AI consultation produces results in 5 seconds, yet doctors are busier than ever

Author: 计算机魔术师 Tags: Frontend

AI-assisted diagnosis / drug discovery / medical insurance payment: the pace of implementation and typical obstacles.

These problems seem scattered, but behind them lies the same proposition: when accuracy is no longer the barrier, agent capability is the real moat for medical AI.

1. Where is the inflection point for medical AI moving from "usable" to "daring to use"?

Policy sets a clear timeline: the 2027 baseline target

In November 2025, the National Health Commission issued the "Implementation Opinions on Promoting and Regulating the Application and Development of 'Artificial Intelligence + Healthcare'" (National Health Office Planning Letter [2025] No. 30), providing a clear timeline: by 2027, establish a batch of high-quality datasets and trusted data spaces for the health sector, form a batch of clinical disease-specific vertical large models and agent applications, and widely apply intelligent assistance for primary care diagnosis, intelligent assistance for clinical specialty diagnosis and treatment decision-making, and intelligent patient visit services in medical and health institutions.

The core value of this document is not in painting a rosy picture, but in slicing the vague concept of "AI healthcare" into assessable milestones.

Got assigned

From an execution perspective, the wording of this document has a clear priority: primary care first, then specialties; assisted diagnosis first, then intelligent decision-making. This means that for vendors looking to enter the market, light consultations, chronic disease follow-ups, and report interpretation at primary care institutions are the policy-encouraged "first mile."

The market votes with real money: 3.5 billion in 2025, potentially exceeding 14 billion by 2030

IDC's research report published in April 2026 provides quantitative data: the AI + healthcare application software market reached 3.54 billion RMB in 2025, and is forecast to reach 14.0 billion RMB by 2030, with a penetration rate of 44.7%.

Behind this number are two signals: growth rate (a five-year compound annual growth rate of about 32%, higher than the average growth rate of the entire software industry) and structural change (the gap between general-purpose large models and medical large models is narrowing, and the moat of pure algorithm competition is becoming shallower).

Big shot nods in approval

More critical is the change in willingness to pay. IDC mentions that "some institutions have shown willingness to pay." From implementation cases, payment scenarios are mainly concentrated in three categories: imaging-assisted diagnosis, structured medical record generation, and departmental knowledge base retrieval. The common characteristics of these scenarios are: clear input-output boundaries, relatively high fault tolerance, and quantifiable verification.

Leading players collectively signal: medical large models enter the scale phase

A Xinhua News Agency report from March 2026 pointed out that AI has been widely applied in fields such as imaging, assisted diagnosis, and personalized treatment plan recommendation, "marking that the AI application in modern medicine has entered a phase of large-scale implementation." During the same period, Tencent Health announced its medical large model product matrix, GE Healthcare exhibited over 50 innovative achievements at CMEF, and NVIDIA released the BioNeMo platform.

The collective entry of leading players signals that medical large models have moved from "technical verification" to "product delivery." But productization and scaling are two different things.

Moving bricks on the construction site is the daily routine

IDC's research offers a more sober judgment: "The real next stop is not a more accurate model, but agent-driven software reconstruction." From experience, many teams are prone to pitfalls at this stage: thinking that connecting a large model to the HIS system counts as "implementation." In reality, HIS system integration is only infrastructure-level work; the real value lies in whether the agent can autonomously complete the closed loop of "consultation-triage-advice-follow-up."

%% title: Three thresholds for scaling medical AI flowchart TD A[Model accuracy exceeds 95%] --> B[Policy 2027 deadline] B --> C[Payer willingness to pay forms] C --> D[Agent reshapes long processes] D --> E[True inflection point for scaling] B -.Policy push.-> C C -.Commercial validation.-> D

Truth locked

Returning to the opening question: why does an AI consultation produce results in 5 seconds, yet doctors are busier? The reason is that most current implementation models are "bolt-on" — AI generates suggestions, and doctors review and modify them. In this model, AI does not reduce the total workload of doctors; it merely shifts part of the work upstream. Real efficiency gains require agents capable of handling the complete consultation-triage-advice-follow-up closed loop.

