AI Phone Answering Versus a Human Answering Service

At 8:04am on a Monday, a three-surgery mixed practice in the West Midlands had eleven calls in the queue and two receptionists, one of whom was also cashing up Saturday’s card terminal. By 8:40 the queue had cleared. Twenty-three callers had cleared it themselves by hanging up. Nobody rang them back, because nobody knew who they were.

That is the problem an ai phone answering dental practice setup actually solves, and it is worth being precise about it, because the vendors are selling something broader than what the technology is good at. AI answering wins decisively on volume overflow. On complaints and on genuine pain calls, a trained outsourced human answering service is still better, and the gap is not closing as fast as the demos suggest.

The metric your phone system reports is the wrong one

Most cloud telephony dashboards (8x8, RingCentral, the Gamma and Vonage reseller packages a lot of UK practices sit on) report answer rate and average wait. Both flatter you. A call answered in 14 seconds that ends with “can you ring back after ten, we’re mad busy” counts as answered.

Three numbers are worth pulling instead, and you can get all three from a month of call logs:

• Abandonment by half-hour band. In practices we have looked at, Monday 08:00–09:00 routinely runs 28–40% abandoned while the daily average sits at 9–12%. The average hides the wall.
• Containment. The share of calls fully resolved without a human, with no callback generated. This is the number AI vendors quote and the number they define most generously.
• Escalation accuracy. Of the calls that should have reached a clinician or a senior person within the hour, how many did.

The last one is where the comparison actually gets decided.

Costing it honestly

Here is a worked example for a 3-surgery mixed NHS and private practice, roughly 6,500 active patients, 1,450 inbound calls per month, of which 38% (about 550) currently overflow to voicemail or abandonment.

Suppose you route only that overflow. Two options.

Moneypenny, alldayPA and Face for Business all price on call bundles rather than a flat per-call rate, so your quoted figure moves with volume and with how much spill you buy. The shape of the answer does not move: overflow handled by AI costs roughly a quarter to a third of overflow handled by outsourced humans, and the gap widens the more calls you throw at it, because AI cost is linear and human cost is linear plus a spill penalty.

For context, a part-time reception hire covering the 08:00–10:00 peak five days a week costs about £790/month at £11.44/hr plus NI and pension. AI overflow at £370 covers the whole month including Saturday and out-of-hours. That is the commercial case, and it is real.

Where AI genuinely wins: the 8:10am wall

Peak load is the one problem where AI is not merely cheaper but structurally better. Twenty simultaneous calls is not a hard problem for a voice agent; it is the same problem twenty times. A human answering service handles your spike by putting you in their queue, because their staffing is shared across hundreds of clients and Monday morning is everyone’s spike.

Routine, high-volume, low-stakes intents are also where containment is achievable: opening hours, address and parking, “am I registered here”, “how much is a hygiene appointment”, confirming or cancelling a booked appointment, taking a card payment link, capturing a cancellation-list request. In a well-integrated build against Dentally’s API or an SOE Exact middleware layer, those intents can reach 60–70% containment. Without write access to the diary, containment collapses to about 35%, because everything becomes a message for reception to action later, and you have paid for an expensive answerphone.

Recall is the other clean win, and it is outbound rather than inbound: an agent that rings 400 overdue patients across a fortnight, books what it can and flags the rest. That belongs in a wider view of front desk and recall automation, where the phone line is one channel alongside SMS and email rather than the whole system.

Where AI loses: the complaint call

A complaint arriving by phone has a legal clock on it. Under the NHS complaints regulations a practice must acknowledge within three working days, and a mishandled first sixty seconds is what turns a complaint into a GDC referral or a Dental Complaints Service file.

Complaint calls do not announce themselves with the word “complaint”. They start with “I was in on the fourteenth and I’ve had to go to another dentist”, or a long silence, or shouting. The caller is testing whether they are being taken seriously. What a good outsourced agent does, and a Moneypenny or Face for Business agent trained on your script genuinely will, is stop talking, take the details slowly, name the practice manager, and give a specific time for a call back. They also flag internally when a caller sounds like they are about to go to the CQC, which is not an intent a classifier is configured for.

Voice agents fail these calls in a characteristic way: they stay pleasant and keep offering to book an appointment. Pleasantness in the face of anger reads as dismissal. That is a reputational cost with a long tail and no line item.

Pain calls: the triage problem nobody demos

This is the one to be blunt about. Here is the kind of output that shows up in real call logs from under-configured deployments.

[09:14:02] Caller: my face has swollen up on the left side and it hurts to swallow
[09:14:05] Agent : I'm sorry to hear that. I can get you into an emergency slot.
                   The next one is Thursday at 2:40pm. Shall I hold that for you?
[09:14:19] Caller: yeah, alright
[09:14:21] Agent : Booked. You'll get a text confirmation shortly.

  duration: 0m 41s   transfers: 0   outcome: appointment_booked
  CONTAINED: true    csat_proxy: positive

Every metric on that call is green. The call is a spreading odontogenic infection with dysphagia, and the answer was a same-day assessment or A&E, not Thursday. The containment rate your vendor reports goes up when this happens.

Dental pain triage is hard for a language model not because the words are unusual but because severity lives in what the caller does not say, and in how they say it. Trismus gets described as “my jaw’s gone stiff”. A patient minimising symptoms because they are frightened of the cost sounds calm. An outsourced human with a one-page red-flag card asks the second question. The AI takes the first answer and books.

You can mitigate this, and you should, with a hard keyword and phrase interrupt that fires before any booking logic:

red_flag_interrupt:               # evaluated on every caller turn, pre-booking
  patterns:
    - swelling|swollen (face|cheek|eye|neck|under my (chin|tongue))
    - can'?t swallow|hurts to swallow|drooling
    - can'?t open my mouth|jaw (is )?stuck|locked
    - temperature|fever|shivering|shaking
    - bleeding (won'?t|hasn'?t) stopped|since (yesterday|last night)
    - knocked out|knocked my tooth out|avulsed
  action: transfer
  target: human_service
  priority: urgent
  whisper: "Possible spreading infection or trauma. Do not book. Escalate."
  on_transfer_fail: play(urgent_ooh_message) + sms(practice_manager)

Note the last line. A red flag that escalates into an unstaffed queue is not an escalation.

The split that actually works

Route by intent and time of day rather than picking one supplier for everything.

Send to the AI: all calls in the 08:00–09:30 and 13:00–14:00 bands once queue depth passes three, every call after 18:00 and at weekends, and any call whose opening intent is hours, location, price list, confirm, cancel, or cancellation list. Give it diary write access or do not bother.

Keep a human service on: anything the red-flag interrupt catches, anything where the caller asks twice for a person, any caller flagged in the PMS as a live complaint or safeguarding record, and the whole of the first ring on out-of-hours emergency numbers. Budget for 15–20% of your routed volume landing here. At £2.20 a call that is the cheapest clinical risk control you will buy this year.

One practical constraint before you sign anything: the AI vendor is processing special category health data, so you need a signed DPA, UK or EU hosting, a retention period you chose rather than inherited, and a DPIA on file. Ask where transcripts sit and for how long. A surprising number of voice-AI startups will answer that question with a pause.

Start by pulling last month’s call detail records and sorting by abandoned-by-half-hour. Listen to thirty of the abandoned calls’ callbacks if you have them. You will know within an afternoon whether your problem is volume, in which case AI fixes it cheaply, or whether your problem is that difficult calls are being handled badly by whoever picks up, in which case no phone system of either kind is the thing to buy first.