AI case notes that protect trust, judgement and client data

Introduction

A useful AI case-note workflow depends on careful recording choices, honest conversations, close review of multilingual errors and clear control of every copy of sensitive information.

Protect the conversation

There is a real administrative problem to solve. Open Government Products (OGP) says its Scribe prototype addresses a problem in which medical social workers could spend 30 minutes to two hours writing notes after a meeting. Scribe records a conversation, generates a multilingual transcript and turns it into a structured summary that can be exported as a draft case note.[1]

CNA reported a Care Corner trial in which senior social worker Sheron Chng completed a case note in about 30 minutes rather than 60 after an English, Mandarin and Singlish conversation. That is one practitioner’s reported experience, not independent evidence of a general effect.[2] NCSS has separately described Viriya Community Services’ use of IntelliBot for recording, transcription and analytical support. It says feedback helped refine the tool for Singlish and case conceptualisation.[3] This agency account is not a controlled evaluation.

These examples show why agencies are interested. They do not settle the harder question: what changes when a client sees a phone recording a conversation about debt, violence, health, immigration status or family conflict?

OGP’s own 2024 write-up identifies this concern directly. Participants in mock testing worried that patients might be less forthcoming when recorded. They thought Scribe might be inappropriate for especially sensitive situations, and some would use it only where rapport already existed.[1]

Protect the choice to continue without recording

A calm explanation should cover what will be recorded, why AI is involved, what outputs will be created, who can access them, how long each will be kept, and what happens if the client says no. The session should proceed without recording and without a penalty to service.

Situations that call for a non-recorded path

Give practitioners authority to choose a non-recorded path when:

• the client declines, hesitates or appears uncomfortable;
• the conversation turns unexpectedly sensitive;
• another person enters whose consent and role are unclear;
• the practitioner cannot explain the workflow in a language or form the client understands;
• recording could change what is safely disclosed; or
• the setting or device does not meet the agency’s approved controls.

A client should also be able to ask for recording to stop without having to justify the request. The practitioner can then continue the session and document it through the approved fallback process.

The SASW Code of Professional Ethics treats audio recordings as client records and says prior consent must be obtained. It also calls for accurate, timely documentation and secure storage and retention under organisational policies.[4] This guide is not legal advice.

Treat every output as a draft

A fluent summary can feel finished before it is correct. That is dangerous in case work because a small wording change can alter meaning: “occasionally” becomes “often”; a question becomes a statement; the speaker is misidentified; an allegation becomes an observed fact.

The practitioner remains responsible for the note. A useful workflow makes that responsibility visible:

1. Label generated text AI-generated draft: not reviewed or approved for filing.
2. Require review before it can be copied or submitted to the case-management system.
3. Separate the client’s words, the practitioner’s observations and professional assessment.
4. Check names, dates, amounts, medication, risk statements, actions and who said what against the audio, using the transcript only as a navigation aid.
5. Record who reviewed and approved the final note, plus when.
6. Make correction possible after filing, without silently overwriting the history.

OGP reported that mock-test summaries could vary between generations and were sometimes verbose or contained minor errors.[1] Design for review rather than assuming a good interface removes it.

Keep the review meaningful

Professional judgement cannot be delegated to a summariser. If a tool describes a client as “coherent”, “resistant” or “distressed”, ask what evidence supports that word and whether it is an observation, inference or model artefact. CNA reported that Scribe could make preliminary assessments in a Care Corner trial.[2] Such wording needs stronger review.

Review multilingual speech for consequential errors

Singapore conversations move between languages, varieties and shorthand. A speaker may use English for an agency process, Mandarin for family relationships, Malay for emphasis and Singlish particles to soften or qualify a statement. Code-switching is not noise. It carries meaning.

Do not ask only, “What percentage of words was transcribed correctly?” Test whether the system preserves facts and stance. Build an error set from approved, non-client material covering:

• names, addresses, dates, money and scheme names;
• negation such as “didn’t”, “never” and “not anymore”;
• pronouns and speaker attribution;
• kinship terms and who did what;
• uncertainty, hearsay and conditional language;
• risk-related phrases;
• local accents, Singlish and code-switching; and
• quiet speech, overlap and noisy rooms.

