AI for social workers: case notes without client names, and which note tools sign a BAA

How can AI be used in social work?

AI is used in social work for paperwork: case notes, the structure of a court report, referral letters, resource lists, supervision prep and plain-language drafts for clients, with the client's identifiers kept out of the chat and put back by hand. In the University of Texas at Austin and NASW survey of 1,179 US social workers, fielded October 2025 to February 2026, "Nearly two-thirds of respondents reported using AI in their current role, most often for routine tasks like correspondence, reports and documentation."

The England numbers come from Social Work England's study, run by Research in Practice in March 2025 with 203 respondents, 155 of them practising social workers, and published in January 2026. Of the social workers, 40% had used AI with their employer's direction and 24% had used generative AI without it. That second row is who this page is written for: a worker with a laptop, an agency template, and no tool the employer has approved.

The one rule that makes the rest safe is NASW's. Its AI and Social Work page, summarising the Legal Defense Fund's "Navigating Artificial Intelligence & Client Privacy", says: "Social workers must vet vendors, minimize identifiers, and comply with HIPAA/NASW ethics." Minimizing identifiers is a habit, not a product. Nothing below asks you to type a client's name, date of birth, address, phone number, case number or school into any chat, whichever tool you pay for. What never to paste into an AI chat is the general list; the social work version is shorter and stricter, because the data is someone else's.

How do you write social work case notes with AI? De-identify, prompt, re-insert

You write case notes with AI in three moves: strip the identifiers, prompt the model with your rough notes and your agency's format, then put the identifiers back inside your own case system. The model never learns who the client is, and the note that goes on file is one you read line by line and signed.

1. De-identify on paper first. Replace every name with a role tag (CLIENT A, CHILD 1, PARTNER B), every place with a type (SCHOOL X, the flat), every date with a relative time (this week, three weeks from the visit). Drop phone numbers, case numbers, employers and any diagnosis the note does not need.
2. Prompt with the format your agency uses (DAP, SOAP, BIRP, or the agency's own headings) and tell the model to add nothing. Most assistants let you save instructions per project; on Whizi a project carries instructions of up to 32,000 characters and 10 pinned files, so the template and the de-identification rules ride on every note without retyping.
3. Read the draft against your rough notes, line by line, and delete anything you did not write. NASW's practice staff say AI-generated notes "are not always accurate" (Galvan, below); this is the step that catches it.
4. Re-insert the identifiers in your case management system, never in the chat.
5. Delete the chat, or keep it only in a tool whose retention you can state.

The prompt I use for a DAP note, with the identifiers already swapped: "You are helping a social worker write a case note from a home visit. The client is CLIENT A, the child is CHILD 1, the school is SCHOOL X. Do not add any names, addresses, phone numbers, dates or facts that are not in the notes. Write a DAP note (Data, Assessment, Plan) under 200 words from these rough notes: home visit 9:30, CLIENT A home, CHILD 1 at SCHOOL X. Flat tidy, food in fridge. CLIENT A sleeping badly since the eviction notice, court date in 3 weeks, has not called legal aid. Reports 3-4 beers most nights, down from last month. SCHOOL X attendance letter, 82 percent this term. CLIENT A agreed to call legal aid this week; I will text the number. Next visit in 2 weeks." The sentence that does the work is the third one. Without it, nothing in the prompt stops the Assessment from carrying things you never observed.

NASW's own practice staff put the check step plainly. In "NASW Tips & Tools for Social Workers: Artificial Intelligence" (NASW Illinois, 10 September 2026), Marla Galvan, LICSW, NASW's Senior Practice Associate for clinical social work, writes: "AI-generated notes or documentation are not always accurate." The members-only tipsheet she points to, "8 Ethical Considerations for the Use of Artificial Intelligence in Social Work" (NASW, 2024), is the document to bring to the conversation about whether your agency allows any of this. The written rule is still coming: NASW's Social Work Blog said on 4 September 2026 that "NASW's Code of Ethics Revision Workgroup is recommending supplemental guidance to help social workers apply the Code of Ethics to the use of artificial intelligence."

What is the best AI tool for social workers?

The best AI tool for a social worker who puts identified client information into it is one that signs a Business Associate Agreement, and as of September 2026 that means a purpose-built note tool: all five US tools we checked sign one, and no general assistant's individual plan does. On price, Blueprint at $0.99 a session and Upheal at $1 a note, capped at $69 a month, are the cheapest way to buy a single note with no monthly fee, and Mentalyc Mini is cheaper past 20 notes a month ($19.99 for 40). OpenAI's help center says a BAA for ChatGPT is only for sales-managed Enterprise and Edu accounts and "we don't offer a BAA for ChatGPT Business"; its free ChatGPT for Clinicians tier, which does carry an in-product BAA, is open to physicians, nurse practitioners, physician assistants and pharmacists, and not to social workers. Anthropic's BAA covers the API and Enterprise and sends Free, Pro and Max users to its consumer pages. Whizi signs none.

