What the agent does
This is for a practice that lives on its appointment book: a dental or physio clinic, a salon, a veterinary practice, a tutoring centre. A missed slot costs the whole fee, and reminder texts are easy to ignore. Every evening the agent reads tomorrow's appointments from the practice's Google Calendar, calls each client who agreed to reminder calls through LumisReach, and writes the front desk a list for the morning. The run has no repository and names two connections: Google Calendar and LumisReach.
It lists tomorrow's events on the bookings calendar with the Google Calendar API's events.list, from midnight to midnight in the practice's time zone with singleEvents set, so a recurring appointment comes back as tomorrow's one occurrence. The practice's booking tool writes each client's phone number and a reminder-consent line into the event description; the agent dials only events whose description has both, and lists the rest as skipped with the reason, so a client with no consent on file is never called.
For each appointment it starts one call with start_ai_phone_call: the client's number in to_numbers, a goal to confirm the appointment at its time or take a preferred new time, instructions that name the practice, the service and the clinician, and an opening_message that says it is the practice's AI assistant calling about tomorrow's appointment. One call per client keeps each call's details correct; LumisReach takes more than one number in a job, but those calls share one goal. It sets retry_attempts to 1 with retry_delay_minutes of 20, max_call_duration_seconds to 180, and a voicemail that gives the time and the practice's number.
Then it polls get_quick_action_status for each job until every target has finished, and sorts each client by the target's status, goal_achieved and summary: confirmed, wants to reschedule (with the times they offered, in their own words from the get_call transcript), cancelled, voicemail left, no answer, failed, or opted out. It changes nothing in the calendar. Moving or cancelling an appointment stays with the front desk, which knows what else is booked.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "Confirm tomorrow's appointments by phone. Practice: {practice}, time zone {time_zone}, callback number {callback_number}.\n\n1. With the Google Calendar connection, call events.list on the calendar '{calendar_id}' with timeMin and timeMax covering tomorrow in {time_zone}, singleEvents true, orderBy startTime. Skip all-day events and events marked cancelled.\n2. For each event, read the client's name, the service and the clinician from the title, and from the description the line 'Phone:' and the line 'Reminder calls: yes'. Call only events with both lines. List every other event as skipped with the reason (no phone, no consent).\n3. For each client to call, start one LumisReach job with start_ai_phone_call: to_numbers [the phone], goal 'Confirm the appointment for {service} with {clinician} tomorrow at {time}, or take the client's preferred new day and time if they cannot make it', instructions 'You are the AI assistant for {practice}. Be brief. Do not give medical or pricing advice. If they want to reschedule, ask for two times that suit them and say the front desk will confirm in the morning. If they ask not to be called, apologise and end the call.', opening_message 'Hi, this is the AI assistant for {practice}, calling about your appointment tomorrow at {time}. Is this {client_first_name}?', retry_attempts 1, retry_delay_minutes 20, max_call_duration_seconds 180, voicemail_enabled true, voicemail_message 'This is {practice} with a reminder of your appointment tomorrow at {time}. If you need to change it, call us at {callback_number}.'\n4. Poll get_quick_action_status for each job until its target is no longer pending or in_progress. For answered calls, read the call record, with its transcript, using get_call on last_call_id.\n5. Write confirmations.md: a count per outcome at the top, then one table row per appointment sorted by time: time, client, service, outcome (confirmed, reschedule, cancelled, voicemail, no answer, failed, opted out, skipped), the client's own words about a new time if any, and the next step for the front desk. Change nothing in Google Calendar and call no number that is not in an event description.",
"connections": [
"google-calendar",
"lumisreach"
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
What comes back is confirmations.md, an artifact on the session, and the run's closing message as the session's result, which gives the counts in a few lines. The table is the front desk's morning list: the confirmed rows need nothing, the reschedule rows quote what the client said they could do, and the voicemail and no-answer rows are the people to try again by hand or by text before they are due in.
