AI for Hospice Teams: Back-Office Jobs Only, and Where It Must Not Go

A hospice can use AI for back-office work: volunteer rotas, fundraising and grant drafts, policy upkeep, staff and volunteer communications, supplier and finance admin. It must never be used for patient care, messages to patients or families, or any bereavement contact. Those are written by people who know the family.

The line is not about whether the tool is capable of writing a gentle sentence. It is. The line is that a hospice's value is that a person turned up. A condolence message that a family later learns was generated costs far more trust than the hour it saved. So the discipline is the same each time: AI prepares work that nobody outside the building will read as personal, and a named person approves it. The rest of this tutorial sorts the jobs into those that fit, those that need care and those that are closed.

Follow me on Instagram@sagnikteaches

Sort every task into three lanes

Print a one-page table and fill it in with your team. It takes an hour and prevents most arguments later.

Connect on LinkedInSagnik Bhattacharya

When a task is unclear, put it in the stricter lane until the clinical lead and the person responsible for data protection agree otherwise. If you want a ready format for a volunteer-facing policy, a sample one-page AI policy for staff and volunteers is easy to adapt.

Subscribe on YouTube@sagnikteaches

Volunteer rotas without the Sunday-night spreadsheet

Hospices often run on volunteers: reception, tea trolley, garden, charity shop, driving, ward befrienders. Rotas break when someone is ill, and the coordinator rebuilds them by hand. Give an assistant the pattern and the people as initials or role codes, and ask for a draft that meets coverage rules.

Draft the volunteer rota for next month. Slots: reception Mon-Fri 9-13 and 13-17 (2 people each), garden Tue/Thu 10-14 (3 people), shop Sat 10-16 (4 people). Volunteers (initials) and availability: [paste initials and days]. Rules: no one does more than 2 sessions a week; V7 and V12 must not be paired (reason confidential); new volunteers (V20, V21) are always with an experienced person. List any slot you could not fill.

Say a coordinator has 38 volunteers and the 13 slots above, which need 30 volunteer shifts a week. A draft that used to take two evenings takes twenty minutes, plus ten minutes of checking against holiday notes. The sample output is a table plus a short list: "Reception Thursday afternoon has one person; V3 or V15 could fill it." The coordinator decides, because she knows that V3 has a hospital appointment that week. Keep the initials key in your volunteer database, never in the prompt. For the message side, volunteer rota messages and recruitment posts has wording you can borrow.

Volunteer adverts, onboarding packs and reminders

Recruitment text is open-lane work. Describe the role honestly, including what is hard about it, and ask for a version at a plain reading level. An advert for a ward befriender should say clearly that training and supervision are provided and that the role involves being around people who are dying. Do not let a tool soften that into "rewarding and uplifting" language; applicants who are surprised leave early.

Onboarding checklists are another good fit: a first-week schedule, a list of documents to collect, a glossary of hospice terms in plain words. The volunteer recruitment and onboarding tutorial goes deeper; here, the point is only that none of it involves patients. A check worth making is whether the checklist includes safeguarding and confidentiality briefings that must be delivered by a person, and whether the draft has quietly dropped them.

Fundraising: careful lane, with a hard rule about stories

Fundraising drafts are where AI is most tempting and most dangerous. The helpful uses are narrow: turning your own bullet points into a letter structure, shortening a long appeal for social media, producing three subject-line options, tidying a thank-you template and checking tone for consistency across a campaign.

The hard rule: no invented stories, quotes, statistics or beneficiaries. If an appeal needs a story, it comes from your supporter-care records, with consent recorded, and is written or checked by someone who knows the family's wishes. A realistic mistake is a draft that includes "Margaret spent her last days watching the garden she loved". If Margaret is a name the assistant made up, you have published fiction under a hospice's name. If she is a real patient, you have used her story without consent. Either way, a reviewer must read every line against a source.

Say the fundraising officer is writing a spring appeal for a new garden room. Her notes: cost $42,000, $28,000 raised, the room will give families a quiet place, planning permission granted, builders start in June. A good prompt tells the assistant to use only those facts and to leave a clear gap where a real story could go.

Write a 250-word appeal letter using ONLY these facts: [paste facts]. Do not invent any person, patient, quote or statistic. Where a real story would go, write [STORY, TO BE SUPPLIED AND CONSENTED]. Warm, plain English, no pressure language, end with one clear way to give.

The output will include the bracket as instructed. That is what you want. The officer fills it from consented material, then gets a second person, often the chief executive, to read the whole letter against the facts. A figure check closes the job: $42,000 minus $28,000 is $14,000 still to raise, and the letter should say exactly that.

Thank-you templates are the other common request, and they show the difference between help and replacement. A before: "Thank you for your generous gift. Your support makes a real difference." An assistant can draft four alternatives in different lengths, but the fundraising officer should pick one, add one concrete line about what the money will do (the garden room, a new bed, a month of volunteer training), and personalise the greeting from her database. Where a donor gave in memory of someone, the thank-you is a person's job, written or at least edited line by line. The template sits in the careful lane; the in-memory reply sits in the closed one.

