A voice receptionist for a clinic: the hard part is taking an email over the phone
Voice agents hold a conversation well. Where they break is simpler than that: writing down an email by ear, over a noisy line. Here is the working approach, on ElevenLabs, with domain checking and booking into Outlook.
People call a clinic while the receptionist is with a patient in the chair. The call rings out, the caller dials the next clinic in the search results. This is close to the ideal job for a voice agent: it always answers, and it collects exactly the details someone needs in order to call back.
The conversation itself is the easy part now. An agent speaks naturally, you can interrupt it, it does not read lists at you. Where it trips is somewhere else, and that place is worth understanding before you commission a system like this.
Where it actually breaks
Taking an email over the phone.
Someone dictates an address by ear. The line is noisy, and sometimes it is a landline. “M” and “N” sound the same, so do “S” and “F”. People swallow dots and forget to mention digits. If the agent just asks for the whole address and writes down whatever it heard, it will get it wrong often, and you will find out when the confirmation email never arrives.
The phone number comes in automatically. The name is easy. All of the risk is concentrated in one field.
Ask for the domain first
The first thing that changes the picture: do not walk through the address left to right, the way it is written.
The agent asks which email provider the caller uses. In practice there are not many answers, and most land in a short list: gmail.com, outlook.com, hotmail.com, yahoo.com, icloud.com. If one of those comes up, the agent simply confirms it: “So that’s at gmail dot com?” If it is a work address, the domain is spelled out and read back.
That closes the longest part of the address with a single question, and leaves only the part that genuinely has to be dictated.
The name letter by letter, confirmed on the way
The second half is collected letter by letter, but in chunks rather than all at once. Every three or four letters the agent reads them back using the phonetic alphabet: “So far I have B as in Bravo, E as in Echo, N as in November — is that right?”
Dots, dashes, underscores and digits get their own explicit question. They are inaudible in a stream of speech, and they are exactly what breaks addresses like john.smith2@.
It is a small move, but it means a correction costs three letters, not the whole address.
Check the domain for real
The third layer is the most underrated one. The assembled address goes out to a check before the agent reads it back to the caller.
The check does two things. It compares the shape of the address against a regular expression. And it asks DNS whether the domain has MX records at all — that is, whether it accepts mail.
The second part catches a real class of error: a misheard “gmaul.com” looks plausible and passes any format check, but it cannot receive mail. The agent gets back a reply saying the domain does not accept email, and asks again about the domain specifically rather than the whole address.
Twice, then stop pushing
The fourth decision matters most to the business, and it is the one usually left out.
The agent asks again at most twice. If the address still does not come together, it does not loop and it does not invent. It says: “That’s fine, I’ll book you in and our team will confirm your email by message or by phone,” marks the record as low confidence, and carries on.
That mark then does work downstream: no calendar invite is sent to the caller, the row in the sheet is flagged, and a receptionist can see which bookings need confirming by hand. The call is not lost, which was the whole point.
Booking into the calendar
Once the name, email, phone and time are collected, the agent repeats everything back in one sentence and asks for confirmation. Then the booking goes into the Outlook calendar.
Before the event is created, the system checks whether the slot is free. If it is taken, the agent does not apologise vaguely — it offers two specific nearby times. If it is free, a half-hour event is created with the name, phone and email in the body, and the same booking is appended as a row in a sheet that works as a simple CRM.
What to take from this, even if you build it yourself
Three things that do not depend on the tools you pick.
Ask for the domain before the name. Confirm three or four letters at a time, not the whole address. And decide in advance what the agent does when it fails: a rule of “twice, then book it with a flag” saves more calls than any improvement in speech recognition.
If you want something like this for your own clinic or salon, get in touch and we can look at it against your actual call flow.