What it genuinely replaced
Answering the same question at 11pm. Chasing a reminder nobody read. Typing out availability for the fifth time that day. 10,793 patient messages answered without anyone opening a laptop.
Routiq Labs · you commented AUTO
Ten automations.
And the failure rate.
Read this first
I sell automation software for clinics, so I have an obvious stake in you thinking this works. The numbers below are from my own production database, including the ones that make me look bad. Read the failure rates before the wins.
The verdict
Where it actually pays
Automation is worth it where the job is repetitive, bounded and reversible. It is not worth it where being wrong costs you a patient.
What it did not replace
Judgement. Every one of the 103 escalations was a conversation a human needed to have, and the automation's only useful job was noticing that fast and getting out of the way.
The list
The ten
What each one replaced, what it has actually done, and the thing that bites. No row here is without a catch.
| Automation | Replaces | So far | The catch |
|---|---|---|---|
| 01Appointment reminders | A person reading tomorrow's list and texting each one | 7,344 queued | The highest volume and the lowest value. It only pays if the reply comes back into the same thread, otherwise you have automated a monologue. |
| 02Answer the reply | Someone watching an inbox all evening | 10,793 patient messages answered | This is the one that actually earns its keep, and it is also the one that needs the most guard rails. See the caveats. |
| 03Book from a text | Phone tag to find a time | 188 bookings from 1,040 calls | Writing to a real diary is the highest-risk automation on this list. Everything else can be wrong and recoverable. This one cannot. |
| 04Post-visit follow-up | The check-in nobody has time to send | 2,129 queued | Easy to make annoying. Once per visit, never on a schedule that ignores whether they replied last time. |
| 05Reactivation | A list of lapsed patients nobody calls | 301 queued | The one people get most excited about and the one most likely to breach marketing consent. Gate it on opt-in, not on last-visit date. |
| 06Missed call, text back | A voicemail nobody checks | 1,040 calls handled | Only useful if the text opens a thread you can actually answer. Otherwise you have replaced one dead end with another. |
| 07Voice message capture | A notepad by the phone | 523 completed calls | Transcription is the easy half. Deciding what is urgent is the half that still needs a person. |
| 08Escalate to a human | Nobody noticing the bot is stuck | 103 escalations | The most underrated automation here. If you build one thing, build the one that admits defeat quickly. |
| 09Daily recap to the owner | Logging in to find out what happened | Running for 3 accounts | Send it once a day at a fixed local hour. Anything more frequent gets filtered, and then you have automated something nobody reads. |
| 10Two-way diary sync | Double bookings and manual re-entry | 559,017 appointments synced | Not glamorous, and everything above depends on it. When the sync gaps, every other automation confidently acts on stale data. |
Both sides
What it is and isn't good at
Equal weight on purpose. A page that only lists the wins is an ad.
Good at
- +Repetitive work with a narrow set of correct answers.
- +Being awake. Most of the value is simply replying at 9pm on a Sunday.
- +Never getting bored on message four hundred.
- +Noticing it is out of its depth, if you build that in deliberately.
Bad at
- ×Anything where being wrong costs a relationship rather than a minute.
- ×Identity. Two patients share a phone number more often than you would think, and software guesses badly.
- ×Knowing when the rule should not apply. That is the whole job of a good receptionist.
- ×Fixing a broken process. It will just run the broken process faster.
Before you quote this
What I'd check
If you take one thing from this page, take this section rather than the table.
- 01
Only 969 of 10,106 queued messages actually sent
That is 9.6%. Most were skipped or disabled by guard rails working correctly, which is the point of guard rails. But 1,733 outright failed, and that is a real number, not rounding. Anyone showing you an automation deck without a failure rate is showing you a demo.
- 02
"Completed" is not "booked"
1,040 voice calls produced 523 completions and 188 bookings. If I quoted the completion number this page would look twice as good and mean half as much.
- 03
One vertical, one country, one PMS family
Allied health clinics in Australia on Cliniko and PracSuite. The shapes generalise, the numbers do not.
- 04
These are cumulative, not monthly
Every figure is a lifetime total across several accounts and several months. Divide before you compare it to anything.
- 05
I have not solved identity
2,863 patients in one account share a mobile with another patient. Phone-only matching is ambiguous for 15% of that book, and I found that out by filing a reply under the wrong person's name.
Method
Numbers pulled live from the production Postgres on 5 August 2026: patients, appointments, chatbot_messages, outbound_messages and voice_calls. No sampling, no projections, no modelled figures. Where something is not running yet it says so rather than being left out.
Nothing on this page wants your email. You already left a comment, which is more than enough. If you want the rest of it, the build notes are here: the nine jobs, building it properly and domain day.