AI for Dementia Care Providers: Life-Story Profiles and Activity Plans

A dementia care provider can use AI to draft three documents: life-story profiles, activity plans and family updates. The safe pattern is consent first, facts supplied by the person, family or staff, AI drafting from those facts only, and a named staff member reviewing everything before it is filed or sent. It must not diagnose, monitor or predict.

Drafting is where staff time disappears. A good profile takes a keyworker an hour or more of conversation and writing, and most services have a backlog of thin ones. AI can turn gathered facts into a warm, readable page in minutes. It cannot know the person, so it must never fill gaps with likelihood: a heading marked "not yet known" is safer than a plausible guess that staff then act on.

Follow me on Instagram@sagnikteaches

What AI is for here, and what it is not for

A life-story profile exists so that anyone caring for a person, including agency staff on their first shift, can offer familiarity: a name they like, a song they know, a way of having tea. When language and memory change, that profile becomes part of how the person is understood. The quality bar is therefore high, and the person's dignity depends on it being true.

Connect on LinkedInSagnik Bhattacharya

The bottom row is a firm boundary. Describing risk, deterioration or what a person "probably" feels is assessment, and an AI draft of that carries more danger than time saved. If you are bringing AI in across a service, the 90-day implementation plan for care homes sets out an order that keeps these boundaries in place.

Subscribe on YouTube@sagnikteaches

Consent comes before the first prompt

Three kinds of consent matter, and mixing them up causes most problems.

1. Consent to create the profile. The person is asked, with support and in a way they can follow. If they cannot decide this for themselves, the representative named in your policy decides, following your provider's assessment and decision process.
2. Consent to share it. A profile helps all staff, but sharing with visiting professionals or in a hospital pack needs its own agreement, ideally recorded as a tick-box on the profile form.
3. Consent to use AI in drafting. Add a plain sentence to the consent form: "Staff may use a computer writing tool to help draft this document from the facts you provide. A member of staff will read and correct it before it is used." If a representative objects, draft by hand.

Record who consented, on what date, and whether the person themselves was involved as far as they could be. If your provider has an AI policy, it should say where to file this. The care home AI staff policy tutorial lists what such a policy should cover.

Which tool, and which plan

For text drafting, the main choices are ChatGPT, Claude and Gemini. For a provider the real choice is between personal and organisation accounts. A ChatGPT Business seat and a Claude Team standard seat both cost $20 a month on annual billing or $25 month to month, with a two-seat minimum. The individual plans, ChatGPT Plus and Claude Pro, are $20 a month (Claude Pro is $17 a month billed annually), and both vendors have free plans. Business and team plans do not train on your content by default, and consumer plans let you turn training off, but consumer plans are not designed for records about identifiable residents.

Practical rules for a provider:

• Use a tool your organisation has approved, on an account the provider owns, not a staff member's personal login.
• Where the plan lacks a data agreement you are comfortable with, enter de-identified facts: no surname, address, date of birth or other detail that could identify the person. First names stay in the final document because staff fill them in after.
• Save the prompt and a template in a shared Project (ChatGPT and Claude both offer shareable Projects on business plans) so each keyworker uses the same instructions.

Confirm retention and training settings on the vendor's own pages before relying on them, and recheck them every quarter.

Building a life-story profile: gather first, then draft

The order matters. AI should be the last step, not the first.

Step 1: a gathering sheet

Use a one-page sheet that the keyworker, family or the person completes in conversation. Keep each answer short and in the person's own words where possible. Typical headings:

• Name they like to be called, and names they do not like
• Family and important people, with relationships
• Where they grew up and key places
• Work and what they were proud of
• Interests, music, television, food and drink, including dislikes
• Daily routine habits, such as tea before washing
• Things that comfort, and things that upset or distress
• Faith or cultural practices, if they want them noted
• What helps when they are anxious, in the family's or staff's words

Step 2: a prompt that cannot invent

Turn the facts below into a one-page life-story profile for care staff, written in the third person in a warm, plain style. Rules:
1. Use only the facts supplied. Do not add details, dates, likes, feelings or "probably".
2. If a heading has no facts, write "Not yet known: ask [person] or family".
3. Do not mention diagnosis, stage of dementia, risk, behaviour or capacity.
4. Keep the person's own phrases in quotation marks.
5. Finish with "Questions for staff to ask": any unclear or contradictory facts.

