The weekly transcript sample
You'll be able to: run a 20-minute weekly check of what the voice agent, the website chat and staff actually said to new patients.
Approving templates and the agent's instructions at setup is not the end of supervision. Supervision is looking at what went out. A practice that never reads its own transcripts finds out what its voice agent says the day a patient complains.
A weekly sample is enough for a practice of up to ten users. Twenty minutes, the same day every week, by the owner, the practice manager or a clinician they name.
The sample (a starting point; adjust to your volume):
- 3 AI voice calls, at least one in Spanish
- 3 website chats, at least one in Spanish
- 3 staff threads, at least one with a handoff to a clinician
- 1 thread where someone other than the patient called or wrote
- Every urgent or crisis trigger from the week
For each, check:
- Did the voice agent say it was an AI assistant and that the call was recorded, before any question, in the caller's language? Did the chat say it was automated?
- Did anyone, the agent, the chat or a person, give clinical advice, promise a result, or say a visit was covered?
- Did anyone confirm that someone was a patient to a caller who was not them?
- Were urgent and crisis triggers handled as set in Lesson 1.4: sequences stopped, the right person alerted, and the alert free of health details?
- Did the person get a human next step within the practice's standard?
- Did the Spanish read as written in Spanish?
- Did any message or preview name a service or condition?
Write down what you found, even when it is nothing. A log of weeks with nothing found is evidence that the practice supervises. A finding gets a fix: coach a staff member, change an approved reply, or adjust the agent's instructions through your named contact or the next quarterly review. Note the fix and the date.
Look for patterns, not just mistakes. Three callers in one week asking whether you offer evening hours is not an error, but it is a finding: a better greeting, a new saved reply, or a question for the quarterly review. The transcripts are the best record the practice has of what new patients want to know.
Handle transcripts as sensitive records. The reviewer opens them inside TheraCRM, does not copy them into email or a chat app, and writes findings without names or health details: "Spanish call, Tuesday, agent answered a question about downtime it should have declined." That is enough to fix it.
Do not skip the sample in a busy week. The weeks the practice is busiest are the weeks shortcuts happen.
How it differs by practice type: at a therapy practice, the reviewer should be a clinician or have one review every crisis trigger. At a dental or med spa practice, the most common finding is usually prices or coverage said too loosely. At a medical practice, it is often referral logistics given to a caller who was not the patient.

Do it in TheraCRM
- Create a recurring weekly task for the reviewer, on the same day each week.
- Use the conversation filters to pull AI calls, website chats and staff threads from the last 7 days.
- Keep a transcript review log (a shared document set up at guided setup): date, item, finding, fix, with no names or health details.
- Send any finding about the voice agent or the chat to your named support contact.
Ask the Guide
Paste this into the AI guide, without patient names or health details:
"Turn these seven checks into a one-page weekly review form we can print, with a column for the finding and one for the fix: [paste the checks]."