How a Home Care Agency Cut Scheduling Time 87% with a $20/month Claude Subscription

Small businesses usually assume a scheduling problem calls for new software or another hire. For one home care agency in Orange County, the fix was a Claude subscription at $20 a month, an audit, and two automations built around software they already had.

A small office team keeps the agency running: one person owns scheduling and complaints, one owns HR and caregiver onboarding, one handles client care and referrals, and the owners cover everything else. Every one of them was spending a large part of the week copying information from one screen into another.

What We Evaluated First

We started with an audit and on-site training, not with a tool. The audit asked three questions: where the hours go, what the agency can spend, and what the data rules have to be.

The workflows

The two heaviest workflows were obvious within a day.

• Filling open shifts. The schedule lives in a HIPAA-compliant care scheduling platform that is decades old. It has no API and no integrations. To fill an open shift, the scheduler had to cross-reference the open visit against every caregiver's weekly availability, years of client notes (preferred caregivers, families who asked that a caregiver not return, language needs, safety protocols), and a separate caregiver roster in Google Sheets with driving limits, allergies, CPR dates, and second languages. She timed it herself: about 5 minutes per household, every week, before a single call was made.
• HR email. The HR inbox sends the same kinds of email over and over: job offers with orientation dates, orientation confirmations, day-before reminders, call-out reminders, attendance and tardiness warnings, time-off approvals, and separation paperwork. Each one meant checking an applicant tracker, the caregiver roster, past emails, and the calendar, then rewriting the same template by hand.

None of this called for judgment the software could not support. All of it needed a person, because the scheduling platform could show the schedule but could not reason across it.

The budget

The agency had no appetite for a second system of record or a platform migration. The constraint was simple: keep the scheduling platform, keep Google Workspace, and keep the recurring cost close to zero. A Claude plan at $20 a month per seat fit. A new scheduling suite or another office hire did not.

The compliance needs

Whether a home care agency is a HIPAA covered entity turns on the definition in 45 CFR 160.103, and non-medical agencies often fall outside it. We did not build on that hope. A standard paid Claude plan does not come with a Business Associate Agreement, so we wrote the rules first:

• Each job gets only the exports it needs, and staff export them by hand. Claude downloads nothing.
• The scheduling platform is read-only to Claude. No logins, no typing, no edits.
• Nothing sends itself. Every output is a proposal or a draft.
• No diagnosis, address, date of birth, or case number goes into a drafted message. Any slip is logged as an incident.

We also wrote down what we would and would not claim, and set a two-week baseline before switching anything on. That plan is why the numbers below are measured, not estimated.

Automation 1: Open-Shift Scheduling

Claude Cowork lets a team save instructions and files in a project and run multi-step tasks against them. We built a project that turns one week of open shifts into a single checklist the scheduler can work through.

What goes in

The scheduler names the Monday that starts the week and uploads three exports from the scheduling platform: the weekly calendar by staff (as Excel, with full caregiver names turned on), the weekly calendar by client (Excel), and the client notes going back as far as the platform keeps them (PDF). Claude reads the caregiver roster directly from Google Sheets through the Drive connector.

Claude also opens the platform's calendar page in the Claude in Chrome extension, after the scheduler signs in, to read the open shifts in the order she sees them. The rules for the browser are strict, because a stray drag on that calendar reschedules a real visit:

• Never log in. If a login page appears, stop and ask.
• Never type into the page.
• Read only. Never drag, reassign, save, or delete.
• If the page cannot be reached, fall back to the export order and say the ordering is unverified.

What Claude does with it

For each open shift, in calendar order, Claude works through the same checks the scheduler used to do in her head:

1. Hard exclusions. Every note saying a caregiver must not return to that client, whether the family or the caregiver asked. These are absolute.
2. Stated preferences. Preferred caregivers, language needs, two-caregiver recommendations, and safety protocols. If a client requires a caregiver of a particular gender, Claude proposes no name at all. Gender is not recorded anywhere in the files, and the instructions forbid guessing it from a name.
3. Who is actually free. Caregivers whose availability covers the full visit with no overlap that day. Labels it cannot parse go on a verify-separately list, never into a recommendation.
4. Competing shifts. When several open shifts draw on the same small pool on one day, nobody is proposed twice, and a short day is called out plainly.
5. Roster rules. Driving limits, preferred-clients-only flags, and weekly hour caps are printed beside the recommendation so the scheduler can check them.

