麦肯锡说仅 1% 医健机构没碰 agent 应用,但可信合规生态还在补课

At the Forum on AI-Powered Innovation in Healthcare, You Mao, deputy director of the National Health Development Research Center, emphasized that AI in healthcare is no longer merely a technical issue. It has become a strategic topic involving public health, industrial upgrading, and national competitiveness.

McKinsey Says Only 1% of Healthcare Organizations Have Not Touched Agentic AI—Yet the Trusted and Compliant Ecosystem Is Still Catching Up

Bilingual Article|English + Chinese
The complete English version appears first.
中文完整版位于文章后半部分,请继续向下阅读。

In July 2026, at the China Internet Conference in Beijing, one message became impossible to ignore:

Healthcare is no longer debating whether to adopt AI.

It is trying to determine how to use AI safely, effectively, and at scale.

At the Forum on AI-Powered Innovation in Healthcare, You Mao, deputy director of the National Health Development Research Center, emphasized that AI in healthcare is no longer merely a technical issue. It has become a strategic topic involving public health, industrial upgrading, and national competitiveness.

Representatives from government, industry, academia, research institutions, and hospitals discussed topics ranging from foundation-model deployment and evidence-based medical agents to clinical decision support.

The question has changed.

It is no longer:

Should healthcare institutions use AI?

It is:

How can they use it well—and keep it trustworthy, compliant, and economically sustainable?

McKinsey’s data illustrates the scale of that transition.


99% Have Entered the Conversation, but Only 19% Have Truly Scaled

According to McKinsey’s Q4 2025 survey of generative AI in U.S. healthcare, approximately 50% of healthcare organizations had already deployed generative AI technologies—double the 25% reported at the end of 2023.

More than 80% had introduced at least one use case to end users, and roughly half had completed their first deployment more than six months earlier.

Agentic AI penetration appears even broader.

Only 1% of healthcare organizations reported no exposure to agentic applications.

In other words:

Approximately 99% had at least explored, tested, discussed, or monitored agentic AI.

But awareness is not the same as operational maturity.

Only around 19% had reached large-scale deployment, while more than half remained stuck in proof-of-concept or pilot stages.

This gap reveals the central challenge facing healthcare AI.

Most institutions already understand what agents could do.

They have not yet solved how to make those agents operate safely, compliantly, reliably, and at a sustainable cost.


What Can Healthcare Agents Actually Do?

McKinsey’s research outlines a broad role for agents across the patient journey.

Before the visit

Agents may help patients:

  • Identify an appropriate provider or care facility;

  • Prepare appointment materials;

  • Organize medical history;

  • Complete preliminary questionnaires;

  • Understand what documents or tests may be required.

During treatment

Agents may:

  • Coordinate patient data across institutions;

  • Retrieve previous records;

  • Summarize medical histories;

  • Assist with clinical documentation;

  • Surface relevant guidelines or potential risks.

After the visit

Agents may:

  • Generate personalized rehabilitation plans;

  • Provide medication reminders;

  • Organize follow-up appointments;

  • Track recovery progress;

  • Escalate abnormal symptoms to clinical staff.

On the administrative side, agents may also:

  • Submit insurance claims;

  • Identify missing documentation;

  • Correct billing records;

  • Draft appeal materials;

  • Follow up on rejected claims.

This is not simply a chatbot answering medical questions.

It is a network of specialized agents participating across the entire healthcare workflow.


Revenue-Cycle Management Shows the Immediate Economic Value

Revenue-cycle management is one of the clearest commercial opportunities.

The sector represents more than $140 billion in annual spending—approximately 3% to 4% of total healthcare-system revenue.

McKinsey estimates that agentic AI could reduce collection-related costs by 30% to 60%.

For a healthcare system with $6 billion in annual revenue, every 1% reduction in collection costs could translate into approximately $6 million in savings.

That is why administrative agents are often deployed before fully autonomous diagnostic systems.

The financial return is easier to calculate, the workflow is more standardized, and the clinical risk is lower.

Healthcare organizations can begin with:

  • Claim processing;

  • Documentation review;

  • Billing reconciliation;

  • Patient scheduling;

  • Insurance verification;

  • Administrative follow-up.

These are high-volume workflows where automation can create immediate measurable value.


The Life-Sciences Opportunity Is Even Larger

The economic potential extends beyond hospitals.

McKinsey estimates that generative AI could unlock between $60 billion and $110 billion in annual value for pharmaceutical and medical-device companies.

Between 75% and 84% of pharmaceutical workflows may be enhanced or partially automated by agentic AI.

