In the first 90 days, an agency owner should set rules and an approved-tool list in month one, then automate two or three office jobs (replies to enquiries, rota messages, recruitment follow-up) in month two, and only in month three pilot help with documentation under a named checker. Measure hours, response times and errors, and spend only once one task proves its worth.
Care is different from most small businesses because the records describe real people's needs, and the people handling them are often on their phones between visits. That is why the order matters: rules before tools, office tasks before care records, and a person's judgement kept on anything that touches a care decision. The plan below is built around a worked example, an agency with 30 carers, and it uses the same shape you can adapt to a smaller or larger team.
The example agency and what the plan is trying to fix
Say a domiciliary agency has 30 carers, around 70 clients and about 600 visits a week. The office has six people: the owner, a registered manager, two schedulers, a recruitment and training coordinator, and an administrator who also does invoicing. The owner's complaints, all illustrative, are familiar. The schedulers spend most mornings on calls and messages about visit changes. New enquiries from families sometimes wait a day for a reply. Recruitment ads are rewritten from scratch each time, and applicants drift away between steps. Policies are out of date. Office hours disappear into typing.
The plan aims at four numbers, set before starting and recorded again at day 90:
- Office hours a week on rota communication (baseline: about 30 across two schedulers).
- Time from a new enquiry to a human, useful reply (baseline: about 20 hours on average).
- Days from a vacancy being advertised to a signed offer (baseline: 38).
- Number of AI-related errors found by a checker, logged by type.
The last measure is deliberate. A plan that only measures speed will drift towards risk. Counting the errors caught tells you whether the checks are working and which tasks need tighter limits.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWho owns what
| Role | Owns in the plan |
|---|---|
| Owner | Budget, the approved-tool list, weekly 20-minute review, final call on stop or continue |
| Registered manager | Anything touching care records, care plans, safeguarding or medication; the red-lines list |
| Lead scheduler | Rota messages and visit-change workflow pilot |
| Recruitment coordinator | Adverts, candidate follow-up, interview scheduling pilot |
| Administrator | Tool register, log of errors found, the numbers dashboard (a simple sheet) |
Notice that there is no "AI lead" who sits outside these roles. A tool owned by everyone is owned by nobody, so each pilot has one owner and one checker, and they are different people.
Days 1 to 30: rules, register and one safe win
Week 1: find out what is already happening
Ask every member of staff, carers included, three plain questions: which AI tools or assistants do you use for work, what do you put in them, and what do you do with what comes out? Expect surprises: a carer who uses a free chatbot to word incident notes, a scheduler pasting client details into a translation tool. The goal is not blame, it is a register. Ask also what built-in AI features exist in the software the office runs every day, including your rota, care planning, phone and email systems, and whether they are switched on.
Week 2: write the red lines and the approved list
The registered manager and owner agree a one-page set of red lines. An illustrative version:
AI RULES FOR [AGENCY] (version 1, reviewed [date])
APPROVED FOR NOW
- [Tool A, business plan]: office drafting, job adverts, policy
drafts, summarising non-identifying text.
- Built-in features in [rota software] once the manager has
checked their settings.
NEVER PUT INTO ANY AI TOOL
- Client names, addresses, health details, care plans, visit
notes, medication, safeguarding information.
- Carer personal details beyond a first name.
NEVER LET AI DECIDE
- Care plan content, risk assessments, medication guidance,
safeguarding judgements, disciplinary outcomes.
CHECK BEFORE SENDING
- Every AI-drafted message to a client or family is read and
edited by a named staff member.
ERRORS
- Anything wrong, odd or risky is logged the same day on
the AI error sheet.This is the document that lets you say yes quickly to later requests, because the boundaries are already written. Run it past your adviser or the body that regulates your agency if you are unsure how your obligations apply. Draft it with AI if you like, but the registered manager writes the final wording.
Week 3: choose the tools, within a budget
Keep the first spend small. Two options that do not need a large commitment: if your office already runs Microsoft 365, Copilot Chat is included at no extra cost with eligible business licences and is a starting point for non-identifying drafting, provided staff sign in with their work account so Microsoft's enterprise data protection applies; or business-plan seats for the office team, where ChatGPT Business lists at $25 per user per month billed monthly or $20 billed annually, from two users (as of November 2026). OpenAI says Business content is not used for training by default, and Microsoft says Copilot Chat prompts and responses under enterprise data protection are not used to train its foundation models, which is a reason to prefer either to free personal accounts. Check the vendor's page when you buy, since plans change.
For six office seats at the annual rate the cost is about $120 a month. Set a ceiling in advance: say $200 a month for the 90 days, covering seats and any add-on, with a review before exceeding it.
