What Hospital AI Cases Teach: 3 Decisions to Make Before Adopting Generative AI

Using hospital AI adoption cases, this article explains how much patient information to give AI, how to store records, and when people should review. It also walks through a hypothetical rollout at a Japanese-owned clinic in Makati under Philippine rules.

Author
AuthorAuthor

AI Engineer · 36+ years in IT · Japanese, based in Manila for 13+ years

What Hospital AI Cases Teach: 3 Decisions to Make Before Adopting Generative AI

Summary

  • In published hospital AI adoption cases, AI drafts documents and organizes patients' questionnaire answers, and doctors and medical staff review the content.
  • Before adopting AI, decide in this order: (1) which patient information to give the AI, (2) what to store and where, and (3) who reviews it and when.
  • In the Philippines, health information is classified as sensitive personal information. Even when processing is outsourced, the clinic remains responsible for how the information is handled.

A search for hospital AI adoption cases turns up examples where discharge summaries take less time to write, or where the wait from questionnaire to consultation got shorter. To bring a similar system into a clinic or a healthcare service, you first need to decide how much patient information the AI receives, where records are kept, and who gives the final review.

This article uses cases published by hospitals to explain the order in which to make these three decisions. The second half presents a case study of a hypothetical rollout at a Japanese-owned clinic in Makati, based on healthcare-related job postings on overseas freelance platforms. It is a fictional example, separate from PH AI Works' actual client work.

AI Adoption Cases Published by Hospitals

Nagoya Medical Center Drafts Discharge Summaries With Generative AI

NHO Nagoya Medical Center set up an environment where every clinical department can use generative AI to write discharge summaries, using a service from Fujitsu Japan. Full operation began on October 31, 2025.

A discharge summary tells the medical institutions and facilities providing post-discharge care about the patient's condition and course of treatment during the hospital stay. It must be completed in detail within two weeks of discharge. The center produces about 16,000 of them a year, and the burden on doctors was a problem.

In a trial in December 2024, the time to write one summary per patient fell from an average of 28 minutes to 8 minutes. The service is accessed over a closed network using a dedicated line. Clinical data is excluded from both generative AI training and cloud storage.

At Iseikai International General Hospital, Doctors Review and Finalize the Drafts

Iseikai International General Hospital introduced a generative AI service linked to its electronic medical records on September 26, 2025. It can draft discharge summaries, nursing summaries, and referral letters in about 30 seconds to 3 minutes per patient.

At discharge, the AI presents a draft, and a doctor reviews the content. After any needed corrections, the draft is finalized as the official record.

Urasoe General Hospital Shortens the Wait for Consultation With AI Questionnaires

In 2019, Urasoe General Hospital became the first hospital in Okinawa Prefecture to introduce "Ubie Medical Navi." Before that, patients filled out a paper questionnaire, then a nurse asked follow-up questions and entered the answers into the medical record.

Now patients answer questions on a smartphone or tablet, and staff transfer the answers into the medical record on a PC. The time from questionnaire to consultation fell from an average of 40 minutes to an average of 20 minutes. The dedicated questionnaire-nurse role was abolished, and the 2 nurses who held it moved to patient support and other work.

The Shared Division of Roles Between AI and People

Across the three cases, the same division of roles appears: AI drafts documents or summarizes questionnaire answers, and doctors and medical staff use that content in their work. In the Nagoya Medical Center case, the hospital also disclosed the data's network path and whether data is used for training or stored.

All three are hospitals adopting services built for medical institutions. When a small clinic or a healthcare app operator builds its own system, it has to decide the roles of AI and people, and how data is handled, one item at a time.

Costs and timelines in industries outside healthcare are covered in a related article. Generative AI for Philippine SMEs: Cut Costs, Save Time

Common Challenges in Healthcare AI Projects

Overseas freelance platforms carry job postings for AI adoption in healthcare. Clients include operators of apps such as nursing-record tools, chiropractic and dental clinics, and companies that provide data services to pharmacies and medical institutions. Requests also come from businesses that run patient-facing apps in the cloud.

The postings often list challenges like these:

  • Confirm before launch that AI features comply with patient-data protection rules (such as HIPAA in the United States)
  • Sign the necessary contracts with the providers of external services, such as AI, SMS, and email, that handle personal information
  • Decide how attachments are handled, and prepare for "prompt injection," where instructions hidden in input try to change how the AI behaves

On the technical side, the basics are encrypting data in the cloud, setting access permissions, and keeping audit logs. For large language models (LLMs), use services you can sign the necessary contracts with. On top of that, put in place records of consent and opt-outs, access permissions that differ by role, procedures for handing off to a staff member, and safeguards that keep the AI from giving professional advice.