Potential applications include:

  • Drug-discovery research;

  • Clinical-trial documentation;

  • Regulatory submissions;

  • Medical-affairs content;

  • Safety monitoring;

  • Commercial analysis;

  • Physician engagement;

  • Patient-support programs.

The opportunity is enormous.

But attractive numbers do not eliminate the difficulty of implementation.


Healthcare Is Still Taking Its Compliance Course

Risk and safety remain the most important obstacles.

McKinsey’s research indicates that 43% of respondents identify risk and security as a major barrier to implementation.

The concerns are familiar but particularly serious in healthcare:

  • Inaccurate outputs;

  • Biased recommendations;

  • Data leakage;

  • Unauthorized access;

  • Regulatory non-compliance;

  • Untraceable reasoning;

  • Unclear accountability.

A wrong movie recommendation is inconvenient.

A wrong medical recommendation can affect a patient’s health.

That difference changes the entire deployment standard.


China Is Beginning to Build a Systematic Agent-Governance Framework

In May 2026, the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly issued the Implementation Opinions on the Standardized Application and Innovative Development of AI Agents.

It was the country’s first systematic policy document focused specifically on agents.

The principles emphasized:

  • Safety and controllability;

  • Standardized and orderly development;

  • Innovation-driven progress;

  • Application-oriented deployment.

High-risk agents—particularly those involved in medical-image diagnosis and surgical assistance—may be subject to filing requirements, mandatory security testing, and clinical validation.

This is an important transition.

The industry is moving from encouraging experimentation to defining the conditions under which agents may be trusted with real responsibilities.


Industry Consensus Is Also Emerging

At the China Internet Conference, Ewell Health and other participants initiated a trusted-AI healthcare application initiative.

The focus was not simply on promoting AI adoption, but on building agreement around four principles:

  • Trustworthiness;

  • Security;

  • Standardization;

  • Sustainability.

That last word matters.

A healthcare agent is not successful merely because it works in a demonstration.

It must continue operating under real budgets, real regulations, real hospital systems, and real clinical accountability.


Compliance Has a Cost

Compliance is not an abstract principle.

It requires people, systems, documentation, and continuous investment.

The cost can be divided into three layers.

1. Foundational compliance costs

These include:

  • Governance policies;

  • Compliance teams;

  • Staff training;

  • Internal procedures;

  • Audit mechanisms;

  • Documentation and reporting.

2. Technical compliance costs

These may include:

  • Security-by-design;

  • Identity and access management;

  • Permission controls;

  • Data encryption;

  • Logging;

  • Model monitoring;

  • Version management;

  • Failure recovery.

3. Scenario-specific costs

Healthcare faces additional requirements such as:

  • Medical-device registration;

  • Clinical validation;

  • Medical-data governance;

  • Algorithm filing;

  • Evidence traceability;

  • Human clinical oversight.

Large tertiary hospitals may be able to absorb these expenses.

Smaller hospitals, community clinics, regional healthcare providers, and independent developers face a much higher barrier.

For them, compliance costs do not exist in isolation.

They are layered on top of model costs, cloud costs, data-management expenses, and Token consumption.


The Token Bill Grows With Every New Agent

Medical-agent ecosystems are becoming increasingly complex.

iFLYTEK’s four-agent clinical system can generate more than 100,000 outpatient records.

Tencent Health has introduced 14 specialized healthcare-agent capabilities.

Medical-image reasoning frameworks may invoke multiple expert models at different stages of analysis.

A single clinical interaction may trigger:

  • Speech recognition;

  • Medical-record generation;

  • Quality review;

  • Risk detection;

  • Guideline retrieval;

  • Evidence verification;

  • Diagnostic assistance;

  • Follow-up planning.

As the number of agents increases, Token consumption can grow rapidly.

The relationship is not always linear.

Agents may repeatedly read the same records, exchange intermediate results, verify one another’s outputs, call specialist models, and retry failed steps.

That creates a difficult equation:

The institutions that would benefit most from AI may be the least able to afford the infrastructure required to run it safely.

This is particularly true for:

  • Smaller hospitals;

  • Grassroots medical institutions;

  • Independent medical developers;

  • Regional health platforms;

  • Specialized AI suppliers.

The next bottleneck is therefore not only intelligence.

It is access to affordable, stable, and compliant computing resources.


Unexhub Is Positioned at This Critical Infrastructure Layer

Unexhub is attempting to restructure how global Tokens and AI computing resources are aggregated, distributed, priced, and settled.

Its core positioning is a computing-power trading center focused on Token aggregation and distribution.

The platform connects fragmented AI supply with demand from enterprises, developers, and intelligent applications.

