All casesNo. 072
Nagano Municipal Hospital Puts 50+ AI Assistants on Its EHR
Assistants pull scoped chart data by patient ID and draft; staff check and finish
- Case no.
- 072
- Primary source
- Hospital press release (PR Times) + Unirita press release and case study
- Source published
- 2026-03-09

1. Why this case
Nagano Municipal Hospital is a 400-bed regional core hospital for advanced acute care in the northern part of Nagano Prefecture, Japan, run by a local incorporated administrative agency. On 2026-03-09 it announced that the generative AI assistants it runs inside the hospital produced work efficiency equivalent to 5,472 hours a year. That is about 456 hours a month, which the hospital equates to about three full-time staff.
The case is useful for four reasons.
First, the AI is wired into the electronic health record (EHR). Staff do not paste patient notes into a chat window. One click on the EHR screen opens the AI tool with the patient’s data already set, and an extraction tool the hospital has used since 2011 pulls only the records the task needs, such as the current admission.
Second, the split of work is stated plainly. The AI gathers information and writes drafts; a person checks the draft and finishes it.
Third, the rollout did not take off on its own. When the platform went live in April 2025, only some doctors used it. Usage across the hospital started to climb in September 2025, after an assistant cut a weekly nursing-team task from about 30 minutes to 3 minutes per patient.
Fourth, the hospital published how it computed the 5,472 hours and broke the figure down by assistant. The breakdown shows where the hours come from: the largest single item is gathering nursing information, and the doctors’ letter-writing assistants sit in the middle of the list.
The sources are the hospital’s press release on PR Times, a press release from the technology partner Unirita, and Unirita’s case study page for its SecuAiGent product, all in Japanese. All figures are the hospital’s or the vendor’s own.
2. Problem
The hospital’s release frames the work against a worsening staff shortage in medical settings. Its stated aim is to give clinicians more time with patients.
The vendor’s material describes where the time went.
- Finding information in the record. Staff spent time searching the EHR’s clinical and nursing records for what they needed and confirming it. Nurses took time to take in the content of nursing records. Doctors took time to pull the needed information from clinical records and to write medical documents from it. Unirita adds that this use of information depended partly on each person’s experience and skill.
- Paperwork that could not be templated away. As an acute care hospital, it produces large volumes of documents from admission to discharge, including discharge summaries and referral letters to other hospitals. The documents follow a set form, but every patient’s course is different. According to the case study page, it was not unusual for a doctor to spend more than 10 minutes per patient on a discharge summary, and this work was often pushed to the evening after clinical duties, which made it a cause of long hours. The page ties this to the reform of doctors’ working hours in Japan and to a tougher operating environment for hospitals.
- Two walls in front of generative AI. Kusano, the senior deputy director who leads the hospital’s DX team, first tried ChatGPT and other general tools on his own and found them useful for summarizing chart information. Putting them into practice ran into two walls, per the case study page. The first was security: handing patients’ sensitive personal information to a cloud AI as it was could not be accepted. The second was practicality: unless the tool was easy for everyone to use in a busy ward, it would not spread.
Confirmed facts: 400 beds, regional core hospital for advanced acute care in northern Nagano; staff shortage named as background; time spent finding and confirming information in clinical and nursing records; discharge summaries and referral letters as the main paperwork; 10+ minutes per patient for a discharge summary was not unusual; work often done in the evening; barriers named = security of sensitive personal data and ease of use.
3. Solution
3.1 An extension of a data pipeline the hospital already had
Unirita had supported the hospital’s data platform since 2011. That platform used Unirita’s no-code ETL (extract, transform, load) tool, Waha! Transformer, and the case study page says large amounts of data extracted from the EHR were already stored in usable form. In one of the two companies’ regular meetings, Unirita proposed extending Waha! Transformer to extract EHR data securely and pass it to generative AI.
The hospital’s system manager in the information systems office, who holds a doctorate in health information management, explains in the case study that he saw this as the breakthrough: rather than using general-purpose AI as it was, the hospital would have Unirita, which already knew its data structure, build a safe gateway. The hospital chose to extend its existing IT assets with a long-time partner rather than buy a packaged product.