The design depends on whether you can sign contracts with every provider whose service handles patient information, whether the work is new development or an audit of an existing app, and whether the AI's job is limited to administrative tasks such as scheduling.

Decide in This Order: Patient Information, Record Storage, Human Review

Once you decide what information goes to the AI, you can sort out what needs to be stored and where. Once you know what is stored, you can decide concretely what staff can review and at which step.

The overall adoption process, including choosing an owner and setting a trial period, is explained in a related article. An AI Adoption Roadmap for Philippine SMEs

(1) Decide the Scope of Patient Information Given to the AI

I start with administrative tasks that can be handled without patient information, such as explaining how to book an appointment. I add features that handle patient information only after confirming the contract and settings of the service being used.

Deciding the scope of features up front has helped me outside healthcare too. When I built a tool that reads prediction-market data, I designed it so it could only read data. Making clear from the start what the tool is trusted to do prevents later misunderstandings, such as whether it can place orders.

When handling patient information in the Philippines, the provisions of the Data Privacy Act (RA 10173) apply.

  • Health information, including past records, is classified as sensitive personal information.
  • Processing this information is prohibited in principle. One exception is when doctors or medical institutions process it for medical purposes with adequate safeguards in place.
  • The PIC (personal information controller) that manages personal information remains responsible for information whose processing is outsourced to a third party.

Advisory No. 2024-04, issued by the Philippines' National Privacy Commission (NPC), applies when personal data is processed in developing or operating AI. It requires that the PIC remain responsible even when processing is outsourced, and that personal data unlikely to help improve the AI be excluded by default.

When passing patient information to an external AI service, check whether you can limit the fields sent to what is necessary, and sign the necessary contracts with every provider whose service handles the information. Where a legal judgment is needed, consult a local lawyer.

(2) Decide What to Store and Where

The records to consider are the inputs and outputs kept by the external AI service and the audit logs and consent records managed by the clinic.

Whether an external service keeps records depends on the service and its settings. Amazon Bedrock, a cloud AI service, handles it as follows.

  • AWS protects the service infrastructure, and users manage the data they handle. This division is called the shared responsibility model.
  • Model providers cannot access Bedrock logs or users' prompts and outputs.
  • Whether inputs and outputs are stored can be set per region, at the account or project level.
  • If you choose none, neither inputs nor outputs are stored on the AWS side, and they are not shared with model providers.
  • With the none setting, calling a model that requires storage returns an error.
  • With aws_review, inputs and outputs are stored so that AWS staff can review them. Some model providers make this setting a condition of use.

When choosing a model, check whether the model you want supports the setting that does not store inputs and outputs.

AWS publishes a list of services that can be used for work covered by HIPAA in the United States. Besides Amazon Bedrock, it includes AWS CloudTrail, which records operations, and AWS KMS, which manages encryption keys. Before handling patient information on these services, the user must configure them to meet HIPAA requirements and sign a BAA (business associate agreement).

On the clinic side, store audit logs and records of patient consent and opt-outs, and set access permissions by role. In my own development work, I split the roles: data and sources are stored in a database, calculations are done in Python, and AI writes the text. That way, the basis for every number can be traced later.

AI service API keys also need managing. Keep them separate from the source code in environment variables or a .env file, and exclude .env from Git with .gitignore.

(3) Decide Who Reviews at Which Step

The NPC guidelines call for a system in which staff with the necessary knowledge and authority can intervene in the AI's processing when needed. In the Nagoya Medical Center and Iseikai International General Hospital cases, doctors also review the AI's drafts and finalize the records.

When I used AI to log my meals, it once counted 2 eggs as 2 grams. I noticed the error only some time later, when I recalculated in Python. Since then, I use a program for calculations. In the tool that reads prediction-market data, I also split the work: AI finds the market that matches the question, and a program calculates the probabilities.

For inquiries, I divide the roles like this:

  1. An AI chatbot receives the inquiry.
  2. Inquiries it cannot handle on the spot are routed by the AI according to their content.
  3. For consultations outside our field, we explain that we cannot help or point the person to another resource.
  4. A staff member reviews the reply and sends it.

For medical inquiries, the AI avoids giving professional advice and hands off to a staff member when needed. Handling of attachments and defenses against prompt injection are checked as well.

Case Study: A Japanese-Owned Clinic in Makati

Suppose a Japanese-owned clinic in Makati introduces a system that uses AI to explain how to book an appointment and passes medical questions to the reception staff. This example is designed around the Philippine Data Privacy Act, based on challenges seen on overseas freelance platforms.

