Build1 publisher3 min readPublished
DenixLabs' dental voice agent picks up only the calls nobody at the desk answered
Across three clinics and a bit over 1,000 real conversations, that overflow rule makes the agent's own call log the first count of ring-outs these practices have had. It is also the only source for the 20 to 30 recovered hours a month.
The Engineer · Build desk
What happened
- Three dental clinics run the DenixLabs voice assistant in production on their existing phone numbers every day, and it has handled a bit over 1,000 real patient conversations.
- The assistant goes live in parallel with the front desk and only takes the calls nobody picked up, with a sit-down review two weeks after launch.
- Each clinic is getting roughly 20 to 30 extra booked hours a month, most of it from calls that used to ring out.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Every deal comes with bespoke modelling work the vendor absorbs: specialist schedules, per-service durations, an approved price list and approved phrasing, and there is nobody in IT at the clinic to hand it to.
- constraint Two vendors in this category cannot be compared on recovered hours, because each number comes out of that vendor's own overflow log and the vendor is the only one holding a count.
- precedent Selling automation into businesses like these means leading with the unmeasured cost. The measurement case has to be built before the software gets evaluated at all.
- capability Lunch, after 18:00 and Saturday morning become bookable windows without hiring anyone to sit on the phone during them.
The overflow rule does two jobs. The desk still gets every call first, so a wrong answer from the model cannot cost the clinic a call a human would have picked up [18]. And because the assistant only takes what nobody answered, its call log is a count of ring-outs [18]. The practice software never produced that number, because a missed call appears neither in the schedule nor in the monthly report [2].
That leaves the reported 20 to 30 extra booked hours a month per clinic as the assistant's own accounting of what it picked up [10].
For those hours to be money the clinic used to lose, two things have to hold. The handoff has to fire on callers who would otherwise be gone, and those callers must not have rung back later. The DenixLabs post on dev.to makes the case for one segment: "Someone with a toothache doesn't leave a message and wait. They hang up and dial the next clinic on their list" [3]. A patient rescheduling a cleaning at 19:15 is a weaker case, since that booking probably survives until morning.
Across three clinics the claim adds up to 60 to 90 extra booked hours a month [1]. Over a twenty-day working month it is one to one and a half extra chair hours a day per clinic [2]. The sample is roughly 340 conversations per clinic, though the post skips the time period [3].
The figures on offer are hours and conversations. Latency and word error rate never come up. The author says he talked to the clinic managers about "patients and phones, not prompts and latency" [21].
Per-clinic setup is a data modelling job. The assistant has each specialist's schedule, which services that specialist offers, and how long each service takes, because a cleaning and a root canal do not get the same slot [6]. Every price the assistant quotes comes off a list the clinic approved [7]. The owner listens to the voice and approves every phrase, and if they do not like it the launch does not happen [17]. All of that comes out of one call covering the schedule, the services and the durations, because these clinics have no IT department [16].
The money line is left to the buyer. "I don't put a revenue figure on that slide, on purpose. Every owner in the room knows their own hourly rate better than I do, and they do the maths in their head faster than I could present it," the author wrote [12]. Nearly a third of the twenty-minute talk went on the invisible leak instead [19].
At the three clinics the teams stayed and nobody lost a job [13]. The author wrote that he would put that ahead of the numbers next time: "In my experience the thing that blocks adoption is not the owner being sceptical, it's the staff being scared, and the owner picks up on that" [15].
What to watch
- Whether DenixLabs publishes containment and booking-error rates next to the recovered-hours figure as the clinic count grows.
- Native missed-call logging in practice-management software would give buyers a baseline independent of the vendor's agent.
- What the two-week post-launch review changes, in particular how long a call rings before the assistant takes it.