AI Transcription and Report Drafting for Disability Support Sessions
Assessment, support-planning and review sessions were recorded on paper and typed up later. We built a structured web app with automated transcription and LLM analysis that drafts the report for the practitioner to confirm.
- Service
- AI Business Applications
- Industry
- Healthcare & Social Care
- Region
- Japan
- Engagement
- Lab-type ODC (dedicated team)
Outcomes
- Paper double-handling eliminated
- AI drafts, practitioner confirms
- Consistent structured records
Stack
Web app · OpenAI · Speech-to-text · Python · Structured data model
Written byOriental Nippon Team
A disability support centre in Japan runs three kinds of structured session: initial assessment, individual support planning, and regular evaluation and support reviews. Each one produces a document that is both a care record and a compliance record.
Sessions were recorded on paper and typed up afterwards, so every session was effectively handled twice, by the same staff who were supposed to be providing support.
The problem
- Double handling. Paper first, then typing.
- Admin on the wrong people. The burden fell on practitioners.
- Lost detail. Writing up after the fact loses what was actually said.
- Re-entry errors. Report data was copied from notes, introducing variation.
| Before | After | |
|---|---|---|
| Session capture | Paper notes | Recorded and transcribed |
| Report writing | Typed from notes | Drafted from the transcript |
| Final record | Whatever was typed | Confirmed by the practitioner |
| Session types | 3, on paper forms | 3, as structured records |
What we built
A web application for all three session types
The structure of each session lives in the system, not on a paper form.
Recording and transcription
Integrated with OpenAI models for analysis, tuned to the terminology actually used in sessions.
Automatic report drafting
Analysed transcripts populate the structured record, and a practitioner reviews and confirms every entry.
From session to confirmed record:
Technology and the role it played
| Technology | Role in this project |
|---|---|
| Web application | Structured records for all three session types |
| Speech-to-text | Transcription of session recordings |
| OpenAI models | Analysis of the transcript into report fields |
| Python | Pipeline and integration logic |
| Structured data model | A defined target for everything the automation produces |
How we delivered it
We mapped each session type with the practitioners who run it, including what each report must contain and why, before building any automation.
- Map the three session types. Report requirements and compliance needs.
- Build the structured record. A defined target for everything the automation produces.
- Capture and transcription. Tested on real session audio.
- Analysis and drafting. Always a draft for practitioner confirmation, never a final record.
- Pilot. A small group of practitioners across a full review cycle.
01Session mapping
2–3 wks
Three session types, report requirements, compliance needs
02Structured record
3–4 wks
Data model, forms, permissions
03Capture and transcription
3–4 wks
Recording, transcription, terminology handling
04Analysis and population
3–4 wks
Draft report data, practitioner review step
05Pilot
3–4 wks
Small practitioner group, full review cycle
Risks and how we handled them
| Risk | How we handled it |
|---|---|
| Unchecked AI records in a care setting | Every draft is confirmed by a practitioner |
| Terminology transcribed wrongly | Tested on real session audio and terminology |
| Reports missing required content | Report requirements mapped with practitioners first |
Outcome
Paper double handling is gone. Administrative effort is lower and records are more consistent.
The decision we would defend most firmly is the review step. In a care setting, an automated record that nobody checked is not an improvement.
What the client keeps
- The application and its source
- Session and report data models
- Transcription and analysis configuration
- Practitioner documentation
Sources and further reading
- Speech to text guide, OpenAI
- 障害者福祉 (disability welfare policy), Ministry of Health, Labour and Welfare, Japan
- People + AI Guidebook, Google PAIR
Facing something similar? This is the kind of work our AI, Data Science & Automation team delivers every day. Talk to us about your stack.
