,剩余环节依然依靠人工推动。
智能体系统则试图连接整套流程。
机构能够采用更加一致的方式:
获取证据
核对数据
记录来源
识别矛盾
上报异常
生成标准化输出
这使尽调流程更容易被复核与审计。
真正重要的收益,未必只是最快的报告变得更快。
而是平均每一份报告的质量都变得更加稳定。
大模型、知识图谱与端到端授信智能体
部分智能尽调平台还会将大模型与知识图谱结合。
大模型负责理解文档、提取信息和生成解释;知识图谱则负责连接企业、股东、高管、司法案件、客户、供应商和关联主体。
两者结合,可以支持更深入的风险穿透。
例如,系统可能发现:
某个客户与供应商由同一实际控制人管理
一起诉讼涉及企业的关键上游供应商
多家关联公司使用相同注册地址
企业在融资申请前突然发生股权变更
财务增长与税务或交易数据明显不一致
知识图谱的价值,不只是把关系画出来。
而是让智能体可以沿着关系继续追踪,判断哪里需要进一步调查。
在合适的工作流中,DeepSeek-R1 等模型可以作为“授信超级智能体”的推理组件,协调数据查询、分析、验证与报告生成。
但最终授信决策仍然需要制度治理、风险规则、人工复核与明确的责任边界。
自动化可以加快流程。
但不能取消责任。
开源框架,让智能尽调不再只是大型银行的专属能力
大型银行有资源建设自有智能尽调平台。
中小型机构和普通企业团队,往往没有相同的研发预算。
DojoAgents 等开源框架,可以降低测试智能体研究流程的技术门槛。
项目支持 Python 3.11 及以上版本。根据环境和模型配置的不同,用户大约可以在 10—15 分钟内拉起一个本地金融研究智能体。
它可以接入多个 API 模型生态,包括:
OpenAI
Anthropic
Gemini
DeepSeek
GLM
也可以通过以下框架部署本地模型:
Ollama
vLLM
这种灵活性让机构可以为不同任务选择不同模型。
例如:
复杂风险判断使用高能力模型
重复字段提取使用低成本模型
敏感客户数据交给本地模型处理
大批量分类使用高速模型
最终报告复核使用更高准确率的模型
模型不再是一项固定配置。
而是工作流中可以被灵活调度的资源。
算力成本:规模化部署的最后一道门槛
Agent Loop 可以把企业尽调从数周压缩到数小时。
但每一次循环都需要消耗 Token。
一份尽调报告可能需要反复调用模型完成:
文档解析
企业主体识别
数据提取
多源信息核验
行业分析
关联关系映射
风险分类
报告生成
结果自检
当少数员工每天处理几份报告时,成本可能仍然可控。
但当一家机构每天运行数千次调用时,模型费用就会成为必须面对的工程和财务约束。
问题开始从:
智能体能不能完成任务?
变成:
机构能不能长期承担智能体规模化运行的成本?