Headings: Name they like | Family and friends | Early life | Working life | What they enjoy | Daily routine | What comforts | What to avoid | Faith and culture.

Facts: [paste the gathering sheet]

Step 3: a worked example with numbers (illustrative)

Say a 24-bed home has 24 residents and 9 profiles are thin or missing. A keyworker takes 20 minutes to interview a family member about a resident and fills the gathering sheet. The prompt returns a draft in under a minute. The keyworker spends 10 minutes correcting it. Then the profile goes to the family member to read, which takes another day. Total keyworker time: about 30 minutes, against roughly 75 minutes writing from scratch. For nine profiles that is nearly seven hours saved across a month, minus time lost if drafts are not checked properly, which is why the checking step is mandatory.

Sample input, illustrative: "Likes to be called [short name], hates [full name]. Worked on the railway for 41 years, started as a porter. Wife (deceased) loved dancing. Daughter visits Sundays. Tea with two sugars, hates coffee. Hums when worried; family says it settles with the radio on. Sings in the shower." The draft might read: "[Short name] (never [full name]) worked on the railway for 41 years, beginning as a porter. ... Not yet known: religion, favourite music." What to fix: the draft may tidy "hates coffee" into "prefers tea". That is a softening of a stated dislike, so restore the exact meaning. Likewise "Hums when worried" must not become "shows anxiety through humming", which is a behavioural interpretation.

Step 4: check with the person and family

Read it aloud to the person if they can engage, and share it with family for corrections. Families often add the best details at this stage: the song, the nickname, the story that always gets a smile. Update the gathering sheet and rerun the draft, so the AI version never drifts from what is on file.

Activity plans: matched to the profile, grounded in what you observe

An activity plan lists options for a person that fit their interests, abilities and routine. AI is useful for generating options, such as five activities for someone who worked on a railway and loves singing. It must not decide what a person can or cannot do. That belongs to staff who know them.

Using only this life-story profile, suggest eight activities for a care home setting. For each, give: the activity, why it links to something in the profile, a simple set-up, how long it takes, and what staff should watch for. Do not state what the person can or cannot do. Do not suggest anything not linked to the profile. Do not mention diagnosis or stage of dementia. Mark each as "Group", "One-to-one" or "Independent".

Profile: [paste approved profile]

Sample output, illustrative, for the same resident: "One-to-one: look through a railway photo book with him. Linked to 41 years on the railway. Set-up: a photo book and a quiet corner. 20 minutes. Watch for: whether photos prompt a story or distress; stop if he seems unsettled." That "watch for" line is the useful part, because it asks staff to observe rather than assume. Treat each suggestion as a trial to log, not a plan to run for a month.

A short template for logging:

• Date and activity tried
• How long the person took part
• What staff observed, in plain words ("hummed along", "turned the page twice", "asked to stop")
• Keep, adjust or drop

After four weeks, the log tells you what works. Paste de-identified log entries (or approved ones in your tool) back into AI to ask for a summary of patterns, but make decisions as a team. For a full calendar rather than individual plans, see planning a care home activities month with AI.

Family updates: kind, factual and checked

Families want to know three things: how their relative is, what they did, and whether anything has changed. Staff often write terse notes. AI can turn those into a warm paragraph. The danger is embellishment: invented feelings, medical-sounding language or reassurance that is not warranted.

A prompt that keeps it truthful

Rewrite these staff notes as a short, kind update for a family member. Use only the facts in the notes. Do not add feelings, health assessments, predictions or reassurance. Do not use words such as "settled", "thriving", "deteriorating", "confused", "agitated" unless they appear in the notes. If a note mentions a fall, injury, medication or health change, do not rewrite it: print "STAFF TO WRITE" in its place. Under 120 words. Sign off with [staff name] only.