Grant applications and impact reports

Funders ask the same questions in different boxes. AI helps by reorganising your own earlier answers to fit a new word count, and by spotting inconsistencies, such as a service described in two ways. Paste only your own published and approved text. Hold back figures about patient numbers until a person has checked them against records, and never paste case-level information. The assistant can draft a logic-model table (activity, output, outcome) from your bullet points, which is tedious by hand.

Leave claims about outcomes to your data. A funder who reads "families report high satisfaction" will ask for the survey. If you do not have one, the sentence goes, however good it sounds.

Policies, procedures and board papers

Hospices carry heavy policy loads: infection control, volunteering, safeguarding, data protection, whistleblowing, health and safety. An assistant can compare two versions of a policy and list what changed, produce a plain-language summary for volunteers, and draft a review checklist. It must not decide what the policy should say on a clinical or regulatory point. A qualified person owns that, and the document carries a named approver and review date.

For board papers, give the assistant your finished figures and ask for a one-page narrative and three questions a trustee might ask. Check each number against the finance system. AI explains numbers well and miscopies them occasionally, so the check is not optional. Meeting minutes work if you provide your own notes without patient detail: ask for actions, owners and dates in a table, and flag anything ambiguous with [CHECK].

A worked grant example, illustrative. A trust wants 500 words on "how the project will be sustained", and the hospice has answered a similar question before at 900 words. Paste the earlier answer and the new word limit and ask for a version that keeps every fact and cuts repetition. The draft comes back at 490 words. The grant writer then checks three things: no number has changed, no new promise has appeared (assistants like to add "we will continue to monitor and evaluate"), and the tone still sounds like the charity. She finds one invented commitment, to publish an annual report, deletes it and submits. The time saved is perhaps an hour per application, and with eight applications a year that is a working day returned.

Retail, events and supplier admin

Many hospices run charity shops and events. These are open-lane. Descriptions for donated items, rotas for shop volunteers, event risk-assessment drafts and supplier emails all suit AI. A shop manager photographing donated stock can ask an assistant for short listing descriptions, then check condition and price herself. An event lead can draft a run sheet for a fundraising evening with no memorial element, such as a quiz night, and a person reviews safety and licensing points. Keep memorial or tribute events in the careful lane, because their wording is personal.

The closed zone, in plain words

These must never go through AI, and your policy should say so without hedging.

• Patient and family communications of any kind, including appointment letters, visiting information and anything in a care plan.
• Bereavement: condolence cards and letters, anniversary messages, memorial invitations, follow-up calls and the scripts for them.
• Clinical notes and summaries, handover notes and anything that records a person's wishes at the end of life.
• Complaints and incident responses, which need a human accountable voice.
• Staff matters such as capability, absence and grievances.
• Chatbots that talk to patients, families or the bereaved. A website bot answering opening hours is different and sits in the careful lane, with a clear hand-off to a person.

Why be so strict about the last one? Because grief arrives in odd places. A bereaved person might type into a website chat box at midnight. A bot that answers with a cheerful FAQ link has failed that person. If you run a chat tool at all, its first message must say a person will reply and give a phone number, and anything emotional goes to a human the same day.

Using AI to support the person who writes, without writing for them

There is a legitimate, modest use in the closed zone: helping a colleague with layout or spelling when they have written something themselves. A bereavement support worker might ask for a plain template skeleton for a memorial-service information sheet that has no personal wording, then fill it in by hand. The test is where the personal words come from. If they come from the tool, the job is closed. If they come from a person and the tool only formats, you are in grey territory and should check your policy and the clinical lead first. When in doubt, do it by hand.

Patient data, nonprofit plans and the AI register

• Never type patient, family, bereaved or staff-case details into a general AI tool.
• Prefer business or nonprofit plans with training off by default, and check the data page of each tool. As of November 2026, OpenAI for Nonprofits prices ChatGPT Business Standard seats at $8 per user per month billed annually ($10 monthly, Premium seats excluded); Claude for Nonprofits lists a standard seat at $8 per user per month with a two-seat minimum; and the no-cost Google Workspace for Nonprofits edition includes the Gemini app with enterprise-grade data protections. Eligibility rules apply, so check each vendor's nonprofit page before assuming your charity qualifies.
• If a records or fundraising system adds AI features, ask the supplier where data is processed and whether you can switch the feature off, and have the data protection lead approve before it is enabled.
• Keep a register of AI uses: tool, purpose, data involved, approver, review date.

Voice and tone checks matter too. Ask whoever knows your supporters best to read AI-assisted fundraising text aloud and mark any line that does not sound like the hospice. Sentimental filler such as "journey" language is a common tell, and supporters notice it.

Ten-minute scenario drill for staff and volunteers

Policies stick when people have practised them. At a team meeting, read these scenarios aloud and ask which lane each sits in and what the person should do. Answers are in brackets for the facilitator.