For healthcare-agent ecosystems, a unified infrastructure layer can address several practical problems:

  • Multiple model providers;

  • Different pricing systems;

  • Separate API formats;

  • Fragmented settlement cycles;

  • Unstable resource availability;

  • High switching costs;

  • Limited access for smaller participants.

A hospital may need one model for medical documentation, another for imaging, a third for evidence retrieval, and a premium model for high-risk review.

Maintaining a separate commercial and technical relationship with every provider creates significant cost and complexity.

A unified aggregation and routing layer can reduce that burden.


Faster Settlement Can Reduce Pressure on Smaller Suppliers

Cash flow is one of the most serious challenges for small and medium-sized AI suppliers.

They may provide computing capacity immediately, while waiting weeks or months to receive payment.

During that period, they must continue covering:

  • GPU expenses;

  • Electricity;

  • Bandwidth;

  • Cloud infrastructure;

  • Operations;

  • Engineering salaries.

Even a technically strong supplier may struggle if the settlement cycle is too long.

Unexhub’s settlement and clearing system is designed to support faster revenue distribution.

The objective is simple:

When computing capacity is consumed, the corresponding supplier revenue can be returned more quickly.

For smaller suppliers, faster settlement can:

  • Reduce working-capital pressure;

  • Improve cash-flow predictability;

  • Accelerate infrastructure reinvestment;

  • Support more stable service delivery;

  • Lower the risk of expansion.

Unexhub also plans early-supplier incentives.

Eligible early participants may receive promotional commission rates as low as 0%, allowing them to retain more of their initial usage revenue while building distribution volume.

The value is not merely “free platform access.”

It may provide:

  • Lower customer-acquisition costs;

  • Faster access to enterprise demand;

  • Shorter payment cycles;

  • More room to test pricing strategies;

  • A faster path to recover infrastructure investment.


API Bidding Could Turn Discounts Into a Market Mechanism

Another important part of the Unexhub model is API bidding.

Traditional model APIs use fixed list prices.

Users pay the same published rate even when infrastructure is underutilized.

A market-based mechanism can allow prices to respond to real supply and demand.

During off-peak periods, developers may access unused capacity at lower prices.

During demand peaks, suppliers may earn higher returns for providing scarce resources.

For users, this may offer:

  • Lower off-peak prices;

  • Greater model choice;

  • More flexible workload scheduling;

  • Reduced dependence on one supplier;

  • Better control of large Token budgets.

For suppliers, it may provide:

  • Higher utilization of idle capacity;

  • Access to broader demand;

  • Better peak-period returns;

  • Faster settlement;

  • Lower channel-distribution costs.

This is more than a promotional discount.

It attempts to return part of Token pricing to market supply and demand.


Medical AI Needs Dynamic Resource Routing

Healthcare workflows do not require the same model for every task.

Healthcare taskSuitable resource
Speech transcriptionFast, low-cost model
Medical-record draftingBalanced language model
Documentation reviewSpecialized quality model
Medical-image analysisMultimodal medical model
Guideline retrievalSearch and evidence agent
High-risk diagnosis reviewPremium reasoning model
Patient follow-upLow-cost conversational model

Sending every request to the most expensive flagship model would make large-scale deployment difficult.

A more sustainable architecture routes tasks according to:

  • Clinical risk;

  • Complexity;

  • Latency requirements;

  • Data type;

  • Budget;

  • Availability;

  • Regulatory requirements.

This is where a Token marketplace becomes more than a low-cost API shop.

It becomes a resource-allocation layer.


Unexhub’s Aggregation and Millisecond-Level Routing

Unexhub is designed to aggregate model and computing suppliers at different scales.

Its ecosystem may include access to mainstream domestic and international models such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Other general and specialized AI models.

The platform can also connect high-quality models and computing resources from individuals, startups, and small or medium-sized suppliers.

Requests can then be routed according to factors such as:

  • Model capability;

  • Price;

  • Latency;

  • Availability;

  • Region;

  • Task type;

  • Service quality.

Millisecond-level routing can be especially valuable in medical workflows.

A low-risk administrative query should not consume the same resources as a high-risk clinical review.

Dynamic routing can reserve expensive capacity for the tasks that genuinely require it.


Pay for Actual Token Usage

Usage-based settlement can further reduce barriers for smaller institutions and developers.

Instead of pre-purchasing large packages from multiple providers, users can pay according to actual Token consumption.

This creates several practical benefits:

  • Start with a smaller budget;

  • Test multiple models;

  • Scale gradually;

  • Avoid unused prepaid resources;

  • Switch models without rebuilding the application;

  • Match spending more closely to real patient or workflow volume.