The session's transcript at GET /o/{org}/sessions/{id}/messages records the calendar read, each start_ai_phone_call body, the status polls and every call record, so a client who says they cancelled on the phone and still got billed for a missed slot can be answered from what the call actually says.
What a person checks before acting: the reschedule requests against what is open in the book, and any row marked failed, which usually means a number in the event description is wrong or out of date. Fix the number in the booking tool, not in the report, so the next evening's run reads the right one.
Run it on every weekday evening
In the console, a Project automation with a schedule trigger runs this prompt at 17:00 in the practice's time zone, a cron such as 0 17 * * 1-5 with timezone set, and Sunday at 17:00 if the practice opens on Mondays. Running at five leaves time for the 20-minute retries to finish before eight in the evening, inside the federal 8:00 to 21:00 window for the client's time zone.
A booking platform that serves many practices can start the runs from its own scheduler instead: one POST /o/{org}/sessions per practice per evening, with a request_id that is a UUID v5 of the practice id and the date, so a retried job does not call anyone twice. With a webhook on the API key, Rush POSTs run.completed when each run ends, and the platform shows that practice's list in its own dashboard.
The use case this job follows, with its automation and trigger: Produce a report or file, no repo.
Why run it on Rush?
Confirmation calls are the front desk's least favourite hour of the day, and the one most often skipped when the clinic runs late. Texts get read by the people who were coming anyway. A call that asks a direct question, and takes a new time when the answer is no, gives the practice a day's notice to fill the slot from its waitlist. Running it as a scheduled agent means it happens every evening, with the same script, and the list is ready before anyone arrives.
Each run has its own computer and a record of every call it placed and what was said, which matters when a practice is asked to show that a client was reminded. The run reaches only the calendar and the calling service it names, and reads the calendar without changing it.
LumisReach bills the calls to the practice's own LumisReach account: $0.20 for each call placed, and each retry is a new call, plus $0.20 a minute billed by the second (its docs). An answered one-minute call costs $0.40, and a client who never answers costs $0.40 with one retry, so each appointment costs at least $0.20, plus $0.20 for each minute someone talks. Rush bills the run's compute and model usage from your credits, and the session's total_cost_usd reads it back.
Frequently asked questions
Do clients need to agree to these calls?
Yes. Get consent to reminder calls when the client books and write it into the event, which is why the prompt dials only events with a 'Reminder calls: yes' line. Under the TCPA, and since February 2024 the FCC treats AI-generated voices as artificial voices, a call like this needs the client's prior consent. LumisReach's site says every number is checked against federal, state and internal Do Not Call lists before dialing (lumisreach.com); its API docs do not repeat it, so the consent check here does not lean on it. This is not legal advice; check your own case, and any health privacy rules that apply, with counsel.
Will it move appointments in my calendar?
No. The prompt reads the calendar and changes nothing; a client who wants another time is listed with the times they offered, and the front desk moves the booking. Letting an agent move bookings on its own risks double-booking a clinician, which a person checking the book avoids.
What does the client hear?
An AI voice that says at the start it is the practice's assistant calling about tomorrow's appointment, gives the time, and asks to confirm. If nobody answers, it tries once more 20 minutes later and leaves a voicemail with the time and your number. Change the wording in opening_message, instructions and voicemail_message.
How do I connect LumisReach and Google Calendar?
Make a LumisReach API key under Settings > Developer API in LumisReach (app.lumisreach.com/settings/api), with a card under Settings > Billing, and paste it in Rush under Settings → Connections → LumisReach. Connect Google Calendar on the same page with the Google account that holds the bookings calendar. The request names them as google-calendar and lumisreach.
More LumisReach jobs
More jobs like this: Produce a report or file, no repo
- Write an SEO Content Brief With Semrush Data (Semrush)
- Find Content Gaps With Ahrefs (Ahrefs)
- Write a Linear Cycle Report (Linear)
- Get a Daily Sales Summary From Square (Square)