Skip to content

What we stopped letting the phone agent do

After more than a year of AI agents answering real calls: the work they handle well, the jobs we took back off them, and the numbers we judge them on.

The first agents we set up were answering after-hours calls in the autumn of 2024. Enough time has passed that we should stop calling it an experiment. Across the offices we look after, an agent now covers the overflow during the day, the whole weekend and every call between six at night and seven in the morning.

February is a good month to write down what we have learned. In dental the January rush is over and the schedule has holes in it. In home services a cold snap fills the phone for three days and empties it for ten. Both are the conditions that show what your answering setup is really doing, because both produce calls nobody is standing next to.

So: what it handles well, what we have taken away from it, and what we look at to decide whether it is worth keeping.

What it handles well

The honest summary is that it is very good at the calls a busy front desk was already handling badly, and no better than a person at anything else.

  • The call that would have gone to voicemail. After hours, during lunch, while both people are with a customer. The competition here is a beep and a recording, and it wins easily.
  • The second and third call at once. A storm week or a promo week produces calls that arrive on top of each other. Nothing gets a busy tone now.
  • Facts. Hours, address, which entrance, parking, whether the practice is taking new patients, whether a service is offered at all.
  • A complete message. Name, callback number spelled back and confirmed, reason, urgency. A message a person can act on in the morning without calling back to ask what it was about.
  • Booking into an open slot on the same calendar the front desk uses, with the same buffer rules, so nobody is double booked.
  • Consistency at ten at night. It asks the same questions in the same order on the four hundredth call as on the first.

What we stopped letting it do

This is the more useful list, because every item on it came from reading a transcript and wincing.

  1. Quoting prices. We started with ranges. Callers heard the bottom of the range and treated it as the price, and the conversation at the desk afterwards was worse than the one that never happened. Now it says the office will confirm cost after they have seen the situation, and it books the visit.
  2. Anything about coverage. It may read the list of plans the practice is in network with, by name. It does not say whether a plan covers a procedure, because it does not know and the caller will hold you to it.
  3. Anything clinical. Symptoms, pain, what to do with a broken tooth, whether something can wait. Straight to the emergency instruction and the on-call number. We wrote this fence before the first script and we have only made it higher.
  4. Cancellations. It may reschedule, because a moved appointment is a kept appointment. A cancellation goes to a person, because a person can often save it, and because an agent that cancels smoothly makes cancelling easier than it should be.
  5. Outbound anything except confirmations. We tried recall calls. They work technically and they feel wrong on the receiving end, and one complaint about a robot calling costs more than the recall was worth.
  6. Continuing with an unhappy caller. The first sign of frustration is a handoff now, not a retry. Repeating a question to someone who is already annoyed is the single fastest way to lose a patient.
  7. Ending a call with no record. Every call leaves a transcript and a recording on the contact, even the ones that were a wrong number. The gaps in the log were where the problems were hiding.

How we judge one

Vendor dashboards report on the agent. We report on the phone line, which is a different thing. The numbers we keep per office, compared against that office’s own previous quarter and nothing else:

  • Answer rate on the whole line, including the calls that never reached the agent because a forward was set up wrong. This one finds more lost work than anything else on the list.
  • Completion rate: the share of calls that ended with either an appointment on the calendar or a message complete enough to act on.
  • Handoff rate, and whether each handoff was correct. Too low is worse than too high. An agent that never hands off is an agent that is guessing.
  • Repeat callers inside a day. Somebody calling back within a few hours is usually telling you the first call failed, and it will not show up anywhere else.
  • Booked work traced back to the call in the CRM, all the way to the appointment being kept, not just made.
  • Cost per booked appointment with the minutes included. Voice is metered. An agent that chats pleasantly for four minutes before booking costs more than one that books in ninety seconds.

We have no target number to hand you for any of these. The only comparison that means anything is the same office against itself, before and after, busy week against quiet one.

The habit that makes it work

Every office that is happy with its agent has someone reading transcripts. Every office that is unhappy stopped reading them in month two.

For the first two weeks we read all of them. After that, a weekly sample plus three automatic buckets: every handoff, every call that ran long, and every call where the caller had to repeat themselves. That last bucket is where the script fixes come from. Fifteen minutes a week, and it is the difference between an agent that gets better and one that quietly drifts as your hours, staff and services change around it.

What to do this month

  • Call your own main number after hours, from a number the office does not know, and try to book something. Then call it at eleven on a Tuesday morning and let it ring.
  • Check every forward: main line, after hours, overflow, the old number on the van or the sign. A broken forward is the most common cause of a line that “does not get calls”.
  • Read last week’s transcripts and mark every place the agent answered something it should have handed over. Put those in the guardrails, not in the script.
  • Update the knowledge with anything that changed since the new year: hours, a new service, a plan you no longer take, a staff name it still uses.
  • Pull completion rate and repeat callers for January and put them next to December. Then leave them alone for a quarter.
  • Set a usage alert on voice minutes before the spring rush rather than after the invoice.

An agent is not a front desk and we have stopped selling it as one. It is coverage for the hours nobody is there and the calls nobody can get to, and it is worth what it costs when someone is reading what it said. That is the way we set up AI answering, and why we keep it inside the same CRM as the calendar and the follow-up, which matters most in dental where the call and the record have to stay together.

Written February 6, 2026, and kept as written. Platforms, features and policies mentioned here are described as they stood at the time.

Today, for your business

This is what we do for clients: AI Agents.

The platforms have moved on since this was written. We will show you where your business stands right now and what to do first.