Under selected supply and promotional conditions, effective API pricing may be as low as approximately 30% of standard market rates.

For high-volume medical agents, the value is not merely saving a few cents per request.

The same budget may support:

  • More patient consultations;

  • More medical records;

  • More evidence retrieval;

  • Additional verification;

  • Longer agent workflows;

  • Wider deployment across smaller institutions.

Of course, low price alone is not enough for healthcare.

Any model-access platform serving medical workloads must also be evaluated for:

  • Data security;

  • Privacy;

  • Reliability;

  • Regional compliance;

  • Auditability;

  • Access controls;

  • Service continuity;

  • Failure recovery.

Affordability becomes meaningful only when it is combined with safe and stable infrastructure.


Where PoloAPI Fits Into the Practical Deployment Layer

Unexhub addresses the broader Token-market and computing-resource infrastructure.

At the application-access level, multi-model aggregation platforms such as PoloAPI can help developers and enterprises connect mainstream models through a unified interface.

PoloAPI supports access to models including:

  • Gemini;

  • Claude;

  • ChatGPT;

  • DeepSeek;

  • Other domestic and international models.

For healthcare developers or enterprise AI teams, this can reduce the need to maintain separate integrations, accounts, and billing systems for every provider.

The practical benefits may include:

  • Unified API access;

  • Centralized usage and cost visibility;

  • Easier model switching;

  • Lower multi-provider integration costs;

  • More flexible backup routing;

  • Enterprise customization;

  • Support for high-concurrency workloads.

Under selected resource plans, effective prices may be as low as approximately 30% of official rates.

For teams running frequent or long-lived agent workflows, this can make the same budget stretch further—supporting more users, more model calls, and more independent verification.

PoloAPI also provides enterprise customization processes covering model selection, integration, permissions, concurrency planning, and deployment.

Its service experience includes large companies, publicly listed enterprises, state-owned organizations, and international customers.

The role is complementary:

  • Unexhub focuses on Token aggregation, distribution, bidding, and settlement;

  • PoloAPI can provide a practical multi-model access and enterprise-integration layer.

Together, these types of infrastructure can help reduce both economic and technical barriers to agent deployment.


Two Industrial Shifts Are Converging

Healthcare agents are turning clinical work into a coordinated workflow.

Medical-record generation, quality review, diagnostic support, evidence retrieval, and patient follow-up may each be handled by specialized agents.

Doctors can spend less time on repetitive documentation and more time on clinical judgment and patient communication.

But the automated workflow can operate only if the underlying resources remain available.

It requires:

  • Stable model access;

  • Sufficient Token capacity;

  • Dynamic routing;

  • Transparent pricing;

  • Faster settlement;

  • Backup suppliers;

  • Reliable infrastructure.

Healthcare-agent companies are answering:

What should the agents do?

Infrastructure platforms are answering:

How can the computing resources required by those agents be delivered sustainably?

The real industrial transformation appears where those questions meet.


The Challenges Remain

The direction is clear, but the transition will not happen overnight.

Healthcare institutions still need to answer difficult questions:

  • Can existing systems support multiple agents?

  • Can data move securely across departments?

  • Are AI recommendations traceable?

  • Can diagnostic evidence be independently verified?

  • Who is accountable when an agent makes a mistake?

  • Can grassroots institutions afford the required Token volume?

  • Can suppliers provide stable capacity during demand peaks?

Medical AI cannot be deployed like an ordinary office chatbot.

It operates in a high-risk environment where mistakes can affect human health.

A mature medical-agent ecosystem will require:

  • Clinical human oversight;

  • Evidence traceability;

  • Permission controls;

  • Independent validation;

  • Model monitoring;

  • Secure data governance;

  • Reliable computing and Token supply.

Technology is moving from isolated intelligence to full-process collaboration.

Governance and infrastructure must catch up.


The Bottom Line

Approximately 99% of healthcare organizations have already encountered agentic AI in some form.

Only around 19% have scaled it.

That gap is not simply a technology problem.

It is a trust, compliance, integration, infrastructure, and cost problem.

Healthcare agents are beginning to rewrite the clinical workflow.

They can generate records, inspect quality, support diagnosis, retrieve evidence, coordinate claims, and guide patient follow-up.

But every agent consumes Tokens, tools, data access, and computing capacity.

Unexhub is attempting to rebuild the Token distribution and settlement logic underneath that ecosystem.

PoloAPI and similar aggregation platforms can help simplify practical multi-model access and enterprise deployment.

One layer defines what healthcare agents can do.