Source: IDC "The Next Stop for Medical Large Models: Agents Are Rewriting Hospital Software," April 2026; National Health Commission "Implementation Opinions on Promoting and Regulating the Application and Development of 'Artificial Intelligence + Healthcare'," National Health Office Planning Letter [2025] No. 30; Xinhua News Agency "AI Healthcare Enters Large-Scale Implementation Phase," March 2026.

These three scenarios are precisely the best lenses for observing medical AI's transition from "usable" to "daring to use." Their common feature is that technical bottlenecks are no longer the core problem; the real threshold lies in how to embed AI into existing clinical workflows.

Imaging assistance: not replacing doctors, but rewriting workflows

Take lung nodule screening as an example. After Xi'an North Hospital introduced AI assistance, the average reading time for radiologists dropped from 20 minutes to 12 minutes, but the overall efficiency gain did not come from AI "reading the film and automatically writing the report" — it came from AI handling the first round of preliminary screening, marking highly suspicious cases for priority review.

Agent runtime overload

%% title: AI imaging assistance workflow reconstruction flowchart LR subgraph Traditional workflow A[CT/MRI scan] --> B[Doctor reads film slice by slice] B --> C[Find suspicious lesions] C --> D[Issue report] end subgraph After AI reconstruction E[CT/MRI scan] --> F[AI preliminary screening and marking] F --> G{Suspiciousness judgment} G -->|Highly suspicious| H[Priority review] G -->|Low suspiciousness| I[Quick pass] H --> J[Doctor targeted confirmation] J --> K[Issue report] I --> K end B -.Heavy workload.-> E

Intensive care: 5-second medical record integration, the "external brain" for clinical decision-making

The application case in the intensive care unit of Peking University Shenzhen Hospital is more representative. Mindray's Qiyuan large model can complete retrospective review and integration of the entire diagnosis and treatment process data within 5 seconds, and generate a structured medical record in 1 minute. This is not simply "faster," but solves a fundamental pain point in clinical decision-making: information fragmentation.

ICU patient data is scattered across multiple systems: ventilator parameters, monitor waveforms, test results, nursing records, etc. AI plays the role of an "information integration layer" here.

Backend system design

%% title: ICU AI-assisted decision-making system architecture sequenceDiagram participant Doctor participant AI Agent participant Ventilator participant Monitor participant LIS participant HIS Doctor->>AI Agent: Request patient current status summary AI Agent->>Ventilator: Get ventilation parameters AI Agent->>Monitor: Get vital signs AI Agent->>LIS: Get latest test results AI Agent->>HIS: Get medication records Ventilator-->>AI Agent: Parameter data Monitor-->>AI Agent: Waveform data LIS-->>AI Agent: Test results HIS-->>AI Agent: Medication records AI Agent->>Doctor: Structured summary + trend analysis Doctor->>AI Agent: Follow-up on specific indicators AI Agent->>Doctor: In-depth explanation

Surgical robots: a new paradigm from assistive tool to precise operation

The application logic of surgical robots differs from imaging and ICU. It is no longer "assisted judgment," but directly intervenes in "assisted operation." Taking the "Tianji" orthopedic robot system used at Beijing Anzhen Hospital as an example, the system digitizes the surgical plan through preoperative planning, and executes the pre-planned path during surgery via a robotic arm, with error controlled at the sub-millimeter level.

Frontier curve

%% title: Surgical robot implementation decision framework flowchart TD subgraph Technical assessment A[Does precision meet clinical requirements] B[Is there sufficient case validation] C[Intraoperative contingency handling capability] end subgraph Engineering assessment D[Is equipment cost within budget] E[Is training period acceptable] F[Compatibility with existing equipment] end subgraph Compliance assessment G[Has NMPA certification been obtained] H[Is liability definition clear] I[Is data privacy compliant] end A --> J{Comprehensive assessment} B --> J C --> J D --> J E --> J F --> J G --> J H --> J I --> J J -->|Pass| K[Pilot deployment] J -->|Fail| L[Reassess or abandon]

The common conclusion across the three scenarios is clear: the real value of medical AI is not in "replacing people," but in "amplifying human professional value." Whether doctors become more relaxed depends on whether AI is truly embedded into their workflows, rather than adding a new operational step.