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 正在建设连接模型供应方、算力供应方、企业用户和开发者的 Token 聚合与分发网络。
它希望把原本由不同厂商分别出售的 Token 能力,转化为可以通过算力交易中心统一聚合、路由、定价、分发和结算的资源。
这套模式可以拆分为三项机制。
第一层:Token 聚合与分发
模型算力分散在:
大型 AI 企业
区域算力中心
垂直模型供应商
中小型 AI 企业
独立开发团队
个人供应方
不少中小型供应方拥有优质模型或闲置推理容量,却缺乏用户流量、计费系统、全球分发渠道和议价能力。
Unexhub 将这些资源聚合进统一供给池,再通过标准化接口提供给企业和开发者。
使用者可以获得更多模型与节点选择,供应方也多了一条商业化渠道。
第二层:加快供应方收益结算
传统企业账期会给中小型模型和算力供应方带来资金压力。
Unexhub 希望根据真实 API 用量,更及时地计算调用收入,缩短提供算力与回收收益之间的周期。
供应方可以更加清晰地查看:
Token 调用量
不同模型的收入
用户需求
路由表现
结算进度
符合条件的早期供应方,还可获得最低 0% 佣金的阶段性权益。
这项权益的价值,并不只是暂时节省平台费用。
它为供应方提供了一段低成本的市场验证期,可以先测试需求、积累初始调用量,并评估模型的商业竞争力,再承担常规渠道成本。
第三层:API 竞价与动态价格
传统 API 模式由单一厂商公布固定价格。
Unexhub 希望引入更市场化的机制,让实际 API 价格可以根据以下因素变化:
当前可用容量
实时市场需求
模型性能
响应速度
节点稳定性
高峰和低谷时段
不同供应方之间的竞争
当算力供给充足时,开发者可以获得更具成本优势的线路。
需求高峰期,能够稳定提供优质算力的供应方,也可以获得更合理的收益。
Token 价格不再只是一张厂商官网上的固定价目表。
而逐渐成为供需匹配后的市场结果。
当 DojoAgents 遇上算力交易中心
DojoAgents 与 Unexhub 解决的是规模化部署中的两个不同问题。
DojoAgents 重构应用层的尽调工作流。
它把数据采集、交叉验证、风险分析和报告生成,变成一条可以重复运行的 Agent Loop。
Unexhub 则尝试重构基础设施层的成本结构。
它通过聚合供给、动态路由、市场竞价和快速结算,降低长期调用模型的成本。
一个回答:
企业尽调怎样实现自动化?
另一个回答:
自动化尽调持续运行时,怎样才能在成本上可承受?
当 DojoAgents 的 Agent Loop 在后台持续采集数据、核验证据、穿透风险并生成报告时,Unexhub 可以尝试将这些模型请求路由至更具成本竞争力的算力线路。
开发者可以按照实际 Token 用量付费,不必提前购买远高于真实需求的固定套餐。
根据模型类型、节点供应、市场竞争和阶段性权益的不同,部分线路最低可享受约为市场参考价格 3 折的成本方案。
对于一份报告而言,价格差异可能并不明显。
但对于每天处理数千次调查、风险核验和文档工作流的机构而言,单位成本下降会持续累积,最终转化为更加明显的月度和年度预算节省。
Unexhub 为开发者与供应方提供什么
Unexhub 致力于聚合海内外主流 AI 模型生态,包括与以下模型相关的服务能力:
Gemini
ChatGPT
Claude
DeepSeek
GLM
独立模型与专业垂直模型
通过统一入口,开发者可以减少对以下内容的重复维护:
多套模型账户
不同 API 格式
多种鉴权方式
分散计费后台
多家供应商协议
手动路由规则
平台规划中的核心能力包括:
多模型 API 聚合
按 Token 实际用量付费
毫秒级智能路由
动态节点选择
API 竞价
集中化用量与预算管理
供应方收益快速结算
企业级接入与定制服务
对开发者而言,价值不只是一个看起来更低的价格。
他们可以根据尽调流程中不同任务的难度,为数据提取、风险推理和报告复核匹配不同模型,并只为真实用量付费。
对供应方而言,通过标准化方式接入分发渠道,可以降低获客、计费与回款成本。
从“更聪明的模型”,走向“真正会干活的智能体”
把企业尽调从 2—4 周缩短到 2 小时,并不只是一场模型能力升级。
它首先是一场工作流重构。
模型不再等待工作人员把所有信息整理好,再粘贴进 Prompt。
智能体开始主动:
判断需要什么信息
调用工具获取数据
比较不同来源
识别矛盾和异常
沿着风险线索继续调查
选择适合的报告框架
输出可供人工审核的初稿
机构工作人员并没有从流程中消失。
他们的工作重点发生了变化。
他们可以减少复制粘贴和机械填表,把更多时间投入异常复核、假设质疑与最终授信判断。
DojoAgents 让尽调工作变得更加自动化。
Unexhub 则希望让支撑这些工作的 Token 成本更加可持续。
这才是真正的转变:
从“模型会写尽调报告”,走向“智能体会执行尽调”;也从“智能体能够运行”,走向“机构能够长期、规模化地负担智能体运行”。