Another determines how the models and computing resources reach the organizations that need them.

The future of healthcare AI will not be decided by model intelligence alone.

It will be decided by whether the entire system can remain trustworthy, compliant, affordable, and continuously available.

The 99% have already entered the game.

The next challenge is learning how to play it well.


中文版

麦肯锡称仅 1% 医疗机构未接触智能体应用,但可信合规生态仍在补课

2026 年 7 月,北京,中国互联网大会。

在“人工智能赋能医疗健康创新应用论坛”上,国家卫健委卫生发展研究中心副主任游茂指出:

人工智能赋能医疗健康,已不仅是技术问题,更是关乎人民健康福祉、产业升级与国家竞争力的重要课题。

会上,来自政府、产业、学术界、研究机构和医疗系统的代表,围绕 AI 模型落地、循证医学智能体和临床辅助诊疗展开讨论。

一个明显变化是:

已经很少有人再问“医疗机构要不要使用 AI”。

所有人都在问:

AI 应该怎么用,才能真正做到可信、合规、安全,并且能够持续运行?

麦肯锡的数据,为这场转型提供了更加清晰的注脚。


99% 已经入局,但只有 19% 真正跑起来

根据麦肯锡 2025 年第四季度开展的美国医疗生成式 AI 调查,约 50% 的美国医疗机构已经部署生成式 AI 技术,较 2023 年底的 25% 翻了一倍。

超过 80% 的机构已经将首批应用推向终端用户,其中约半数在六个月前便完成了初步部署。

智能体 AI 的渗透率更值得关注。

仅有约 1% 的医疗机构表示完全没有接触智能体应用。

换句话说:

约 99% 的医疗机构,至少已经关注、尝试、测试或讨论过智能体 AI。

但“接触过”不等于“真正跑起来”。

只有约 19% 的机构进入大规模部署阶段,超过半数依然停留在概念验证或试点环节。

这一断层揭示了医疗 AI 当前最核心的问题:

大多数机构已经知道智能体能够做什么,却还没有完全解决如何让智能体安全、合规、稳定并且以可持续成本运行。


医疗智能体究竟能做什么?

麦肯锡的研究,较为完整地勾勒了医疗智能体在患者全流程中的作用。

就诊前

智能体可以帮助患者:

  • 匹配合适的医疗机构和护理资源;

  • 准备预约材料;

  • 整理既往病史;

  • 完成预问诊;

  • 明确就诊前需要准备的检查和文件。

就诊中

智能体可以:

  • 协调跨机构病例数据;

  • 调取历史病历;

  • 总结患者医疗记录;

  • 辅助临床文书生成;

  • 提示相关指南与潜在风险。

就诊后

智能体可以:

  • 生成个性化康复建议;

  • 提醒患者用药;

  • 安排复诊;

  • 跟踪恢复情况;

  • 在发现异常症状时通知临床人员。

在行政端,智能体还可以:

  • 自动提交理赔;

  • 识别文书缺项;

  • 修正账单记录;

  • 起草申诉材料;

  • 跟进被拒绝的保险申请。

这已经不只是一个回答医学问题的聊天机器人。

而是一套覆盖医疗流程不同阶段的专业智能体网络。


收入周期管理,最先体现直接经济价值

收入周期管理是医疗智能体最容易产生明确商业价值的场景之一。

这一领域每年支出超过 1,400 亿美元,约占医疗系统收入的 3% 至 4%。

麦肯锡估算,智能体 AI 可以将催收与回款相关成本降低约 30% 至 60%。

对于一家年收入 60 亿美元的医疗系统而言,催收成本每降低 1%,便可能节省约 600 万美元。

这也是为什么许多医疗机构优先部署行政型智能体,而不是直接上线高度自主的诊断系统。

行政流程更加标准化,经济回报更加容易计算,临床风险也相对较低。

医院可以优先从以下场景开始:

  • 理赔处理;

  • 文书审查;

  • 账单核对;

  • 患者预约;

  • 保险验证;

  • 行政随访。

这些都是调用量高、规则相对清晰,并且能够快速衡量投入产出的流程。


生命科学领域的空间更大

智能体带来的经济价值并不只存在于医院。

麦肯锡估算,生成式 AI 每年可为制药与医疗器械行业释放约 600 亿至 1,100 亿美元的经济价值。

制药行业约 75% 至 84% 的工作流,可以通过智能体 AI 得到增强或部分自动化。

潜在应用包括:

  • 药物研发;

  • 临床试验文档;

  • 监管申报;

  • 医学事务内容;

  • 药物安全监测;

  • 商业分析;

  • 医生沟通;