What to decideStage 1 (start with administrative tasks)Stage 2 (handle patient information)
Patient informationLimit what the AI receives to general information needed for booking guidance and similar tasksConfirm contracts and settings, then give the AI only the fields the work requires
Record storageConfirm the setting that keeps the external AI service from storing inputs and outputsStore audit logs and consent records, and set access permissions by role
Human reviewHand medical questions off to reception staffDoctors and staff review and finalize the AI's drafts

In Stage 1, the AI handles guidance that can be answered without patient information, such as reception hours and how to book. Medical questions are directed to a desk staffed by reception. With general guidance only, staff can get used to the system without leaving medical judgments to the AI.

Before moving to Stage 2, check whether you can sign the necessary contracts with every provider whose service handles patient information, such as AI, SMS, and email. At the same time, set up record-storage settings and audit logs, and confirm points that need a legal judgment with a local lawyer.

FAQ

Q: How much time did hospital AI adoption save?

At Nagoya Medical Center, the time to write a discharge summary per patient fell from an average of 28 minutes to 8 minutes. At Urasoe General Hospital, after AI questionnaires were introduced, the time from questionnaire to consultation went from an average of 40 minutes to an average of 20 minutes.

Q: Can patient information be handled with AI in the Philippines?

Health information is classified as sensitive personal information, and processing it is prohibited in principle. One exception is when doctors or medical institutions process it for medical purposes with adequate safeguards in place. The clinic remains responsible even when processing is outsourced, so consult a local lawyer before actually handling it.

Q: Is patient information sent to an external AI service used for training?

Whether it is used for training or stored depends on the service, the contract, and the settings. In the Nagoya Medical Center case, the system does not use clinical data for training and does not store it in the cloud. Before sending patient information, confirm how the service you use handles it.

Q: Can a small clinic get started?

You can start with tasks that can be handled without patient information, such as explaining reception hours and how to book. While staff get used to the system, you can confirm the contracts and settings needed to handle patient information.

Conclusion

Hospital AI adoption cases show more than time savings: they also show the scope of information given to AI, how records are handled, and when people review the output. When adopting AI in a clinic or healthcare app, decide what information the AI receives, sort out how it is stored, and then decide who reviews at which step.

In the Philippines, the clinic remains responsible for health information even when its processing is outsourced. Start with administrative tasks that do not need patient information, and widen the AI's scope only after confirming the necessary contracts and settings.

PH AI Works offers free consultations on AI adoption for Japanese companies and small and medium-sized businesses in the Philippines. You can consult us from the pre-adoption stage about which tasks to give to AI and how to handle record storage and staff review.

References and Sources

About the author

Author
Author

Founder / AI Engineer (36+ years in IT)

  • ●From Tokyo · based in Manila for 13+ years
  • ●36+ years in IT (development, SEO, AI)
  • ●Google Certified AI Professional
  • ●IBM Certified Generative AI Engineer
  • ●AI chatbots, RAG & AI agent development
  • ●Physical AI Engineer

A Japanese AI engineer with 36+ years in IT and 13+ years on the ground in the Philippines. I write from hands-on experience to help Japanese companies adopt AI that actually delivers results — chatbots, workflow automation, AI agents, and AI-driven marketing. Feel free to reach out in Japanese or English.

Free AI Consultation

Tell us your challenges and we'll propose the right AI adoption plan for your business.

Book a Free 30-Minute Consultation

Related Articles

AI Case Study

AI Model Vendors Now Sell Implementation Too

OpenAI and Anthropic now sell implementation support as well as AI models. What a Japanese-affiliated local subsidiary in the Philippines should decide in-house first, before choosing an AI partner.

8/18/2026

AI Case Study

Contracting for AI While Prices Keep Falling

Write a model number into a contract and it is obsolete in six months. Learn how to design AI contracts for switching instead, with capability-based specifications, price-revision terms, and migration clauses.

8/17/2026

Meta Cut 750,000 Under-16 Accounts in Australia
AI Case Study

Meta Cut 750,000 Under-16 Accounts in Australia

Meta removed access to over 750,000 under-16 accounts in Australia using AI-assisted age checks. What this shows about depending on one platform for reach, and what a Philippine company should review.

8/16/2026

When an AI Calls Your Shop to Ask About Stock
AI Case Study

When an AI Calls Your Shop to Ask About Stock

AI now phones shops on a buyer's behalf to check stock. Decide what your phone answers before it reaches you: opening hours, products carried, how stock is confirmed, and whether prices may be quoted.

8/15/2026

Handling News That Is Not Settled Yet
AI Case Study

Handling News That Is Not Settled Yet

A reported $6 billion acquisition talk is not a completed deal. How to act on industry news before it is confirmed, from sharing templates to keeping your AI vendor switchable.

8/14/2026

What a 69-Page AI Report Does Not Show
AI Case Study

What a 69-Page AI Report Does Not Show

OpenAI's 69-page report says frontier firms generate 8.3x as many output tokens, yet a table on page 35 shows no statistically significant link to revenue per employee, and what that means for your adoption decision.

8/13/2026