1. A volunteer wants to paste a family's feedback email into a chat tool to get a summary. (Closed: it names a family and may refer to a patient. Read it and summarise it by hand, or remove every identifying detail with the manager's agreement.)
2. The shop manager wants an assistant to write 30 item descriptions. (Open: donated goods, no personal data; she checks condition and price.)
3. A trustee suggests a website chatbot for supporters. (Careful: only for opening hours and giving, with an immediate route to a person and sign-off from the data protection lead.)
4. A nurse asks if AI can draft a handover note to save time. (Closed: clinical records follow the clinical lead's rules, not the back-office rule.)
5. The finance lead wants a plain-English version of the budget for trustees. (Open or careful: figures checked by the finance lead before it is circulated.)
6. A colleague asks for a kind message to send to someone whose relative died last week. (Closed: write it yourself, even if it is short and imperfect. Imperfect and human beats polished and generated.)

Record the session in the AI register. Repeat it when someone joins and when a new tool is added.

A filled-in policy paragraph you can adapt

Illustrative wording for a staff and volunteer handbook:

Our hospice may use AI tools to help with back-office tasks such as rotas, drafting and summarising documents that contain no personal information about patients, families or colleagues. We never use AI to write to, or about, patients, families or people who have been bereaved, or to produce clinical records. Everything AI helps to draft is read and approved by a named member of staff before it is used. If you are unsure whether a task is allowed, ask your manager first and do it by hand in the meantime.

Share it in your induction and revisit it when a new tool appears. Include a line that nobody will be blamed for reporting an AI mistake promptly; otherwise mistakes stay hidden until they matter.

Minutes and actions from a team meeting, a small example

Meeting notes are a quiet win in the open lane when they contain no names of patients. Say the volunteer services team meets monthly and the coordinator types rough notes: "garden rota gaps in July, need 2 more for Thursdays, shop till upgrade quote $1,250 vs $980, agree cheaper if it takes card, fundraising quiz 14th need raffle prizes, background checks for 3 new starters outstanding". She pastes them with this request: list decisions, actions (owner, due date) and open questions, and mark anything unclear with [CHECK].

A realistic draft returns three decisions (cheaper till if it takes card payments), four actions with owners left blank because none were named in the notes, and an open question about which three starters are outstanding. Note what the tool did right: it did not guess owners. If it had written "Action: coordinator to chase checks by Friday", you would reject that line, because nobody agreed a date. Circulate the final version within 24 hours while memories are fresh. The whole job takes ten minutes instead of forty.

Timing three jobs to see whether AI earns its place

Pick three jobs, time them for two weeks before and after, and note any errors found at review. Rota building, grant reformatting and policy comparison often save the most. If a job saves under 30 minutes a month, drop it; the review time and the risk are not worth it. If your team finds the checks more tiring than the original task, that is a signal too. Care organisations run on goodwill, and any tool that adds worry loses it.

If you would rather work through the lanes with someone, that is what my AI implementation consultation covers; get in touch.

Questions hospice managers ask about AI

Can AI help write letters to bereaved families?

No. Bereavement letters, condolence messages and any contact with a family after a death should be written by a person who knows them. AI can help with the layout of a standard template and with checking spelling of a name you supply, but the words and the decision to send are human.

Is it acceptable to use AI for fundraising appeals?

Yes, for drafting and tidying, with a person approving every word and every claim. Never let it invent a patient story, a quote or a statistic. Real stories need the person's or family's consent and come from your own records, written or checked by someone who knows the context.

Do small hospices need paid AI tools?

Often not at first. A paid assistant on a nonprofit or small-team plan covers most back-office drafting, and some suites already include AI. Check whether your charity qualifies for nonprofit pricing, and read the data terms before staff start.

What should a hospice AI policy say about patient information?

That no patient, family or bereaved person's details are ever typed into a general AI tool, that any AI feature in a clinical or records system is approved by the clinical lead and data protection lead first, and that staff report mistakes without blame.

Further reads

• Community Centre Volunteer Form: AI Sign-Ups and Rota Reminders — A sign-up form and reminder flow for volunteers.
• Volunteer Onboarding Checklist for Charities: Build One With AI — A volunteer onboarding checklist you can build with AI.
• Google AppSheet for Nonprofits: Volunteer Rota App Without Code — A no-code rota app if spreadsheets keep breaking.
• AI for Church Administration: Rotas, Notice Sheets and Room Bookings — Rotas and notices handled by a similar volunteer-led team.
• How Funeral Directors Can Use AI for Admin With Care — Another sector where tone is the whole job.
• AI for Fertility Clinics: Admin Jobs Only, and Where It Must Stop — Another sensitive care setting with an admin-only rule.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI for Nonprofits help article; Claude for Nonprofits page; Google Workspace for Nonprofits edition help page; OpenAI data controls FAQ; Anthropic commercial model-training help article; checked 4 November 2026.