  • 患者支持项目。

市场空间非常诱人。

但漂亮的数字,并不能自动解决落地问题。


医疗行业仍在补合规这一课

风险和安全,仍然是最重要的障碍。

麦肯锡研究显示,43% 的受访者将风险与安全列为落地过程中的主要问题。

这些担忧包括:

  • 输出不准确;

  • 模型偏见;

  • 数据泄漏;

  • 未授权访问;

  • 监管不合规;

  • 推理过程无法追溯;

  • 责任边界不清晰。

一次错误的电影推荐,只会影响用户体验。

一次错误的医疗建议,却可能影响患者健康。

这种差异,决定了医疗智能体必须采用完全不同的部署标准。


我国正在建立更加系统的智能体治理框架

2026 年 5 月,国家网信办、国家发改委、工信部联合印发《智能体规范化应用与创新发展的实施意见》。

这是我国首部聚焦智能体的系统性政策文件。

文件强调以下原则:

  • 安全可控;

  • 规范有序;

  • 创新驱动;

  • 应用牵引。

涉及医疗影像诊断、手术辅助等高风险领域的智能体,可能需要接受备案管理、强制安全检测与临床验证。

这意味着行业正在从鼓励试验,逐步走向明确智能体获得真实业务权限的条件。


行业共识也正在形成

在中国互联网大会上,易联健康集团等单位联合发起“AI 赋能医疗健康可信应用倡议”。

倡议围绕四个方向建立共识:

  • 可信;

  • 安全;

  • 规范;

  • 可持续。

其中,“可持续”尤其重要。

一个医疗智能体在演示中表现良好,并不代表它已经真正落地。

它必须能够在真实预算、真实监管、真实医院系统和真实责任体系中持续运行。


合规是有成本的

合规并不是一句抽象口号。

它需要人员、系统、文档和持续投入。

相关成本大致可以分为三层。

第一层:基础合规成本

包括:

  • 制度建设;

  • 合规人员配置;

  • 员工培训;

  • 内部流程;

  • 审计机制;

  • 文档与报告。

第二层:技术合规成本

包括:

  • 安全设计;

  • 身份和访问管理;

  • 权限控制;

  • 数据加密;

  • 日志记录;

  • 模型监控;

  • 版本管理;

  • 故障恢复。

第三层:医疗场景特定成本

包括:

  • 医疗器械注册;

  • 临床验证;

  • 医疗数据治理;

  • 算法备案;

  • 证据链追溯;

  • 医生人工监督。

三甲医院或许能够消化这些成本。

但对中小医院、基层医疗机构、区域医疗服务商和独立开发者而言,这构成了很高的准入壁垒。

因为合规成本并不是单独存在的。

它还要叠加模型费用、云资源费用、数据管理成本与 Token 消耗。


每增加一个智能体,Token 账单都在增长

医疗智能体生态正在变得越来越复杂。

科大讯飞的四类诊疗智能体,每天可生成超过 10 万份门诊病历。

腾讯健康推出了 14 项医疗专属智能体能力。

医学影像推理框架可能在分析不同阶段调用多个专家模型。

一次普通诊疗,可能触发:

  • 语音识别;

  • 病历生成;

  • 质量审查;

  • 风险检测;

  • 指南检索;

  • 证据验证;

  • 诊断辅助;

  • 随访规划。

随着智能体数量增加,Token 消耗可能快速上升。

这种增长也并不总是简单线性的。

多个智能体可能重复读取同一份病历、交换中间结果、相互验证、调用不同专家模型,并对失败步骤进行重试。

这形成了一个非常现实的矛盾:

越需要通过 AI 改善服务的机构,反而越可能缺乏安全运行智能体所需要的基础设施预算。

这对以下参与者尤其明显:

  • 中小医院;

  • 基层医疗机构;

  • 独立医疗开发者;

  • 区域医疗平台;

  • 专业 AI 供应商。

下一个瓶颈不再只是模型智力。

而是如何获得价格可控、稳定可靠并符合监管要求的算力与 Token。


Unexhub 正卡在这一关键基础设施节点

Unexhub 正在尝试重构全球 Token 与 AI 算力的聚合、分发、定价和结算方式。

其核心定位,是建立面向 Token 聚合与分发的算力交易中心。

平台连接分散的 AI 供给,以及企业、开发者和智能应用产生的需求。

对医疗智能体生态而言,统一基础设施可以缓解一系列实际问题:

  • 模型供应商数量过多;

  • 定价体系不同;

  • API 格式割裂;

  • 回款周期过长;

  • 算力供应不稳定;

  • 模型切换成本高;