Notes: [paste notes]

Sample notes, illustrative: "Ate all lunch. Sat in garden 30 min. Joined singing group, sang along to two songs. Didn't want cake at afternoon tea, asked for toast, had it." Sample output: "[Short name] had a good lunch and enjoyed half an hour in the garden. He joined the singing group this afternoon and sang along to two songs. He asked for toast instead of cake at afternoon tea, so we made him some." Check: is "good lunch" accurate? The notes say "ate all lunch", so say so. Was "enjoyed" supported? Replace with "spent half an hour in the garden" unless staff recorded enjoyment. This is the single most common fix: AI turns observations into feelings.

The routing rule matters as much as the wording. Anything about falls, health changes, medication, incidents or distress is written by a senior staff member, not drafted by AI, and follows your own incident-disclosure process. For ideas on the daily-report workflow in another setting, see how nurseries use AI for parent updates, which uses the same facts-first approach.

Hard cases: no family, painful history and cultural detail

The tidy example above has a daughter who visits on Sundays and a stack of facts. Many residents have neither, and the gathering step is where the care happens.

When nobody can supply the story

Some people have no family nearby, or family who are estranged. Start from what the person says now, from records the provider holds with consent, and from staff who have spent real time with them. Mark every entry with its source ("told us on 3 March", "from the referral letter") so the profile makes clear what is certain and what is second-hand. A short profile built on honest sources is better than a long one that guesses. When AI is drafting, add one line to the prompt: "Keep the source tags in brackets after each fact." Staff can then see which sentences rest on thin evidence.

When the past includes things that hurt

A good profile includes what to avoid. Family may tell you that a person lost a child, served in conflict, or was in an institution as a young person, and that certain topics cause distress. Hold those facts carefully. Put the guidance in the profile as practical advice ("please do not ask about the war; talk about the garden instead") rather than in a narrative paragraph that every visitor reads. Ask the family and, where possible, the person how much they want written down. Where a detail would be painful to see in a document that agency staff and visitors can open, leave it to a verbal handover and a short flag.

Faith, language and culture

Do not let AI generalise. A prompt that says "a resident from a particular community" invites stereotype. Supply the specifics: which prayers matter, which foods are avoided and why, which language the person falls back on, how they want to be addressed. If the person's first language differs from the staff's, ask the family for a short list of phrases and what they mean, and add it to the profile verbatim. Do not use AI translation for comfort phrases without a speaker of that language checking them, because a mistranslated greeting can sound like a command.

A one-page hand-over summary for hospital or agency staff

When a resident goes to hospital, or when agency staff cover a shift, a one-page summary of the profile helps more than the full document. AI is good at compressing a checked profile, provided it is told to drop nothing safety-critical and add nothing new.

Compress this approved life-story profile into a one-page summary for a visiting member of staff. Keep: the name they like and any name they dislike; three things that comfort; three things that cause distress; communication tips as written; food and drink preferences. Do not add anything. Do not summarise any clinical or medical content: print "See care plan" in its place. Maximum 150 words.

Sample output, illustrative: "Call him [short name], never [full name]. Comforts: tea with two sugars, the radio on, talk about trains. Dislikes: coffee. Hums when worried. Daughter visits Sundays. See care plan for health information." That page can sit in a hospital bag. Review it the same way as the profile: someone who knows him reads it and signs.

Keeping profiles alive

A profile written in week one is out of date by month six. Set a review date on every profile, for instance every six months or after any significant change. At review, staff note what has changed ("no longer enjoys the garden, prefers the lounge"), family add anything new, and you rerun the draft from the updated gathering sheet. The prior version is archived, not deleted, so a person's history is intact.

A quick sum shows why this matters. If a 24-bed home reviews each profile twice a year, that is 48 reviews. At roughly 15 minutes each with AI help and a checking pass, that is about 12 hours across the year, or one hour a month. Done by hand from scratch it would be closer to 30 hours. The saving is modest on paper, but the real value is that reviews actually happen, because they are no longer a half-day writing task that gets postponed.

The review step: who reads what, and what they check

A review step that exists only on paper fails in a busy week. Make it a short list that takes under five minutes per document.

1. Source match: every sentence traces to a fact on the gathering sheet or the staff notes.
2. No interpretation: no emotions, no clinical terms, no predictions, no "always" or "never" unless the source says so.
3. Dignity check: would the person, or their family, be comfortable reading every line?
4. Consent check: consent is recorded for this document and this use.
5. Sign-off: the reviewer's initials and date are on the filed version, and the draft chat is deleted.