The hospital’s release gives the timeline. In June 2023 it built a generative AI engine on a private cloud and set up a secure link to EHR data. In a joint trial with the data platform vendor, it succeeded in linking EHR data directly to generative AI in August 2023. The vendor’s product from this work is SecuAiGent, a generative AI cloud service, which the case study page says became available as a platform at the hospital in April 2025.
3.2 How an assistant gets its data
The case study page describes three design choices.
- Patient-scoped extraction. Based on the patient ID, Waha! Transformer extracts the exact data needed from the EHR, such as only the current admission, and hands it to the AI. The page describes this as retrieval-augmented generation (RAG) combined with ETL and claims it structurally prevents the AI from answering without a factual basis. That is the vendor’s claim, not an independent evaluation.
- Data protection. Per the vendor, personal information entered is not used to train the AI and no logs are kept on the cloud side. The Unirita release says the product places its own learning module in a local or private cloud environment. The hospital’s release describes the setup as one that lets staff use generative AI in real time and at high speed while taking care over the safety of clinical information.
- One click from the chart. To remove re-entering patient IDs and copying and pasting, SecuAiGent opens with one click from the EHR screen and the needed information is set automatically. Kusano explains in the case study that for clinicians a few seconds of extra effort is a barrier, so a tool that worked entirely inside the EHR was an absolute condition for it to spread.
3.3 More than 50 assistants, each for one job
Following the needs of doctors, nurses, other medical staff, and administrative staff, the hospital built more than 50 generative AI assistants, each for a specific use. The hospital’s release lists the main uses:
- drafts of referral letters (medical information provision documents) and discharge summaries
- an immediate one-pass summary of past records when an emergency patient is being accepted
- broad information gathering that nurses need at the start of a shift and that care teams need for their work
- transcription of patient explanations and case conferences from audio, and writing the records from them
The case study page adds care-information gathering for chemotherapy patients, records of informed-consent discussions, and help for pharmacists writing patient-education leaflets. For discharge summaries, the page says the AI summarizes the past five years of clinical records in one pass.
The assistants were built by the information systems office and the hospital’s DX promotion team, Team DIGITAL 2.0. The team began in February 2023 as Team DIGITAL and was later renamed. Kusano leads it, the system manager acts as its accelerator, and Sumi, a hematologist, acts as its evangelist. According to the hospital’s release, the team developed more than 50 assistants over about two and a half years from the August 2023 link.
3.4 Who does what
Kusano describes the working rule in the case study: the AI is not trusted with everything; a person checks the draft the AI produced and finishes it. He says that because this way of working took hold, the hospital gained large efficiency while keeping accuracy.
3.5 Getting people to use it
When the SecuAiGent platform was ready in April 2025, use started with discharge summaries. At first, only some doctors used it. Sumi explains in the case study that there is a real gap in IT literacy in medical settings, and that at first people did not understand what the tool could do or how it would make their work easier.
Team DIGITAL 2.0 responded with what the page calls grassroots activity built on concrete successes rather than tool explanations:
- sessions where staff could try generative AI hands-on, to lower the psychological barrier
- several sessions to share ideas for using it in clinical work
- a call for “digital ambassadors” across the hospital, staff interested in DX who joined Team DIGITAL 2.0 as partners. The team’s membership grew from about 10 at first to about 50 now.
The case study dates the turning point to September 2025. Once a week, a care team gathers the progress of its target patients from the records for urinary care. That took about 30 minutes per patient. With a dedicated prompt in SecuAiGent, it took 3 minutes. Per the page, this convinced the nurses that AI could make their own work easier, and the usage rate across the hospital began to rise quickly. No usage figures are given.
Confirmed facts: Waha! Transformer data platform since 2011; generative AI engine on a private cloud June 2023; direct EHR link August 2023; SecuAiGent platform ready April 2025, first used for discharge summaries; patient-ID-based extraction, e.g. current admission only; one-click launch from the EHR screen; 50+ assistants built by the information systems office and Team DIGITAL 2.0 over about 2.5 years; AI drafts, a person checks and finishes; team membership grew from about 10 to about 50 after the ambassador call; turning point September 2025 with urinary-care information gathering (about 30 min → 3 min per patient, weekly task).
