ChatGPT for Healthcare Update: Computer-Use Agents, Clinical Search, and Over 20 Plugins

ChatGPT for Healthcare Update: Computer-Use Agents, Clinical Search, and Over 20 Plugins

OpenAI updated ChatGPT for Healthcare on November 16, 2026, introducing computer-use agents, enhanced clinical search, and over 20 plugins. This article summarizes the work it can assist with, and the considerations regarding data security, permissions, and costs.

OpenAI updated ChatGPT for Healthcare on November 16, 2026, focusing on three key areas: enhancing "computer-use agents" to enable AI to perform multi-step tasks within existing software used by medical teams; improving "trusted clinical search" to find, assess, and cite medical evidence; and expanding the medical plugin ecosystem, with over 20 new plugins announced.

This news is worth noting for hospitals, clinics, and related management teams in Taiwan, but there's no rush to adopt it fully. It addresses common pain points in the medical field: fragmented systems for patient records, scheduling, insurance, and administrative tasks that require users to copy and paste information across different screens. The real question is: these capabilities are primarily designed according to U.S. regulations and software ecosystems—what tasks are truly appropriate to try, and which permissions and data should not be entrusted to AI agents?

What's New in This Update

The official explanation states that medical teams often need to piece together information from electronic health records (EHRs), internal protocols, scheduling systems, and insurance portals to complete care and administrative tasks. The direction of ChatGPT for Healthcare is to integrate the tools already utilized by the team, along with the medical evidence needed for decision-making, into a single work platform.

The term computer-use agents refers not to opening another chat window, but to allowing the AI to operate existing software: reading screens, clicking, entering data, and moving between applications. The official statement explains that this update features a model more suitable for clinical and administrative workflows that emphasizes safety and cost, and offers access control and audit logs for institutions to manage deployments. The demonstration scenario uses synthetic patient data: clinical staff in the ChatGPT application use voice commands to request the AI's help in tracking an MRI application; the AI reviews the patient's records, drafts medical orders and responses for the patient, which are then sent for review.

On the same day, OpenAI also explained the improvements in clinical search. The official statement mentions that the updated search will find, evaluate, and cite medical evidence, allowing clinical staff to match relevant evidence to actual problems and check sources. In a blind test, physicians reviewed answers to real clinical questions (including multiple conversation rounds); the official statement indicated that in over 1,300 reviews, the updated experience received higher marks in all eight rating dimensions, including factual accuracy, safety, usability, and source authority.

Regarding plugins, the official list includes new medical plugins covering clinical care, care coordination, and operations, and specifies that they can be used in ChatGPT for Healthcare, as well as accessed through ChartGPT within EHR workflows. Examples of tasks provided include: viewing patient records and internal protocols, drafting prescriptions for physician review, arranging patient contacts, checking insurance eligibility, and analyzing medical costs and operational performance. It should be noted that most of the plugins listed belong to the U.S. healthcare and insurance ecosystem, which does not imply that Taiwanese hospital information systems or health insurance processes are compatible.

What Tasks Are Most Helpful

Based on the official demonstrations and collaborations, tasks that may be more likely to see early assistance involve cross-system, repetitive tasks, and those that allow "drafting before human submission", rather than directly replacing clinical judgment.

Scenarios closer to practical application include: organizing referral information, transferring documents from one system to another, preparing pre-authorizations for insurance, tracking post-discharge follow-ups, and confirming appointments and eligibility before returns. The official statement also noted collaborations with organizations like Abridge, Ambience, Cigna Healthcare, and CVS Health to apply AI in referral processing, document transfer between systems, and compiling information needed by care teams. Another accompanying explanation uses the example of MRI applications: the AI first reviews medical records and existing images, drafts medical orders, medical records, and patient responses, and then prepares a pre-authorization application for the insurance portal based on the same data, with every step designed for human review.