Week 4: one safe win with no client data
Pick a task that carries no personal data and has an obvious before and after. Recruitment adverts are ideal. The coordinator gives the assistant the role, hours, pay band and the agency's real selling points, and asks for three versions of an advert and a screening-questions list. Illustrative before and after: the old advert is a paragraph of generic phrases; the new one states shift patterns, travel arrangements, training provided and how the interview works, in plain language. The coordinator edits for accuracy, because a promise in an advert is a promise to a candidate. AI for recruiting carers: ads, screening and follow-up goes deeper.
Days 31 to 60: three office jobs, each with an owner and a measure
Job 1: enquiry replies within the hour
Families enquire at odd hours, and the first reply shapes whether they stay. The administrator builds a reply template set for the five most common enquiries: price and funding questions, availability, what a first visit involves, how you vet carers, and how to complain. The assistant helps write and tighten the templates once; a person sends each reply and personalises it. If you later add an after-hours answering service, see AI answering for home care agencies: after-hours call-outs for what to script and what to hand to the on-call manager.
Measure: median time to first human reply. Illustrative target: from 20 hours to under 4.
Job 2: rota and visit-change messages
Scheduling takes the most hours, and most of it is messaging. The lead scheduler lists the 12 message types sent in a typical week: a visit moved, a carer running late, a new client intro, a cancelled visit, a request to cover, and so on. The assistant turns them into short templates with fields: [carer first name], [time], [reason]. The schedule system fills the fields. No client details go into the assistant, only the template text. Where your rota software has its own messaging or AI features, ask the vendor what they do with the data before enabling them; what the AI features in care software add lists the questions.
Measure: scheduler hours a week on messages. Illustrative target: from 30 to about 20.
Job 3: candidate follow-up
A vacancy is lost when a good candidate hears nothing for three days. The coordinator sets up a simple sequence of messages: application received, interview invitation, reminder the day before, post-interview update, offer or decline. The assistant drafts them; the coordinator personalises and sends. If your applicant tracking or recruitment software can send them automatically, use that. Measure: days from advert to offer. Illustrative target: from 38 to 30.
A sample prompt and what you would fix
Write five short email templates for a home care agency to send
to job applicants: (1) application received, (2) interview
invitation with [date], [time], [place], (3) reminder the day
before, (4) after interview, we will decide within [number] days,
(5) polite decline. Warm, plain English, under 90 words each.
Do not promise a job, pay or hours. Use [first name] and
[recruiter name].Illustrative output for template 4: "Hi [first name], thank you for coming in today. We enjoyed meeting you. We aim to let you know within [number] days. If you have any questions in the meantime, just reply to this email. Kind regards, [recruiter name]." What you would fix: check that "within [number] days" is a deadline you can actually meet, and remove any line that sounds like an offer if one slipped in.
Days 61 to 90: documentation help under a named checker
Only now do you approach the sensitive material, and only for tasks where a person stays accountable. The registered manager chooses one pilot. Reasonable candidates are tidying the wording of the office's own policy drafts, preparing the structure for audits, or helping a supervisor write up supervision notes from their own bullet points. Carer visit notes are the most sensitive, because they are part of the care record and must reflect what was observed. If you pilot anything there, do it with a handful of carers, a tool approved for personal data under your data-protection terms, an explicit rule that the carer reads and corrects every word, and the registered manager reviewing a sample each week. Turning visit notes into family updates and what a provider must check before AI helps with care plans go into those risks in detail.
Care plans and risk assessments themselves stay with the registered manager and the people who assess the client. AI may help format or proofread a plan that a person has written, within your rules, but it should not generate assessments or decide anything. If you are unsure whether a use is acceptable, it is not yet.
The mid-point check
At day 45 and again at day 75, the owner and registered manager review the error log. Categorise each entry: wrong fact, invented detail, wrong tone, data in the wrong place, rule broken. Three wrong-fact errors in one template means the template needs a tighter prompt or a different checker; a rule broken twice in one week means a retrain, not a new tool.
One scheduler's morning, before and after
Numbers on a sheet are easier to believe when you can picture them. Here is an illustrative Tuesday for the lead scheduler, first without and then with the Job 2 templates in place.
Before. At 7:40 a carer texts that she is unwell. The scheduler checks who is free, rings two carers, rewrites the visit list, then types five separate messages to clients and families explaining that a different carer is coming, each worded slightly differently and each taking three to five minutes, then takes two calls from families who found the wording unclear. By 9:30 the morning's planning work has not started. Time spent on notifications alone: about 40 minutes.