4. Impact
4.1 The headline figure and how it was computed
| Item | Figure | Basis |
|---|---|---|
| Work efficiency | 5,472 hours a year | Hospital release, 2026-03-09 |
| Monthly average | about 456 hours | Same |
| Full-time equivalent | about 3 full-time staff | Hospital’s own conversion; basis not stated |
| Who | 111 staff who actively use the assistants | Surveyed group |
| Period | usage from April 2025 to January 2026 | Same |
The method, per the hospital: the time saved per run was calculated from a survey comparing before and after introduction, then multiplied by the number of times each assistant was run. Assistants in use for less than 12 months were adjusted based on their actual results. The 5,472 hours are therefore an annualized estimate built from self-reported time savings, not a measured total for the ten months.
4.2 Where the hours come from
The release includes a chart of annual hours by assistant, with each assistant’s time reduction rate plotted on a second axis. The chart names 23 assistants, fewer than the 50-plus built. It prints no values, so only the order can be read:
- The largest is a nursing-information batch-collection assistant.
- Next come two transcription-and-summary assistants (meeting minutes and nursing conferences), then a batch summary of a patient’s visit history.
- The doctor-document assistants sit in the middle: the medical information provision document is ninth and the doctor’s discharge summary eleventh. A prompt-making assistant appears twelfth.
- The smallest is information gathering for pre-operative nursing rounds.
The plotted reduction rate varies widely from assistant to assistant and does not follow the order by hours.
4.3 Two task-level figures from the vendor
Unirita’s release gives two before-and-after figures:
- discharge summary draft: an average of 10 minutes to 25 seconds per person
- information gathering for team care: about 30 minutes to about 3 minutes per patient
The 25 seconds is the time to produce a draft, which a doctor then checks and finishes. The hospital’s chart plots its own reduction rate for the discharge-summary and urinary-care assistants, and those rates are visibly lower than these pairs would suggest.
4.4 Quality, as reported by staff
The hospital’s survey asked about efficiency and the quality of care. Shares of respondents:
- both efficiency and quality of care improved: 45.6%
- quality of care improved: 3.8%
- efficiency improved: 48.1%
- no change: 2.5%
The hospital summarizes this as about 98% of respondents feeling an improvement in efficiency or quality. These are perceptions, not measured clinical outcomes.
4.5 Overtime and the target
Per Unirita, inpatient numbers rose 5% year on year while hospital-wide overtime hours were on a downward trend, and Unirita says the AI contributed to the reduction. No overtime figure is published. The case study page also says the hospital set a target of creating 7,800 hours of work time a year at the start of the fiscal year and is moving toward it.
Kusano, quoted in the hospital’s release, regards the direct link between the EHR and generative AI as one of the more advanced efforts in Japan, and calls the 5,472 hours the result of getting back time with patients that clinicians should be spending.
5. Insight
The pipeline came before the AI. The hospital had been extracting EHR data with the same ETL tool since 2011. Generative AI was added on top of that pipeline as a new consumer of data already in usable form. That is what made patient-scoped extraction and one-click use possible. A hospital whose EHR data can only be reached by copying and pasting would start from the security and usability walls this hospital had already worked around.
The hours came from reading as much as from writing. The project began with discharge summaries, the obvious doctor paperwork. The largest single item in the breakdown is gathering nursing information, and transcription assistants rank next. Hours are minutes saved per run times the number of runs, so a task that many nurses repeat every shift or every week can outweigh a task that takes longer but runs less often.
Adoption came from a nursing task, not the flagship. Discharge summaries went live first and stayed with some doctors. What changed the hospital’s usage was a weekly care-team task cut from about 30 minutes to 3 minutes per patient, plus an ambassador drive that grew the team from about 10 members to about 50.
The work split is draft and sign-off. The AI collects records and writes first drafts. The clinician checks and finishes them. The hospital credits that rule for keeping accuracy while saving time.