Every note in the output has to trace back to a line in the uploaded files. If it is not in the files, it does not go in.

What comes out

One checklist for the week, grouped by day. Each line shows the client and time slot, the recommended caregiver in bold, backups, and the notes that justify the choice. Duplicate or suspect shifts are marked as a hold instead of filled. A counter shows how many shifts are left. Checking a box changes nothing in the scheduling platform; the scheduler still makes every assignment herself.

Result: scheduling time dropped from about 5 minutes per household to 39 seconds, an 87 percent cut.

Automation 2: HR Email Drafting

The second project lives in the HR mailbox. HR types a one-line request ("send an offer to [name]", "confirm orientation for [name]", "draft a verbal warning for [name]"), and Claude leaves a draft in Gmail that is ready to review.

How it works

Claude reads four sources in full on every run: the applicant tracker, the caregiver roster, a departures sheet, and the HR mailbox itself. From those it works out each person's status (offer sent, orientation scheduled, no response, hired, and so on), picks the matching template, and fills in the details.

• Seventeen fixed templates. Offers, orientation confirmations and reminders, call-out reminders, a full warning ladder, time-off and availability approvals, resignation acceptance, and separation paperwork. Claude uses the agency's wording verbatim. If a request does not fit a template, it says so and stops.
• Drafts only. Claude never sends, forwards, or schedules a message, and never adds a signature. HR adds their own signature and presses send.
• Never guess. If an email address, rate, or date is not in a source, the placeholder stays bracketed and gets flagged.
• Thread-aware. Claude checks for an existing draft before creating another, and replies inside the right thread.
• The warning ladder comes from the mailbox. Before drafting a warning, Claude builds a history of what the caregiver has already been sent on that topic and drafts the next level. If HR asks for a level that skips a step, Claude asks first.

The morning run

A scheduled task runs every morning at 7:00. It handles departures first (calendar reminders for final paperwork and paychecks, plus ticking the departures sheet only when the mailbox shows the email was actually sent), then drafts tomorrow's orientation reminders and any confirmations applicants have replied to, then drafts offer letters with two orientation dates drawn from the agency's standard windows. Warnings, resignations, and time off are never drafted on a schedule. They wait for a person to ask.

No part of this screens, ranks, or scores applicants or employees. California's automated-decision rules under FEHA are one reason, and keeping hiring judgment with people is the other.

Result: time spent on the HR emailing process dropped by 40 percent.

Writing the Rules Before Switching Anything On

Both projects run on written instructions that say exactly what Claude may read, what it may write, and when it must stop and ask. The agency also agreed to a short quality check: three drafted outputs a week, each scored on four points: accurate against the source record, on policy, complete, and sendable after a light edit. It is the same approach we describe in how to write an AI usage policy for a small business. People move faster when the answer is written down.

For the broader question of what belongs in a general-purpose assistant, is ChatGPT safe for business data covers the reasoning we applied here.

The Result

The audit's estimate of $5,000 to $7,000 a year in recovered office time is still labeled an estimate. The 90-day measurement plan will replace it with a measured range, every input shown, so the agency can check the arithmetic. If the numbers do not move, the plan says that gets reported as a "not yet," and we revise the jobs before the agency spends anything more.

The next workflow on the list is customer service: the referral reports the client care team builds by hand each week.

What to Take From This

If your office runs on an old system that will not integrate with anything, you do not have to replace it to automate around it. The steps that worked here transfer to most appointment-driven businesses:

1. Time the work before choosing a tool.
2. Keep the existing system of record. Export from it, and treat it as read-only.
3. Write the data rules first, and decide what you will measure before you build.
4. Draft everything, send nothing automatically, and keep sensitive decisions with people.

That sequence is the substance of our assessment work, and for teams that want the automations kept running as software and rules change, On-Going Support covers it. If you are weighing this against hiring, AI consultant vs. hiring an in-house AI engineer lays out the comparison.