After. The scheduler still decides who covers each visit, because matching a carer to a client depends on knowledge of both. But the notification is a template with fields: "Hello [first name], [carer name] will be visiting at [time] instead of [usual carer]. They have been briefed on your routine. Call us on [office number] if you have any questions." The rota system or the scheduler's saved text fills the fields, and the scheduler checks the five messages before sending, about two minutes each. Time on notifications: about 10 minutes, and the clearer wording means fewer call-backs. The saving, around 30 minutes on a bad morning, is not from AI doing the work. It is from having already used AI once to write good templates, so the repeating message needs no thought.
That distinction is worth telling the team. Most of the gain in a first 90 days is from standardising and speeding messages that were being written from scratch, and the assistant's contribution is the first draft of the standard.
Teaching carers the rules without scaring them
Carers are on phones all day and many already use assistants at home. A five-minute conversation at a team meeting in month two does more than a policy PDF. Say three things: you may use the approved tools for the office tasks they are approved for, never put a client's name, address or health details into any AI tool, and if you are unsure, ask the registered manager. Give two real examples: a template written without names, and a case where a name in a prompt was the mistake. Then repeat the three rules in the supervision cycle for the next quarter. Keep the tone plain and non-punitive, because staff who fear blame stop mentioning mistakes, and the error log is your early-warning system.
A short line for the privacy notice and welcome pack
Clients and families are entitled to know how their information is used, and the explanation need not be long. An illustrative paragraph, to be checked against your own legal duties and your adviser:
How we use AI tools: Our office team sometimes uses AI tools to
help draft general letters, templates and policies. We do not put
your name or personal or health information into these tools.
Anything that affects your care is written and checked by
our care staff. If you have any questions, please speak to
[registered manager name] on [office number].Only publish this once it is true. If a pilot later involves personal data, update the notice and follow your data-protection obligations before it starts. Put the same paragraph in the welcome pack, so the two never disagree.
Referral sources and local relationships
One more office job that fits month two: following up with the people who send you clients. If you take referrals from social workers, hospital discharge teams or local community groups, a short, personal thank-you and update keeps the relationship warm. The assistant can draft the standard message; the owner or manager adds a line about the specific referral without any client detail. Illustrative example: "Thank you for the referral last week. We have started visits and will keep you updated on availability for further referrals. If it helps, our current capacity is [number] new packages." A line like that, sent monthly, costs ten minutes of review.
Budget at a glance
| Item | Illustrative monthly cost | Notes |
|---|---|---|
| Six office seats on a business chat plan | About $120 annual billing, $150 monthly billing | Check the vendor page; two-seat minimum applies |
| Free business-suite Copilot Chat | $0 if you already hold a qualifying business plan | Check data settings before use |
| Rota or care software AI add-on | Vendor quote | Ask what data the feature uses |
| Staff time on training and review | About 3 hours per person across the 90 days | The largest real cost |
| Contingency | $50 | For a second tool or a test |
Ninety days at roughly $200 a month is $600 of licences, plus about 18 staff-hours across six people. If the schedulers save ten hours a week, the hours repay themselves in the first month. Treat the savings as time handed back to the office, not as a reason to cut staff; the point is faster replies and fewer missed follow-ups.
The approved-tool list, filled in
| Tool | Approved for | Not approved for | Owner | Reviewed |
|---|---|---|---|---|
| Business chat assistant (office seats) | Adverts, templates, policy drafts, summaries of non-identifying text | Any client or carer detail | Administrator | Day 30, day 90 |
| Rota software's messaging feature | Standard visit-change messages with fields | Free-text client information until checked | Lead scheduler | Day 45 |
| Free personal chatbot accounts | Nothing for work | All work use | Owner | Rule restated monthly |
The administrator keeps this list and the error log in one shared sheet. Add a new tool only when the owner and registered manager have both approved it in writing.
A week-by-week rhythm that fits a busy office
- Monday, 20 minutes: owner and registered manager read last week's error log and the numbers sheet.
- Wednesday, 15 minutes: the pilot owner shows one real output and what they changed.
- Friday, 10 minutes: administrator updates the numbers sheet: hours, reply times, vacancies, errors.
That is 45 minutes a week of management time. It is the minimum that keeps the plan alive; without it, pilots drift and the rules decay.
The day-90 decision
On day 90 the owner reviews each pilot against its number and chooses one of four outcomes: keep and expand, keep as is, pause and fix, or stop. A pilot that hit its target with few errors expands. One that saved time but produced repeated errors gets tighter limits. One that did neither stops, and that is a good result because you learned it cheaply. Then pick the next two jobs: perhaps invoicing queries, training-matrix reminders or audit preparation, using the same pattern of owner, checker, measure and review date. The wider picture of what is possible is in the care-home version of this plan, which suits agencies that also run supported living.
Where the plan goes wrong in agencies
Starting with care records. The most common mistake is beginning with notes and care plans because they feel like the biggest time sink. They are also the riskiest. Start with adverts, replies and rota messages, then earn the right to go further.