Make the reviewer a different person from the drafter wherever you can. When the same keyworker writes and reviews, errors that "feel right" slip through. If the person has a representative, the representative sees the profile before it is relied on by staff.

Mistakes that show up in practice

• The invented song. A draft says a resident "loves wartime dance-band songs" because the facts mention the war years. Staff play it, the resident becomes upset. Fix: preferences come only from named sources, and anything inferred is listed under "Questions for staff to ask".
• The profile that sounds like a diagnosis. "[Short name] often becomes confused in the afternoon" is not a life-story fact. Fix: the prompt bans clinical words, and the checker looks for them.
• The shared login. A care assistant uses a personal free account because the organisation has none. Fix: provide an approved tool and a short policy, so people are not improvising.
• The unchecked family update. A cheerful AI paragraph is sent about a day when a resident had a fall that staff recorded in a different system. Fix: the "STAFF TO WRITE" rule and a second-person check on every update.

Setting it up over four weeks

1. Week 1: agree the boundaries with your manager and data lead; choose an approved tool; add the AI sentence to consent forms.
2. Week 2: build the gathering sheet and the profile prompt; test on three profiles with staff who know the residents.
3. Week 3: add the activity prompt and the logging template; run the first trials.
4. Week 4: pilot the family-update prompt for one unit, with a manager reading every output.

Measure two things: keyworker minutes per profile, and the number of corrections you make per draft. If corrections do not fall after the first dozen documents, the prompt or the gathering sheet needs improving, not the tool. For a wider view of AI in the service beyond documents, improving daily note wording with staff review uses the same rules for a more frequent task. If you would like help deciding where AI fits in your organisation, that is what my AI implementation consultation covers, through the contact page.

Questions care providers ask about AI and dementia care

Who gives consent for a life-story profile?

Where the person can decide, they do, with support. Where they cannot, the legal representative or the person your policy names does, following your provider's assessment process. Record who agreed, when, and what exactly they agreed to, including any use of AI tools.

Can AI tell us what a person likes from their care notes?

It can summarise what staff and family have written, but it cannot know what a person enjoys. Treat any 'preference' it infers as a question to ask the person or family, not a fact for the profile.

Is it safe to put a resident's details into ChatGPT?

Not into a free or personal account. For anything identifiable, use a tool your provider has approved with a data agreement in place, or work from de-identified facts. Check your own policy and your regulator's expectations before changing the setup.

Can AI predict decline or flag risk in a resident?

It should not be used that way here. Monitoring and diagnosis are clinical functions with their own regulation. The scope here is drafting documents that people review, not assessing the person.

Further reads

• Care Home Newsletter Ideas: AI-Drafted Family Updates With Consent — Family newsletters drafted with AI, with consent built in.
• AI Home Care Services: Turning Visit Notes Into Family Updates — Turning home-care visit notes into family updates.
• AI Consultant for a Care Home: Do You Need One, and What to Ask — Whether you need outside help and what to ask.
• Care Home Enquiry Form Template: Capture and Answer Leads With AI — Capturing and answering enquiries from families.
• Care Home Mandatory Training Tracker: Renewal Dates and AI Overdue Alerts — Track renewal dates, including AI-policy training.
• Creating Home Exercise Programmes With AI for Physio Patients — Reviewed handouts written with AI, from a clinic setting.
• Care Home Menu Planner: Four-Week Cycle With AI and Dietary Checks — A four-week care home menu cycle drafted with AI, with an allergen matrix, shopping list and a cook or dietitian sign-off step.
• Use AI to Write a Care Plan Review From Daily Notes, Safely — Draft a care plan review summary from a month of daily notes without exposing identities, then trace every line back before sign-off.
• Can AI Write Care Plans? What a Home Care Provider Must Check First — Provider-level checks before AI drafts a home care plan: accuracy, consent, de-identification, manager sign-off and an audit trail.
• How Can AI Help in Care Homes? Ten Back-Office Uses — Ten care home back-office jobs where AI saves real time, each with an example, a check and a clear line on what stays with staff.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT pricing pages (individual and business), Claude pricing page, Anthropic privacy centre article on model training (all checked 4 November 2026); general dementia-care documentation practice.