  • 中小参与者接入困难。

一家医院可能需要病历模型、影像模型、循证检索模型和高风险复核模型。

如果分别与每家供应商签合同、开账户、开发接口并处理账单,技术和运营成本都会快速上升。

统一聚合与路由层,可以显著降低这一负担。


快速结算,缓解中小供应商资金压力

现金流,是中小 AI 供应商最严峻的问题之一。

供应商立即提供算力,但收入可能要在数周甚至数月后才能到账。

在等待回款期间,它们仍然需要支付:

  • GPU 成本;

  • 电费;

  • 带宽;

  • 云端基础设施;

  • 运维成本;

  • 工程团队工资。

即使供应商技术能力很强,过长的账期也可能让业务难以持续。

Unexhub 的金融结算清算体系,强调更快速的调用收益分配。

其目标可以概括为:

算力被使用后,对应收益能够更快回到供应端。

对中小供应商而言,快速结算可以:

  • 缓解周转资金压力;

  • 提高现金流可预测性;

  • 更快投入新增算力;

  • 提升服务稳定性;

  • 降低扩张风险。

Unexhub 还计划推出面向早期供应商的激励机制。

符合条件的早期入驻方,在推广阶段可享受最低 0 佣金费率,从而在积累分发规模时,尽可能保留初期调用收入。

这项权益的价值并不只是“免费入驻”。

它还意味着:

  • 更低的客户获取成本;

  • 更快触达企业需求;

  • 更短的结算周期;

  • 更大的价格测试空间;

  • 更快回收算力投入。


API 竞价,让优惠成为市场机制

Unexhub 另一项重要机制,是 API 竞价模式。

传统模型 API 通常采用固定报价。

即使算力处于空闲状态,用户仍然要按照公开价格调用。

市场化机制则可以让价格根据真实供需变化。

在闲时,开发者可以用更低价格调用未被充分利用的资源。

在需求高峰,供应商则可以因提供稀缺资源获得更高收益。

对用户而言,可以获得:

  • 更低的闲时价格;

  • 更多模型选择;

  • 更灵活的任务调度;

  • 更低的单一供应商依赖;

  • 更可控的大规模 Token 开支。

对供应商而言,则可以获得:

  • 更高的闲置资源利用率;

  • 更广泛的市场需求;

  • 更合理的高峰期收益;

  • 更快速的结算;

  • 更低的渠道分发成本。

这不再只是一次促销打折。

而是尝试让部分 Token 定价权回归市场供需。


医疗 AI 需要动态资源路由

医院并不是每一个任务都需要最昂贵的模型。

医疗任务更适合的资源
语音转写快速低成本模型
病历起草均衡型语言模型
文书质量检查专业审核模型
医学影像分析医疗多模态模型
指南检索搜索与循证智能体
高风险诊断复核高端推理模型
患者随访低成本对话模型

如果所有请求都交给最昂贵的旗舰模型,规模化部署将很难持续。

更加合理的架构,应根据以下因素进行动态路由:

  • 临床风险;

  • 任务复杂度;

  • 延迟要求;

  • 数据类型;

  • 使用预算;

  • 算力可用性;

  • 合规要求。

此时,Token 交易市场不再只是提供低价 API。

它开始承担资源分配层的角色。


Unexhub 的模型聚合与毫秒级路由

Unexhub 计划聚合不同规模的模型与算力供应商。

平台生态可覆盖海内外主流模型,包括:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 其他通用和专业 AI 模型。

个人开发者、初创团队及中小供应商提供的优质模型和算力,也可以进入平台生态。

请求可以根据以下条件被动态分配:

  • 模型能力;

  • 价格;

  • 延迟;

  • 可用性;

  • 地区;

  • 任务类型;

  • 服务质量。

毫秒级路由对于医疗工作流尤其重要。

低风险行政查询,不应该占用与高风险临床复核相同的昂贵资源。

动态路由可以把高端算力保留给真正需要它的任务。


按实际 Token 使用量结算

按照实际 Token 消耗付费,还能进一步降低中小机构与开发者的使用门槛。

用户不必分别向多个模型供应商预购大量套餐,而是根据真实调用量结算。

这种方式可以带来:

  • 从更小预算开始;

  • 同时测试多款模型;

  • 根据需求逐步扩容;

  • 避免预付资源闲置;

  • 切换模型时无须重建应用;

  • 让费用与真实业务量更加匹配。

在部分供应和推广条件下,实际 API 价格最低可达到市场标准价格的约 30%。

对高调用量医疗智能体而言,这并不只是单次请求节省几分钱。

同样预算可能支持:

  • 更多患者咨询;