OGP says Scribe supports local languages and speech varieties; its earlier project page specifically listed English, Malay, Chinese and Singlish, while mock testers requested dialect support and reported speaker-identification problems.[1] NCSS says Viriya’s IntelliBot was refined to account for Singlish.[3] These are useful product and programme descriptions. They are not proof that every service context, dialect or microphone condition will be handled safely.

Govern the whole information trail

Recording creates more than a final note. There may be local and uploaded audio, a transcript, model input, drafts, logs, backups and a case-system export. Map every copy before a pilot.

PDPC’s social service sector guidance says organisations should notify purposes and obtain valid consent unless an exception applies. It requires reasonable security arrangements, says personal data should be reviewed regularly, and explains that the PDPA does not prescribe one fixed retention period. Data should cease to be retained, or be anonymised, when its original purpose is no longer served and retention is no longer needed for legal or business purposes.[5]

An agency engaging a data intermediary still has obligations for personal data processed on its behalf.[5] PDPC’s AI guidance recommends data minimisation, suitable controls and impact assessments where useful. The deploying organisation bears primary responsibility for meeting its PDPA obligations.[6]

Turn that into operational questions:

• Where are audio, transcript and drafts stored and processed?
• Does any vendor use them for model training, testing or product improvement?
• Which staff and vendor roles can access each copy?
• What is the retention reason and deletion trigger for each artefact?
• Do deletion requests propagate to caches, backups and subprocessors where applicable?
• Can the agency retrieve relevant personal data for an access or correction request?
• What happens after consent is withdrawn, and what records must still be kept for valid reasons?
• Can the agency investigate an error or breach without retaining everything indefinitely?

Do not promise a client an absolute right to immediate deletion. PDPC guidance sets retention around purpose and legal or business need, and access or correction can be subject to exceptions.[5] Get the agency’s data protection officer and relevant professional leads to approve the actual wording.

Measure what happens to practice

Minutes saved matter, but they are not a sufficient outcome. A pilot can reduce writing time while weakening candour or creating a queue of unreviewed drafts.

Measure a balanced set:

• documentation time, including consent, review and correction;
• proportion of sessions where recording is offered, declined, stopped or judged unsuitable;
• client understanding and comfort, gathered without pressure;
• consequential error rate by language mix and error type;
• corrections made before and after filing;
• time from session to approved note;
• practitioner attention, cognitive load and overtime;
• completeness and usefulness judged through structured peer review;
• near misses, inappropriate access and deletion failures; and
• whether freed time is actually returned to client work, supervision or recovery.

Set stopping rules before launch. Examples include repeated loss of negation, poor speaker attribution in high-risk material, staff bypassing review, unclear deletion, or evidence that clients are withholding information. A pilot is allowed to conclude that recording is unsuitable for some programmes.

A safer pilot pattern

Start with synthetic conversations, then staff role-play, and only later consider a small approved live pilot if governance, professional and consent requirements are met. Keep ordinary note-taking available.

Choose a narrow service context. Complete a data-flow map and impact assessment. Agree which sessions must never be recorded. Train practitioners to explain the tool and recognise hesitant consent. Test consequential errors, not only average transcription quality. Require human approval and audit it. Set separate retention rules for audio, transcript, draft and final note. Review early results with practitioners, the data protection officer, supervisors and, where appropriate, client representatives.

The central design choice is simple: use AI to offer a draft, not to transfer accountability. Time-saving is worthwhile only if the conversation remains safe enough to be honest and the record remains the practitioner’s considered work.

Source note

Source note: Product and agency pages are first-party descriptions. Reported outcomes are attributed and not presented as independent proof.

About the author

Darren writes for Social Tech Guild about practical, responsible uses of technology in Singapore’s social service sector.