Workflow integration is the real difficulty

3. Agents, the real battlefield for medical AI

95% accuracy is just the entry ticket; agents are the watershed

Two research reports published by IDC in April 2026 give a clear conclusion: the accuracy of medical large models has improved from 80% to 95%, but this is only the entry ticket, not the moat.

The gap between general-purpose large models and medical large models is narrowing. IDC's actual measurements show that on tasks such as standardized clinical Q&A and medical record summary generation, the accuracy difference between leading general-purpose models and specialized medical models has shrunk to within 5 percentage points. But when task complexity rises — for example, requiring cross-modal integration of imaging, test reports, and medication records — the shortcomings of general-purpose models begin to show.

General-purpose model is assigned work it's not good at

An Agent is defined in the IDC report as: an independent AI unit capable of perceiving the environment, making decisions, calling tools, and executing operations. It possesses three core capabilities: memory storage (retaining patient context across sessions), tool invocation (interfacing with medical interfaces such as HL7/FHIR), and task planning (decomposing multi-step clinical decisions). These three capabilities are precisely what general-purpose models lack most when extending into medical scenarios.

%% title: General-purpose model vs. medical agent capability comparison flowchart LR subgraph General-purpose large model A[Information retrieval] B[Single-turn Q&A] C[Text generation] end subgraph Medical agent D[Multimodal integration] E[Tool invocation] F[Task planning] G[Continuous memory] end A -->|"Insufficient coverage" --> D B -->|"Lacks execution closure" --> E C -->|"Cannot handle multi-step" --> F D --> G E --> G F --> G

Agent capability confirmed by IDC as the watershed

This gap is not a performance problem, but an architectural problem. General-purpose models are passively responsive; medical agents need proactive collaboration.

IDC measurements reveal: the gap between general-purpose and medical models is narrowing

IDC's measurement data reveals trends on two levels.

The first level is accuracy convergence. On public medical datasets such as MIMIC-IV and CMeEE, the clinical text understanding capabilities of general-purpose models like GPT-4 and Claude have reached 85%-90% of the level of medical vertical models. This means that the ceiling for paths relying solely on model fine-tuning is lowering.

The second level is capability divergence. General-purpose models excel at information retrieval and single-turn Q&A, but are not good at multi-step reasoning and tool collaboration. Medical agents, on the other hand, need to complete the "perception-decision-execution" closed loop in real clinical environments.

Whoever can reshape long-process business with agents wins

One characteristic of medical scenarios is: long processes, high fault-tolerance requirements, multi-role collaboration. A complete diagnosis and treatment process involves multiple nodes: registration, consultation, examination, diagnosis, treatment, follow-up, etc.

The real value of agents lies in connecting these scattered nodes. Taking chronic disease management as an example, traditional AI can only provide assistance at a single point — such as image reading or medical record quality control. Agents can cover the entire process from outpatient prescription to examination appointment, from result return to prescription adjustment.

Tencent Miying's practice is a typical case. Its AI imaging platform has evolved from a single assistive diagnostic tool into a multi-module intelligent system integrating imaging reports, follow-up records, and clinical decision support.

Multi-system integration is the hardest part of engineering

But the core obstacle to agent implementation lies not in algorithms, but in two engineering problems: interface standards and trust mechanisms. Medical system interface standards (HL7, FHIR, DICOM) are severely fragmented, and data formats across different vendors' systems are not unified. More critically, the fault tolerance of medical decisions is extremely low; every operation executed by an agent needs to be traceable, auditable, and rollbackable.

The case of Peking University Shenzhen Hospital's ICU is not an isolated one. At Xi'an North Hospital, AI compressed aortic dissection imaging diagnosis from 15-20 minutes to 3 minutes; in lung nodule screening, AI helped radiologists reduce workload by 30%-50%. The efficiency gains are real. But the counter-intuitive phenomenon of "doctors being busier" reveals the core contradiction in scaling medical AI: tool-level efficiency gains do not equal workflow-level burden reduction.

Efficiency improved, but the workload didn't decrease

4. Three key questions for practitioners

Why hasn't your scenario been implemented yet?