  • 更多病历生成;

  • 更多循证检索;

  • 更多独立复核;

  • 更长的智能体流程;

  • 更广泛的基层部署。

当然,医疗领域不能只看价格。

任何服务医疗场景的模型与 API 平台,都必须接受以下方面的评估:

  • 数据安全;

  • 隐私保护;

  • 模型可靠性;

  • 地区合规;

  • 审计能力;

  • 权限控制;

  • 服务连续性;

  • 故障恢复。

只有低价与安全稳定同时存在,价格优势才真正有意义。


PoloAPI 在实际多模型接入层的角色

Unexhub 解决的是更加底层的 Token 市场、算力分发、竞价与结算问题。

在具体模型接入层,PoloAPI 等多模型聚合平台则可以帮助开发者和企业,通过统一接口连接海内外主流模型。

PoloAPI 可接入:

  • Gemini;

  • Claude;

  • ChatGPT;

  • DeepSeek;

  • 其他国内外模型。

对医疗开发者和企业 AI 团队而言,这可以减少分别维护多套接口、账户和账单系统的负担。

其实际权益包括:

  • 统一 API 接入;

  • 集中查看使用量与费用;

  • 更方便地切换模型;

  • 降低多供应商维护成本;

  • 提供备用路由;

  • 支持企业级定制;

  • 满足高并发业务需求。

在部分资源方案中,实际调用价格最低可达到官方价格的约 30%。

对需要长期运行智能体的团队来说,同样预算可以覆盖更多用户、更多模型调用以及更多独立验证。

PoloAPI 还提供模型选型、接口集成、权限配置、并发规划和部署等企业级定制服务。

相关服务经验已覆盖大型公司、上市企业、国有企业和国际客户。

两者的角色可以形成互补:

  • Unexhub 关注 Token 聚合、分发、竞价和结算;

  • PoloAPI 提供更加直接的多模型接入与企业集成能力。

这类基础设施可以共同降低智能体部署的经济和技术门槛。


两场产业变革正在汇合

医疗智能体正在把诊疗工作改造成一条协同工作流。

病历生成、质量检查、诊断支持、循证检索和患者随访,都可能由专业智能体分别承担。

医生可以减少重复文书工作,把更多时间重新投入临床判断与患者沟通。

但这条自动化流水线要想持续运行,底层资源不能中断。

它需要:

  • 稳定的模型接入;

  • 足够的 Token 容量;

  • 动态路由;

  • 透明价格;

  • 快速结算;

  • 备用供应商;

  • 可靠基础设施。

医疗智能体企业正在回答:

智能体应该完成什么?

算力与模型基础设施平台则在回答:

支撑智能体的资源,应该如何被持续、低成本地提供?

两条线索交汇之处,才是真正的产业变革。


挑战依然存在

方向已经明确,但转型不会一蹴而就。

医疗机构仍然必须回答:

  • 现有系统能否承载多个智能体?

  • 数据能否安全地跨部门流动?

  • AI 建议能否被追溯?

  • 诊断证据能否独立验证?

  • 智能体犯错时由谁负责?

  • 基层机构能否负担所需 Token?

  • 高峰期供应商能否提供稳定算力?

医疗 AI 不能像普通办公聊天机器人一样部署。

它运行在高风险环境中,错误可能直接影响患者健康。

成熟的医疗智能体生态必须同时具备:

  • 医生人工监督;

  • 证据可追溯;

  • 权限控制;

  • 独立验证;

  • 模型监控;

  • 安全数据治理;

  • 稳定的算力和 Token 供给。

技术正在从单点智能走向全流程协作。

治理与基础设施也必须同步补课。


写在最后

约 99% 的医疗机构已经以某种形式接触智能体 AI。

但只有约 19% 真正进入规模化部署。

这道鸿沟并不只是技术问题。

它同时是信任、合规、系统对接、基础设施与成本问题。

医疗智能体正在重写诊室流水线。

它们能够生成病历、检查质量、支持诊断、检索证据、处理理赔并跟进患者。

但每一个智能体,都在消耗 Token、工具调用、数据权限与算力。

Unexhub 正在尝试重写这一生态底层的 Token 分发与结算逻辑。

PoloAPI 等聚合平台,则可以简化实际的多模型接入与企业部署。

一层定义医疗智能体能够做什么。

另一层决定模型和算力如何抵达真正需要它们的机构。

医疗 AI 的未来,不会只由模型智力决定。

它还取决于整个系统能否保持可信、合规、可负担,并且持续稳定地运行。

99% 的机构已经入局。

接下来的问题,是如何打好手中的牌。