Outside the US the shape is different. Beam's Magic Notes, the tool English councils buy, sells an organisation licence: £30,000 a year for up to 6,000 hours of recording with unlimited users on its May 2026 G-Cloud price list, so it arrives through your employer or not at all. CaseworkAI is free, takes no account, says notes are deleted after delivery and is registered with the UK Information Commissioner; it is not a HIPAA product and does not claim to be.

Where a general assistant fits is everything that never needed the identifiers: the de-identified note, the court report skeleton, the referral letter with placeholders, the resource list you will verify, supervision prep and translation drafts. Whizi, which publishes this page, is that kind of tool, with GPT, Gemini and Claude on one login and web search on by default. Where it loses for a social worker: it is not a case management system, it signs no BAA and claims no HIPAA certification, and what you type goes to the model provider you picked, so an identified client record belongs in it no more than in ChatGPT Plus. Its uploads expire after up to 30 days and Whizi says it does not use your content to train Whizi-owned models, which is the retention answer to have ready when a supervisor asks.

Can AI write a court report, a referral letter or a resource list?

AI can build the skeleton of a court report and draft a referral letter from your bullet points, and it should never be the source of a single fact, date or phone number in either. The structure is where it earns its keep; the content is yours because the signature is.

For court, work from the template your court expects rather than a model's idea of one. In England and Wales that is the Social Work Evidence Template, published by ADCS with Cafcass (revised April 2021, guidance updated November 2023) to "support a consistent, analytical approach to presenting social work evidence to the family courts." Its final statement runs: the child's details, the social work chronology, analysis of risk and protective factors, child impact analysis and the child's views, analysis of parenting capability, analysis of wider family and friends capability, the proposed care plan with its realistic options analysis, the range of views of parties and significant others, and a statement of procedural fairness. US courts vary by state and county, so the template is whatever your agency's legal team last approved.

The useful prompt is structural. Paste your de-identified chronology as bullets and ask the model to sort them into date order, flag any gap longer than a month, and say which section each event belongs in. Do not ask it to write the analysis. The SWET final statement closes with "The facts in this application are true to the best of my knowledge and belief and the opinions set out are my own," and a court can ask you where any sentence came from. This page is a workflow, not legal advice; what goes in a statement is between you, your manager and your legal team.

A referral letter is the safest drafting job on this page because its shape is fixed: who you are, why you are referring, what has been tried, what you are asking the service to do, how urgent it is, and how to reach you. Give the model those six bullets with CLIENT A in place of the name, ask for 250 words in plain English, then paste the result into your own letterhead and add the identifiers there.

Resource lists are where the model fails you. Ask a general assistant for the food banks near a postcode and it returns names, addresses and phone numbers in the same confident voice it uses for everything, and nothing in the output tells you which ones are real. On 13 May 2026 MIT Technology Review reported Gemini handing out a software engineer's personal WhatsApp number as the customer service contact for a payment app; the same mechanism produces a plausible helpline number that belongs to nobody. Turn web search on so the list arrives with source links, then verify every phone number against the organisation's own site, or by calling it, before a client sees it. How to fact-check AI answers is the longer routine.

Should you use AI to translate for clients?

Use AI to translate your own plain-language letters and reminders, and never as the translation of record for anything that decides a client's rights or benefits. That line is not mine. The US health and human services civil rights rule, 45 CFR 92.201(c)(3), in force since July 2024, says: "If a covered entity uses machine translation when the underlying text is critical to the rights, benefits, or meaningful access of an individual with limited English proficiency, when accuracy is essential, or when the source documents or materials contain complex, non-literal or technical language, the translation must be reviewed by a qualified human translator."

In practice: consent forms, safety plans, court dates, benefit decisions and anything a client will sign go to a qualified translator, and a live interview gets an interpreter rather than a phone held between you. What the model does well is the layer before that. Rewrite your English so a ten year old could follow it, then translate that draft so the interpreter or translator starts from something clear. Keep the client's name out of it; the letter needs a greeting, not an identity. Use AI to translate covers the models and the back-check that catches a bad draft.

How do you prepare for supervision with AI?

Supervision prep is the one job on this page that needs no client data at all, because the material is your own thinking. Write what you would write in a reflective log, role tags instead of names, and send: "Here is a case I feel stuck on, with no identifying details. Ask me the five questions a good supervisor would ask, one at a time. After my answers, list what I have not considered and which of my assumptions is doing the most work." You arrive with your own answers on paper, which is what the hour is for.

The same de-identify, prompt, re-insert routine is the one AI for nurses uses for handover notes; the identifiers change and the rule does not. The decision rule for everything else: if the task needs the client's identity to be useful, it needs a tool with a BAA and your agency's sign-off. If it does not, a general assistant and a set of role tags will do, and most of a social worker's paperwork is the second kind.