Between technical usability and commercial sustainability lie three gates.

The first gate is data access cost. IDC's measurement report published in April 2026 points out that the accuracy gap between general-purpose large models and medical large models is narrowing — leading medical large models have broken through the 95% baseline. This means single-point accuracy is no longer the barrier. The real threshold is: can your scenario obtain structured, fully annotated, compliance-ready datasets?

The head of the IT department at a certain top-tier hospital admitted at the 2026 VB100 conference that when they procured an AI-assisted diagnosis system, the biggest cost was not the software license fee, but the 6-month HIS/PACS interface renovation and manual annotation cost. "We spent 3 million, of which 2.2 million was data governance."

The second gate is workflow embedding depth. If an AI consultation system only runs independently as a "bolt-on," doctors need to repeatedly switch interfaces, manually transcribe conclusions, and perform secondary verification — this is not efficiency improvement, but increased cognitive load. The truly effective embedding method is workflow reconstruction: AI intervenes as an "assistant" rather than a "decision-maker."

Tencent Miying's medical imaging platform adopts this layered architecture: AI completes lesion annotation and priority sorting, and doctors only need to review high-risk cases. This design shifts radiologists from "frame-by-frame reading" to "focused review," improving per capita efficiency by about 40%.

Interface renovation costs three times more than model training

%% title: Three gates of AI healthcare implementation flowchart TD subgraph Data layer A[Structured data acquisition] --> B[Compliance annotated dataset] B --> C[95%+ accuracy baseline] end subgraph Workflow layer D[Bolt-on invocation] -.Inefficient.-> E[Manual transcription + secondary verification] F[Embedded reconstruction] --> G[AI preliminary screening → Doctor review] G --> H[Per capita efficiency improved by 40%] end subgraph Organizational layer I[Process adjustment resistance] --> J[Incentive mechanism reconstruction] J --> K[Commercial sustainability] end C -.Just the entry ticket.-> D C --> F E -.No qualitative change.-> I H --> J

The third gate is incentive mechanism reconstruction. Are medical insurance payers willing to pay for AI-assisted diagnosis? Will hospital management include AI usage rates in performance evaluations? Deloitte's 2025 Global Healthcare Industry Outlook survey shows that nearly 70% of healthcare institution C-suite executives list "optimizing operational efficiency" as an annual priority. But equally important, over 60% of respondents indicate that budget constraints are the biggest obstacle.

During the policy dividend period, which sub-sectors are worth betting on?

Policy provides a clear timetable. According to the implementation opinions issued by the National Health Commission in November 2025, by 2027, intelligent assistance for primary care diagnosis, intelligent assistance for clinical specialty diagnosis and treatment decision-making, and intelligent patient visit services will be widely applied in medical and health institutions.

This means the next 18-24 months are a window period. From the perspective of implementation certainty, the following three directions have the highest priority:

Imaging-assisted diagnosis. This is currently the most commercially mature track. For highly standardized scenarios such as lung nodule screening, breast cancer screening, and fundus lesion detection, AI has completed the capability leap from "assisted screening" to "quantitative reporting." The competitive focus in this track has shifted from "accuracy" to "workflow integration capability" and "medical insurance payment integration."

Intensive care and emergency decision support. The case of Peking University Shenzhen Hospital demonstrates the value density of this type of scenario: 5 seconds to integrate the full medical record, 1 minute to generate structured documentation. This type of scenario is characterized by strong timeliness requirements and low fault tolerance.

Chronic disease management and patient follow-up. The commercialization path for this type of scenario is clearer: patient-side APP + doctor-side dashboard + AI follow-up agent. Google Cloud's 2026 industry trends report points out that the comprehensive analysis capability of multimodal LLMs combining patient medical records, imaging, and laboratory reports is reshaping the way clinical teams work.

Policy gives time, market gives space

But it is necessary to be clear-headed: although high-end scenarios such as drug discovery and surgical robots have greater technical imagination space, their commercialization cycles are longer, regulatory thresholds are higher, and single implementation costs are higher. For teams with limited resources, it is recommended to prioritize tracks with "high certainty + fast payback."