Unowned side doors. A carer uses a free tool on a personal phone because nobody told them not to. The week-one question and a one-line rule prevent most of it.
Features switched on by default. A software update adds an AI summary of visit notes. Ask the vendor, and turn it off until you have reviewed it.
No checker. A template sent without review that quotes the wrong visit time to a family harms trust quickly. The rule that a named person reads every outgoing AI-drafted message costs minutes.
If you would rather map the first two jobs and the rules with someone alongside you, that is what my AI implementation consultation covers: get in touch. The three-month sequence above is complete enough to start without it.
What owners ask before starting an AI plan
Can carers use AI on their phones for visit notes?
Not until you have an approved tool, a rule on what goes in, and a person who checks the output. Visit notes are part of the care record, so wording must reflect what the carer observed. Start by using AI to tidy office-written text, then pilot carer-facing use with a small team.
Do I need a technology budget on day one?
Not a large one. Many agencies can start with the AI already in their email suite or rota software, or with a handful of business-plan seats for the office team. Spend after the first month shows which task is worth paying for, and set a monthly cap before you begin.
Should I tell clients and families I use AI?
Be honest and plain. Add a short line to your privacy notice and welcome pack saying how AI is used in the office, that a person checks anything that affects care, and who to ask. Follow your regulator's and data-protection requirements; take advice if you are unsure.
What if my care software already has AI features?
Treat them like any other tool: check what data they use, whether it is switched on by default, and who reviews the output. Add them to the approved list only after that check. Software vendors' AI features often arrive in updates without announcement.
How do I know the 90 days worked?
Compare three or four numbers from before and after: office hours spent on scheduling and chasing, time from enquiry to first reply, time to fill a vacancy, and the number of AI-related errors caught. If the numbers did not move, change the plan, not the target.
Further reads
- Home Care Agency Employee Handbook: Draft With AI, Then Review — Draft the staff handbook with AI, then review it.
- Non-Medical Home Care Leads: An Enquiry-Reply Routine Using AI — An enquiry-reply routine for non-medical home care.
- Home Care Agency Marketing Plan: Build One With AI and Local Data — Build a marketing plan with AI and local data.
- AI and Senior Living: Where a Small Care Home Should Start — Where a small care home should start with AI.
- How Can AI Help in Care Homes? Ten Back-Office Uses — Ten back-office uses of AI in care.
- HIPAA Compliance Policy for a Home Care Agency: Build It With ChatGPT — Draft a privacy and data policy with ChatGPT.
- Home Care Agency Welcome Packet: Build the Admission Pack With AI — Build the admission pack, including the AI disclosure line.
- Private Duty Home Care Referral Sources: Follow-Ups Drafted by AI — Referral follow-ups drafted by AI, reviewed by the owner.
- Adult Day Care Software for a Small Centre: Billing and AI Notes Checklist — A grouped checklist for a small adult day centre choosing software for attendance, transport, billing and care notes, with AI drafting kept under review.
- AI for Disability Services: Service Agreements and Incident Logs — Draft service agreements from your own template and structure incident logs with AI, while keeping participant details out of the prompt.
- AI for Home Health Agencies: Faster Replies to Hospital Referrals — A referral intake checklist, ready-to-send missing-information requests and a response-time log, with AI under a signed business associate agreement.
- AI in Care Homes UK: A Staff Policy Covering GDPR, Consent and CQC — A checklist-style AI policy for a care home: approved tools, data protection, consent, staff phones, oversight and the evidence to keep for CQC.
- CQC Mock Inspection Template: Run One and Summarise Findings With AI — Run a care home mock inspection by CQC key question, capture evidence, then use AI to summarise findings into a prioritised action plan.
- CQC Statement of Purpose Template: Draft Yours With AI and Check It — Draft a CQC statement of purpose with AI from facts you supply, then check every line against your registration, sign it off and set review dates.
- Care Home Tenders UK: AI Drafting From Your Evidence Library — Build an evidence library once, then let AI draft tender answers from it, with word limits, social value and a quality review that stops invented outcomes.
- Caregiver Onboarding Checklist: Pre-Start Checks and Induction With AI — A four-stage caregiver onboarding checklist, plus the prompts that turn it into a dated induction plan for each new starter.
- Caregiver Retention: Plan Check-Ins and Exit Interviews With AI — A check-in calendar, exit-interview questions and a safe way to find themes in them, with firm limits on monitoring staff.
- Domiciliary Care Rota Template: When Software or AI Fills the Gaps — A domiciliary care rota template with travel time and continuity rules, plus how AI can suggest gap cover while a coordinator makes the call.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI ChatGPT Business pricing page; Microsoft Learn pages on Copilot Chat licences and enterprise data protection (all read 4 November 2026). All agency figures are illustrative.