AI Agents Entering Healthcare: From Medical Documentation to Clinical Decision-Making — Racing Toward a Billion-Dollar Market

In the spring of 2026, the hype surrounding AI agents in the healthcare sector is rapidly intensifying.

From sustained capital infusion and intensive product iterations by enterprises to the urgent demand for cost reduction and efficiency improvement on the hospital side, a trend is becoming increasingly clear: medical AI is evolving from standalone tools toward a collaborative phase centered on intelligent agents.

Yet beneath the hype, the core question remains unresolved: Is the current level of technical maturity of AI Agents sufficient to support large-scale clinical implementation? As large language models begin to equip themselves with “claws,” where lie the boundaries of intelligence in healthcare scenarios?

Recently, the 18th session of the “China Innovative Healthcare Asset Lounge” Trading Roundtable, co-hosted by VCBeat and Wei Jie Yao, focused on“Multi-Billion-Dollar Market: How Can AI Agents Break Through in New Directions for the Healthcare Sector?”In an expanded discussion, multiple frontline industry and capital participants offered their assessments.

"As an AI doctor assistant already deployed in over 100 top-tier tertiary hospitals,"Xue Chong, General Manager and Founder of Quanzhen MedicineHe has a deep understanding of the “last mile” in technology implementation. In his view, an excellent Agent is by no means as simple as building a workflow and connecting to a large language model.

Xue Chong pointed out that, due to data security concerns, hospitals generally require private deployment. However, general-purpose large language models (LLMs) consume substantial computational resources, significantly reducing inference efficiency after deployment and thereby compromising the user experience for physicians. To address this, Quanzhen’s solution is “model compression”—training specialized, small-scale models for specific tasks such as medical insurance coding and clinical documentation. This approach minimizes computational costs while maintaining accuracy. Achieving this requires the feedback loop and post-training of large volumes of high-quality real-world data, enabling the models to operate both rapidly and accurately within hospital environments.

Tairex, also spun out from the Institute for AI Industry Research (AIR) at Tsinghua University, faces another challenge with its “Agent Hospital”: how to ensure that the high-performing technologies showcased in academic papers do not underperform in real-world settings.

Qiao Yuchen, Head of Product at TairexSummary of the "Four Gaps" from R&D to Clinical Practice:Data Integration Gaps(Hospital data is highly fragmented, far from a laboratory environment);Clinical Reasoning Gaps(AI is more like “memorizing answers” than “making judgments”);Reality-Scenario Disconnect(The patient's expression is complex and does not respond according to standard input);Human-Machine Collaboration Gap(There is a discrepancy between AI logic and physicians' work habits).

To address these challenges, Agent Hospital was designed from the outset to cover the entire patient journey—pre-consultation, during consultation, and post-consultation. It integrates specialized medical knowledge into its training and gradually cultivates clinical reasoning by reviewing negative samples. Furthermore, the agent technology underlying Agent Hospital possesses sustainable evolutionary capabilities, allowing it to continuously receive feedback during real-world use and progressively adapt to physicians’ workflows and complex clinical scenarios. Currently, the system has been deployed in multiple partner hospitals, where continuous feedback-driven iterations are gradually bridging the gap between laboratory technologies and clinical needs.

As the technological foundation becomes increasingly solid, a deeper question emerges: What are the fundamental differences between this wave of AI Agents and the previous generation of medical AI?

Li Yingjie, Executive Director of Inno Angel FundIt clearly delineates the essential differences between the two waves. The previous generation of medical AI resembled point solutions; taking AI-based medical imaging as an example, it performed only the single task of image interpretation and result output, addressing merely a small link in the diagnostic and treatment chain, thereby offering limited improvement to overall experience and efficiency.

This generation of AI agents has acquired the capabilities for proactive interaction, long-horizon reasoning, and execution of complex tasks, enabling them to participate in the entire healthcare journey—from preliminary consultation to post-diagnosis follow-up—much like a “junior physician.” This leap in capability has also driven a qualitative transformation in business models, shifting from the monetization challenges that plagued previous-generation AI to viable models such as subscription-based services and performance-based pricing. Consequently, the market size has expanded from the tens of billions to the hundreds of billions.

Regarding this view, the host,Yan Jingjing, Founding Partner of Probes CapitalShe also expressed agreement. In her view, as technology gradually achieves a form of “democratization” at certain levels, the core of project evaluation returns to who can more accurately grasp the pain points of payers and who possesses stronger product engineering capabilities. She further pointed out that the previous wave of AI largely remained at the level of NLP parsing, whereas the current wave features genuine reasoning and complex task-processing capabilities, enabling it to address problems and cover scenarios that are incomparable to those of the past.

Commercialization is a question that all AI healthcare companies must answer. Quan Zhen Medicine and Tairex have each presented their own paths of exploration.

● A B2B Logic of Pay-for-Performance, Starting with “Medical Record Documentation”

Xue Chong chose to address “medical record documentation,” the most painful pain point for physicians, a decision underpinned by deep-seated logic. As a former physician, he is acutely aware that administrative paperwork is one of the primary drivers of physician burnout. Meanwhile, medical records, which capture the complete diagnostic and treatment process, represent a “gold mine” for training more advanced medical AI systems.

Based on this assessment, he believes that future business models will shift from traditional software sales to performance-based pricing—where value metrics such as reduced insurance claim deductions, lowered medical risks, or freed-up physician time can be quantified and converted into fees hospitals are willing to pay. However, practical constraints remain: B2B operations cannot bypass hospitals’ tendering, project approval, and procurement processes, which often take six months or even longer, misaligning with the rapid iteration cycle of AI products. In his view, lighter-weight approaches, such as service-based or per-capita subscription models, may be necessary in the future to shorten implementation timelines.