The officials also released evaluation figures, which can help understand how they tested but should not be seen as a guarantee for Taiwan's site. They stated that in reviewing the electronic health record inbox tasks using GPT-6 Astra, 99% of 325 evaluated tasks were deemed clinically safe; out of 646 usability assessments, 93% were rated by physicians as "very helpful" or "extremely helpful". In terms of administrative processes, they cited the independent benchmark HealthAdminBench from Stanford researchers, pointing out that GPT-6 Astra received a score of 94.3% and GPT-6 Luna received 91.8%, with Luna's cost per task at 8.5 cents—98% lower than Astra when assessed under this benchmark and with maximum inference effort. These are the testing conditions and results released officially; actual performance will still vary based on system screens, data quality, and manual review processes.

For Taiwanese institutions, a more pragmatic starting point is usually not to aim for complete integration of all clinical systems at once, but rather to select a cross-system administrative process with manageable failure costs that requires manual approvals to test. For example, allowing the AI only to draft authorization requests or referral summaries, which must first be confirmed by a responsible person or physician before submission.

Considerations Before Implementation: Data Security, Permissions, and Costs

Regarding data security, the officials stated that the content shared with ChatGPT for Healthcare will not be used to train models; institutional data will also not be used to train OpenAI's base models. If protected health information is to be processed in accordance with HIPAA workflows, institutions must use qualifying services covered under their business associate agreement (BAA) with OpenAI, applying necessary security and retention settings, and extending protections to connected systems. These are the official conditions under U.S. regulations and not an automatic compliance guarantee for Taiwan. Taiwanese institutions still need to review the Personal Data Protection Act, medical personal data, and internal cybersecurity policies themselves to confirm how data is handled, where it is stored, what liability arises from incidents, and whether the contract truly covers the functionalities and systems connected for actual use. The official statements did not specify whether Taiwan has fully opened its access or the locations where data is stored.

On the permissions side, computer-use agents can read screens, click, and input data, which means they can potentially access everything they are allowed to. The officials recommend that institutions first identify cross-application workflows, set the applications the agents can access, and clearly define which steps must undergo human review. ChatGPT for Healthcare provides role-based access control and audit logs; the partnership with Ambience describes their practice: when medical record data is unclear, the AI will pause to ask clinical staff, getting consent before proceeding to submit. This kind of "draft, pause, and approve" design should be viewed as a basic requirement for implementation, not an optional feature. Clinical judgments, medical orders, and formal responses to patients should remain the responsibility of licensed medical personnel.

Regarding costs, the officials mentioned that this update also focuses on affordable models and provided the task costs mentioned above, but those are not the pricing in Taiwan. Actual expenditures will depend on workspace plans, API usage, whether plugins are charged separately, and the human resources required for manual reviews, system settings, and exception handling. Officials also clarified that ChatGPT for Healthcare is a governed workspace intended for use by clinical, research, and administrative teams; ChatGPT for Clinicians is free for qualified and validated clinicians in the U.S. for reviewing medical evidence and preparing clinical documents. Both can sign BAAs with OpenAI to support HIPAA-compliant workflows. Computer-use agents (in ChatGPT Work and Codex) and the improved clinical search are stated by the officials to be available immediately, but still subject to workspace settings; new medical plugins must be activated by workspace administrators. For purchasing, the officials noted that direct purchases are available, or through private proposals on AWS Marketplace. The eligibility for free plans, whether plugins can be used in Taiwan, and subsequent usage costs should all be confirmed with the officials or distribution channels; estimates based on U.S. scenarios may not be appropriate.

Therefore, when evaluating, it’s important not only to ask, "Is it powerful?" but also to address three very specific questions: Which systems is the AI allowed to access? Which actions must not be done without human intervention? Is there sufficient documentation for traceability in case of errors? If plugins or systems are not compatible with existing processes in Taiwan, even if the model performs well under official benchmarks, it doesn’t mean your hospital or clinic can immediately save administrative time.

If enterprises or medical institutions wish to assess whether such tools are suitable for their existing workflows, AI.com.tw can assist in inventorying implementation methods and considerations. Consultations can be arranged via LINE: @119m.

Original official text: