AI Use in Long-Term Care Facilities: Where Resident Data Must Not Go

A long-term care facility can safely use AI for work that contains no resident information: activity ideas, rotas, job adverts, policy drafts, supplier emails and general family newsletters. Resident data must never go into a general chat tool. Anything describing a real resident's health, care, behaviour or family belongs only in a system you have a contract with.

The line is easy to state and hard to keep, because the dangerous cases look harmless: a care assistant tidying a handover note, a manager asking for help wording a family email about a fall. Both feel like admin, and both carry a resident's health details into a tool the facility has no contract with. Once the information in the building is sorted into zones, those cases stop being judgement calls.

Follow me on Instagram@sagnikteaches

Three zones for everything your facility holds

Sort the information in your building into three zones. The zone, not the task, decides which tool is allowed.

Connect on LinkedInSagnik Bhattacharya

Two rules make the table work. First, when in doubt, a piece of information goes in the higher zone. Second, mixing zones moves the whole document up: a rota (amber) with a note saying "cover for Mrs P's one-to-one" (red) is now red. The data classification tutorial explains the method for any business if you want a template to copy.

Subscribe on YouTube@sagnikteaches

Why "I removed the name" is not enough

Most accidental leaks happen after someone has tried to be careful. The name is gone; the person is still identifiable. In a small facility, a handful of facts point to one resident.

Here is an illustrative before and after. A care assistant wants help writing a handover summary.

Before (not safe): "Mrs [surname] in room 14 had a fall at 03:10, was on blood thinners, daughter [first name] was called, wants to move to the ground floor."

After the name is removed (still not safe): "Resident in room 14 had a fall at 03:10, on blood thinners, daughter called, wants to move to the ground floor." Room number, time, medication and family details together still point to one person, and the content is still health information.

Safe version, with no resident detail at all: "Draft a short handover template with headings for falls, medication changes, family contact and room requests, with space for a time and a signature." The template is useful, and a person fills in the real details inside the care record system.

A quick test for staff: "Could a colleague work out who this is from this text alone?" If yes, it stays out. Typical identifiers that slip through include room numbers, dates of birth, rare conditions, a distinctive incident ("the lady who left through the fire door"), a family member's name, and a photograph's background.

The jobs that need no resident data

There is plenty to do on the green side. These are the admin jobs that save time without ever touching a record.

• Activity ideas and calendars. Seasonal, seated, low-noise, themed, for mixed abilities. Ask for ideas and a sample week; add real residents' preferences yourself, in your own system. The activities calendar tutorial covers a month with resident input collected on paper.
• Job adverts and interview questions. Describe the role and shift pattern; remove anything unique to a current employee.
• Policy and procedure drafts. A first draft of a visiting policy, a hand-hygiene poster or a laundry procedure. The home's manager reads and adapts it; the model does not know your regulator's current wording.
• Training quizzes. Ten questions from a policy you paste in, with the answers, for a team meeting.
• Supplier and maintenance emails. Chasing a boiler contractor, a food order, a mattress delivery.
• Rota drafts, without leave reasons or any resident-specific cover note.
• General family newsletters. Opening times, an event, a new menu rotation, written without any individual's name or story.

Menu planning sits on the border. A cycle menu is green; the residents' dietary needs and allergies are red. The four-week menu planner keeps those apart: the tool drafts dishes, and the cook checks them against the dietary record inside your own system.

A filled-in example: a job advert with nothing sensitive

Prompt: "Write a job advert for a part-time night care assistant at a 40-bed long-term care facility. Two nights a week, 20:00 to 08:00, experience preferred but training provided. Warm, plain tone, under 200 words. Do not promise pay figures." The result is a clean, usable advert in seconds. What you fix: any invented benefits, any claim about qualifications you do not require, and any pay or location detail. Nothing in the prompt identified a person, so the tool choice hardly matters.

Jobs that need resident data, and what they require

Some of the most useful AI work involves the red zone: summarising daily notes into a weekly review, drafting an incident report, preparing a family update with real detail, or transcribing a care-planning meeting. These can be done, but only inside a tool you have contracted for that purpose. Four conditions decide it.

1. A contract that names the product and the plan. In health settings that is usually a business associate agreement, or the equivalent data-processing terms where you operate. Vendors differ: OpenAI's help centre says it does not offer a BAA for ChatGPT Business, offers one for the API, and handles ChatGPT Enterprise and Edu through sales. Anthropic's BAA covers specific services such as its API and Enterprise plans once the Primary Owner has activated it; Claude's pricing page marks the Team plan "No" for that health-data offering, and the commercial article is separate from its guidance for personal plans. Google Workspace requires an administrator to accept a BAA before protected health information is used, and third-party add-ons are outside it. Read each vendor's current page; these change.
2. Training off, retention known. Business plans are generally not trained on by default; personal plans often have a switch that a staff member can flip back. Check how long chats are kept: OpenAI's help centre says a ChatGPT Temporary Chat may be kept for up to 30 days for safety purposes, while an ordinary chat stays in the history until someone deletes it or a workspace retention policy removes it.
3. Access control. Who in your organisation can see what was typed? Can a departed employee still see it?
4. A human reads the output. A summary that omits a medication change is worse than no summary.

If a tool cannot tick all four, it is a green-zone tool only. The longer privacy discussion is in how to stop AI tools training on your data.

An approved-tool test your administrator can run in an afternoon

Staff will use whatever is on their phone unless the facility offers something better. Make "approved" easy to answer with a one-page checklist and a short list posted in the office.

1. Name of the tool and the exact plan.
2. Which zone it is approved for: green only, amber, or red.
3. Contract on file, with the date it was signed and who signed it.
4. Where data is stored and for how long; whether it is used for training.
5. Who has accounts, and how an account is removed when someone leaves.
6. What the tool must not be used for.
7. Date of next review (a calendar reminder every six months, because plans change).

Here is the checklist filled in for an illustrative green-zone entry. Tool: a business-plan chat assistant, two seats. Approved for: green zone only (activities, adverts, policy drafts, quizzes). Contract: the vendor's standard business terms, saved in the shared drive on the day of signing, signed by the administrator. Data: training off by default on this plan, retention checked on the vendor's page and noted. Accounts: the administrator and the activities coordinator; accounts removed on the leaving checklist. Must not be used for: anything about a named or identifiable resident. Review: every six months. A sheet like that takes ten minutes to fill and settles most arguments, because "is this allowed?" now has a written answer.

Keep the approved list short. Three tools that everyone knows beat fifteen nobody remembers. For the questions to put to a vendor, what to check in a privacy policy gives the wording to look for.

A 52-bed facility works it out, with numbers

Here is an illustrative picture of an operator doing this properly. A 52-bed facility has 70 staff over three shifts. The administrator asks every team lead one question: "What do your people already do with a chat tool or phone app?" The answers (all figures illustrative):

Of 27 uses, four put red-zone data into a chat tool outright and eleven more (family emails and translations) risked it. The administrator does not ban the tool. They publish a one-page rule, add the nine activity-idea users to the approved business plan, stop the handover-note practice that day, and book a team meeting to explain why. Total cost: an afternoon of listening, one short meeting, and the seats on the approved tool. The facility ends up with fewer risks and more productive use, which is the usual result of asking staff before banning anything.

Photos, voice recordings and family messages

Three kinds of material cause trouble because they do not look like "records". A photograph of an activity shows faces, names on a whiteboard and sometimes a medication trolley in the background. A voice recording of a handover or a family call holds names, conditions and tone. A message to a family about a resident is a record the moment it describes the resident.

The safe pattern for family messages is a template with blanks. An illustrative prompt for any approved chat tool: "Write a gentle, plain-English template for telling a family that a resident has had a minor fall and has been checked by the nurse on duty. Use placeholders for the family member's name, the resident's name, the time and the follow-up. Do not include medical advice. Keep it under 120 words." The sample output is a short message with [family name], [resident name] and [time] in square brackets. A member of staff fills those in inside the care record system or email client, never in the chat window. What you fix in the draft: any reassurance the nurse has not given ("there is nothing to worry about"), and anything that sounds like a diagnosis.

For photographs, make the rule about the camera as well as the tool. Do not upload resident photos to an image or chat tool for captions, and keep the consent status for each resident next to the photo library, not in someone's memory. For voice, treat any recording of a handover, call or meeting as red-zone content: it goes only into a transcription service you have a contract with, and the residents' meeting minutes tutorial shows how to keep minutes useful with no personal detail in them.

Agency staff, volunteers and students

Your rules are only as good as the newest person's knowledge of them. Agency and bank staff often work a handful of shifts and may carry habits from other places, including using a personal chat app to tidy notes. Volunteers and students on placement may not have had any data training at all. Add three lines to the induction pack: which tools are approved, a plain sentence on the red zone, and the name of the person to tell if something goes wrong. Ask the agency whether its own staff guidance covers AI tools, and put the expectation in the booking terms.

A quick example of how this shows up. A student on placement photographs a whiteboard of the day's resident priorities to "type it up later" with a free app. Nobody told her not to. The fix was a laminated card in the staff room and a five-minute conversation, not a disciplinary. Most breaches of this kind are gaps in information rather than bad intent.

Staff rules to print and put on the wall

Staff will read a rule that fits on one side. Adapt these:

1. Use only the tools on the approved list, and only for the zone shown next to each.
2. Never type a resident's name, room number, date of birth, condition, medication, behaviour, family detail or photo into any tool that is not marked red-approved.
3. Removing a name does not make it safe. If someone could tell who it is, it stays out.
4. Never use a personal phone or personal account for anything about a resident.
5. AI output is a draft. A person reads it and is responsible for it before it reaches a family, a record or a regulator.
6. No AI tool gives clinical advice, changes a care plan, or decides on medication, restraint, safeguarding or end-of-life matters.
7. If you put something in the wrong place, tell your manager the same day. Nobody is disciplined for reporting honestly.

Rule 7 matters more than it looks. A facility where staff fear blame will never hear about the mistake until it is much bigger.

If resident data has already gone somewhere it should not

It will happen eventually, so plan for it. First, stop and note what was entered, in which tool, on which account, and when. Second, tell the manager and the data lead the same day. Third, check the tool's retention and deletion options: delete the chat, and where a vendor offers a deletion request, send it. Fourth, decide with the data lead whether the incident must be reported under your own procedures or to a regulator or the families affected; that is a judgement for your administrator and, where needed, a qualified adviser. Finally, record what changed: a clearer rule, a better approved tool, a short retraining. The incident log is the evidence that you take it seriously.

Checking the rules work after three months

A policy nobody follows is worse than none because it creates false confidence. Do three simple checks each quarter. Ask a sample of staff which tools they use and for what, in a relaxed conversation, not an interview. Look at your account list for the approved tools and remove anyone who has left. Count incidents and near-misses; a rising count of near-misses reported is a good sign, because people are talking. If the checks show staff still using personal accounts for red-zone work, the fix is usually a better approved tool or an easier route, not a stricter rule.

Before any new job joins the approved list, write down three things: which zone its data sits in, who reads the output, and what a wrong answer would cost a resident. A job that cannot be described in those three lines is not ready.

Questions from administrators and care managers

Can staff use ChatGPT for care notes?

Not on a personal or standard chat account. Care notes identify residents and describe health, so they belong only in a system covered by a signed business associate agreement or equivalent contract. Staff can use a chat tool for generic tasks that contain no resident detail.

Is removing the resident's name enough?

No. A room number, a date of birth, a rare condition, a family member's name or a distinctive incident can identify someone without the name. Treat any note describing a real resident's health as identifiable even after the name is gone.

What AI jobs are safe in a care facility?

Jobs with no resident information: activity ideas, staff rota drafts, job adverts, policy drafts, supplier emails and training quizzes. Family newsletters are safe if they contain no individual's details. Everything touching an individual's record needs an approved, contracted tool.

What should happen if someone pastes resident data into the wrong tool?

Stop, note what was entered and where, tell the manager the same day, and follow your incident procedure. Do not try to hide it. The sooner the record is deleted and the vendor's retention rules are checked, the easier the fix.

Who approves a new AI tool?

One named person, usually the administrator or data lead, using a short checklist: contract, data location, training use, retention, who can access it. Staff must not approve their own tools.

Further reads

• Care Home Medication Audit: A Monthly Template and AI Action Log — An audit routine where AI helps with the log, not the decisions.
• How Funeral Directors Can Use AI for Admin With Care — Another sensitive-records setting where admin AI needs care.
• Is It Safe to Put Customer Data Into ChatGPT? — What consumer and business plans do with what you type.
• Questions to Ask Before Buying AI That Touches Client Data — Vendor questions for any tool near sensitive records.
• Telling Patients You Use AI Note-Taking: Consent Wording That Works — Consent wording when a tool listens or writes notes.
• Is an AI Scribe Worth It for a Small Private Clinic? — How to judge a records tool against its cost and risk.
• AI Use Case Template: Score Every Idea on One Page — Score each new AI job for value and risk on one page.
• AI and Senior Living: Where a Small Care Home Should Start — A small care home's starting order for AI: family updates, activities, rotas and enquiry replies first, safeguarding and resident records kept human.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Help Center on business associate agreements and Temporary Chat retention; Anthropic privacy centre on business associate agreements and model training for commercial customers (checked November 2026); Google Workspace help on health-data compliance; vendor privacy settings pages. Check current terms before relying on them.