Where is the boundary from "technically usable" to "commercially sustainable"?

The scaling of medical AI is not a matter of making the model a bit more accurate, but a matter of who can make doctors willing to use it and daring to use it. This determines the division of three key boundaries.

Boundary 1: The legal red line between assistance and substitution. Under China's current regulatory framework, AI cannot serve as an independent diagnostic entity. All AI-assisted diagnosis systems must operate under the premise of "doctor-led," and AI output must be confirmed by a licensed physician. This means product design must reserve a "manual review" step.

Boundary 2: The engineering trade-off between data privacy and model generalization. The high sensitivity and dispersion of medical data cause technical routes such as "federated learning" and "privacy computing" to face dual pressures of performance and compliance during implementation. The MaaS model demonstrated by the Wenzhou Health Information Center provides a feasible path: aggregating city-wide medical institution data through a regional integration platform to create high-quality datasets. This "data usable but not visible" model is becoming the standard solution for regional-level AI medical projects.

Boundary 3: The mismatch between the speed of technological iteration and the speed of organizational change. The technical iteration cycle of large models has shortened to 3-6 months, but the informatization construction cycle of hospitals is typically 18-36 months. This mismatch requires product architecture to possess the capabilities of "model replaceability" and "process configurability."

Technology runs fast, organization can't keep up

%% title: The triple boundary of commercial sustainability stateDiagram-v2 [*] --> TechnicalVerification TechnicalVerification --> AssistivePositioning: Legal boundary confirmed AssistivePositioning --> DataCompliance: Privacy computing solution selected DataCompliance --> OrganizationalAdaptation: Training + process reconstruction OrganizationalAdaptation --> CommercialSustainability: ROI verified OrganizationalAdaptation --> TechnicalVerification: Model iteration too fast DataCompliance --> TechnicalVerification: Federated learning performance insufficient AssistivePositioning --> TechnicalVerification: Regulatory policy changes

The real moat of medical AI is shifting from "algorithm accuracy" to "agent capability." IDC's measurements reveal a key trend: the gap between general-purpose models and specialized medical models is narrowing; the future watershed lies in who can reshape long-process business with agents.

For practitioners, the current strategic question is not "can AI be used," but "why hasn't your scenario been implemented yet." The answer usually lies not at the model layer, but at the data layer, workflow layer, and organizational layer.

Frankly speaking, the golden window period for medical AI is approximately 18-24 months. Policy has given a clear timetable, the market is voting with real money, but players who can truly cross the chasm from "technically usable" to "commercially sustainable" are still scarce. The next stage of competition will not take place in laboratories, but in hospital IT departments, medical insurance bureau payment catalogs, and doctors' daily work.

References

[1] AI Healthcare Enters Large-Scale Implementation Phase. https://www.news.cn/tech/20260304/5abc83145dba4124b3ba85e90bff04d7/c.html [2] News Center_GE Healthcare | GE HealthCare (China). https://www.gehealthcare.com/zh-cn/china-news [3] Cloud Healthcare and Life Sciences (HCLS). https://cloud.google.com/solutions/healthcare-life-sciences?hl=zh-CN [4] Implementation Opinions on Promoting and Regulating the Application and Development of "Artificial Intelligence + Healthcare". https://www.nhc.gov.cn/guihuaxxs/c100133/[REDACTED]/d1a42ae835c743b9b3e83ac0253c3e9f.shtml [5] 2025 Global Healthcare Industry Outlook | Deloitte China. https://www.deloitte.com/cn/zh/Industries/life-sciences-health-care/perspectives/2025-global-health-care-executive-outlook.html [6] AI Purpose-Built for Healthcare and Life Sciences. https://www.nvidia.cn/industries/healthcare-life-sciences [7] How Large Models Help Improve the Quality and Efficiency of Medical and Health Services - Tsinghua SEM. https://www.sem.tsinghua.edu.cn/info/1173/37293.htm [8] 2026VB100 | The "Last Mile" of Medical AI Large Models: Only Real Implementation Has Value. https://bydrug.pharmcube.com/news/detail/9a530f1ae67f0613f533cc87a6747241