● Comprehensive scenario coverage, leveraging "gap-filling" capabilities to establish essential B-side demand

Qiao Yuchen categorizes the paying customers of AI Agents into three groups: C-end patients, B-end hospitals, and D-end doctors. The B-end segment boasts strong payment capacity and high stickiness, but it is characterized by long decision-making chains and heavy customization requirements. Although the C-end market offers greater potential and more flexible business models, it suffers from low willingness to pay, and building trust in medical scenarios is particularly challenging.

In practical implementation, beyond conventional benefits such as “efficiency improvement,” the value of AI is also significantly reflected in its ability to “fill gaps.” Qiao Yuchen shared a typical case: when a pulmonologist was seeing a patient with chest tightness, the AI, through pre-consultation information, learned that the patient had previously undergone coronary stent placement and reminded the doctor to take this medical history into account when forming a diagnostic and treatment plan. The expert noted that specialists tend to be confined to their own domains in clinical practice, whereas AI effectively addresses this blind spot—this capability is precisely what makes AI an essential necessity in hospital settings.

● Path Selection from an Investment Perspective—Deepening B2B Operations Before Extending to B2C

From an investment perspective, Li Yingjie’s assessment is more straightforward. He believes that the core of healthcare remains hospitals as strong trust carriers; only when products are validated within in-hospital scenarios can they establish true competitive barriers. In contrast, To-C products face low entry thresholds, suffer from severe homogenization, and exhibit low usage frequency, making it difficult to build a stable business model.

However, this does not mean that there is no room for To-C models. In his view, a more realistic path is to first demonstrate value in To-B scenarios before extending to broader consumer markets—once a product has established trust and reputation within hospitals, expansion to the C-end will follow naturally. From the current perspective, this approach is being validated by an increasing number of companies.

As AI becomes involved in clinical diagnosis and treatment, the issue is no longer merely about “usability,” but rather two more serious concerns: whether data is compliant and whether results can be trusted.

To address issues of interpretability and “hallucinations,” Xue Chong proposed a practical distinction: quality control and early-warning scenarios require only results, whereas clinical diagnosis and treatment must be supported by evidence. To this end, Quanzhen Medicine has adopted a two-tier strategy: first, post-training the model on real-world clinical data to align it more closely with physicians’ decision-making logic; second, incorporating retrieval-augmented generation to simultaneously call upon the latest literature and knowledge bases when producing outputs. In essence, this shifts AI from merely “providing answers” to “presenting evidence.”

In terms of data security, Tairex chooses to mitigate risks at the source. Qiao Yuchen explained that during the training phase, Agent Hospital extensively uses “virtual patient” data generated based on professional medical knowledge, thereby minimizing reliance on real-world private data. In the practical application phase, strict adherence to legal and regulatory requirements is maintained in system construction, data isolation, and access control, ensuring that these issues are addressed through concrete engineering implementation.

Li Yingjie added from a due diligence perspective that investment institutions will focus on three key points: whether the data sources are compliant, whether there is a clear regulatory pathway for medical device registration, and whether the company itself possesses the necessary qualifications to conduct related business. “In the heavily regulated healthcare industry, compliance is a prerequisite for survival, not an option.”

When AI truly enters clinical practice, these issues will not be automatically resolved by technological advancements. They are not mere optimization targets, but matters of life and death.

As the discussion draws to a close, the conversation naturally turns to the future: How far will medical AI agents evolve by 2026?

Trend 1: Agents will be equipped with “Claws.”“Previously, AI had ears (ASR), eyes (OCR), and a mouth (TTS), but it lacked hands.” Xue Chong believes that with the maturation of technologies such as OpenClaw, AI agents will be able to truly control computers, automatically performing tasks like clicking and data entry in electronic medical record systems, thereby completely freeing doctors’ hands from the keyboard.

Trend 2: Multi-Agent Collaboration Will Become Mainstream.Qiao Yuchen predicts that medical AI will evolve from single-agent systems to multi-role collaboration, with multiple intelligent agents—such as doctors, patients, and nurses—participating jointly, potentially even enabling cross-departmental “multi-agent consultations.” However, he also cautions that open-source technologies like OpenClaw must be applied with care in healthcare settings, requiring rigorous safety assessments.

Trend 3: Bridging the Digital and Physical Worlds.“AI agents will move from purely virtual environments into the physical world,” Li Yingjie envisions. In the future, AI will not only handle online preliminary consultations but also integrate with fully automated equipment in clinical laboratories, enabling an automated workflow from “pre-consultation” to “pre-testing.” In broader healthcare-related fields such as drug discovery and biomanufacturing, AI will become increasingly intelligent and autonomous.

Trend 4: Shifting from Passive Response to Proactive Service.Yan Jingjing depicted a typical scenario: by the time patients return to the consultation room after completing their examinations, AI has already integrated all data and provided multiple diagnostic and treatment recommendations, leaving physicians to make only the final judgment. She believes that the advent of such proactive agents will mark the dawn of the next era.

From the “hands” of medical record documentation, to the “brain” of clinical decision support, and further to the “autonomous execution” that may eventually take over workflows, AI Agents are reshaping healthcare scenarios at a visibly rapid pace. Despite ongoing challenges in technological implementation, commercial monetization, and data compliance, a new healthcare ecosystem centered on “intelligent agents” is accelerating toward us.