Limits. The 5,472 hours are an annualized estimate from self-reported time savings multiplied by run counts. They come from 111 active users, not all staff, and the breakdown names 23 assistants, fewer than the 50-plus built. The basis for “about three full-time staff” is not stated. Quality results are staff perceptions. The overtime claim has no figure. The vendor’s before-and-after pairs do not match the hospital’s own per-assistant reduction rates, and the claim that the extraction design prevents hallucination is the vendor’s. The hospital and the vendor both present the project as a model, and the vendor now sells the product built from it.
One-line takeaway: A hospital got generative AI into daily clinical work by attaching it to a patient-scoped data pipeline it already had, launching it from inside the chart, and letting nurses’ repeated information-gathering tasks, not only doctors’ documents, carry the adoption.
6. Implementation hypotheses (not in the source)
6.1 Each assistant as a prompt plus an extraction recipe
Basis: The case study page says Waha! Transformer extracts data from the EHR based on the patient ID, for example only the current admission, and that the discharge-summary assistant summarizes the past five years of records. Illustrative screens on the page show a short number entered as the message and replies laid out in fixed sections, such as a history summary capped at 500 characters followed by an 18-item urinary-care checklist, with a note that the recipient can be changed with ”@”.
The likely shape is that each of the 50-plus assistants pairs two things: a fixed prompt with an output template, and an extraction definition that tells Waha! Transformer which tables and which time window to pull for a given patient ID. The discharge-summary assistant would pull the past five years of records; the urinary-care assistant would pull the past week’s nursing and fluid records. The one-click launch would pass the patient ID from the open chart, and the ”@” selector suggests staff switch between assistants in the same chat window. If adding an assistant mostly means writing a new prompt and a new extraction query on a pipeline that already existed, that would explain how a small team built more than 50 in about two and a half years, and why a prompt-making assistant appears among them.
6.2 A possible check pass, and the final step left to people
Basis: Kusano says the AI’s draft is checked and finished by a person. In the illustrative urinary-care screen, the assistant’s reply opens by saying the content above has no problems, so it reposts the same output. The illustrative chemotherapy screen ends with a note that final judgments rest with the doctor and nurse.
The opening line in the sample reply may reflect a second pass, in which the assistant is asked to check its first draft against the extracted record and either correct it or repeat it unchanged, though on an illustrative screen it could also be leftover text from a demo conversation. The sources describe data flowing from the EHR into the assistant but not back. The likely workflow is that the clinician reads the draft, edits it, and enters the final text into the EHR document themselves, so the official record is always written by a person. That would keep the official record in a person’s hands and fits the rule the hospital describes.
6.3 How the hours were counted
Basis: The hospital multiplied survey-based time saved per run by each assistant’s run count and adjusted assistants used for less than 12 months based on actual results. The vendor says no logs are kept on the cloud side.
Run counts per assistant most likely came from usage records kept on the hospital side of the platform, counted per assistant rather than per patient, since the vendor rules out cloud-side logs. Minutes before and after would come from asking users of each assistant how long the task took before and how long it takes now. For assistants launched partway through the April 2025 to January 2026 window, the monthly run rate would be scaled to twelve months. Under that method, the 23 assistants in the chart are probably the ones with enough runs and survey answers to estimate, which would explain why the chart lists fewer than half of the assistants built.
7. Sources
- Nagano Municipal Hospital — press release on PR Times (Japanese): Nagano Municipal Hospital demonstrates 5,472 hours a year of work efficiency from generative AI (title translated)
https://prtimes.jp/main/html/rd/p/000000002.000179216.html
2026-03-09 (accessed 2026-11-17) - Unirita — press release (Japanese): Nagano Municipal Hospital creates 5,472 hours a year with Unirita’s secure generative AI SecuAiGent (title translated)
https://www.unirita.co.jp/press/2026/0416_press002.html
2026-04-16 (accessed 2026-11-17) - Unirita — SecuAiGent case study (Japanese): SecuAiGent case study: Nagano Municipal Hospital (title translated)
https://www.secuaigent.com/case-nagano
Publication date not shown on page; announced in the Unirita press release of 2026-04-16 (accessed 2026-11-17)