What an AI receptionist can and cannot do

Two weeks ago OpenAI held its first developer day and shipped GPTs, custom versions of ChatGPT that anyone can build by describing what the thing should do and uploading a few files. No code, no developer, about twenty minutes. The demos were good enough that we have had the same call from owners several times a week since: can this answer my phone.

Separately, and for a few months now, short videos have been going around of an AI voice picking up a call, talking to a caller like a person and booking an appointment with nobody in the room. Some of those demos are real. Some are heavily staged. Either way, the technology moved a long way this year, and the gap between a clean demo and a Tuesday at a busy front desk is still wide.

We have been testing this on our own lines rather than on client phone numbers. Here is where it honestly stands.

A custom GPT is not a receptionist

Worth separating, because the announcement has caused some confusion. A GPT is a chat window with your instructions and a few documents attached. It answers in text, inside ChatGPT, to somebody who is logged in and went looking for it. It does not pick up your phone. It does not know today’s schedule. It does not write to your calendar unless somebody wires that up, and it is not on your website unless you build that separately.

That does not make it useless. The GPTs we are actually getting value from are internal:

• Drafting review replies in the practice’s voice, for a human to send.
• Turning a messy set of notes into a plain-English explanation of a procedure for the website.
• Answering staff questions about your own policies, with the policy documents uploaded.
• Writing the first draft of a new patient email sequence.

One rule we have put on all of it: nothing about a specific patient or customer goes into a general-purpose chat tool. Not names, not charts, not photos. If the job needs that data, it needs a vendor agreement, and a chat window on somebody’s laptop is not one.

What an AI answering system genuinely does well today

Phone answering is a different product from a chat assistant, and the good ones are narrower than the demos suggest. Where they earn their money right now:

1. After hours and weekends. The 7pm caller currently gets voicemail and then calls the practice down the road. An AI that answers, takes the details and texts a booking link is competing with nothing.
2. Overflow when both lines are busy. Lunchtime and the first hour of the morning are where calls go missing in every account we look at.
3. The same six questions. Hours, address, parking, whether you take a given insurance, whether you see children, what to do in an emergency. These are scripted, checkable facts.
4. Taking a structured message. Name, callback number, reason, urgency, spelled back for confirmation. A message in a form beats a voicemail nobody transcribes.
5. Texting a link while the caller is still talking. This is the single highest-value trick, and it is the one most demos skip.

What it cannot do, and pretending otherwise costs you patients

• Insurance verification and pricing. “Do you take my plan” is not a yes or no question. An AI that guesses at coverage creates an angry conversation in the chair three weeks later.
• Anything clinical. A caller describing pain needs a human triaging them, today.
• An upset caller. The moment somebody is angry or frightened, a bot is an insult. Handoff has to be immediate and it has to work.
• Names, spellings and street addresses over a bad connection. This is where the demos quietly fail. Every system we have tested still mangles a percentage of them, and a wrong callback number is a lost patient.
• Interruption. People talk over the hold-music voice. The better systems handle it now. None of them handle it the way your front desk does.
• Knowing when it is out of its depth. The transfer rules have to be written by you, tested by you, and checked weekly.

There is also the compliance side, which for a dental practice or a clinic is not optional. The moment a caller says why they are calling, the system is handling health information, so the vendor needs a business associate agreement and the recordings need somewhere appropriate to live. Some states require every party on a call to consent to being recorded. And callers should be told they are speaking to an automated assistant, because the ones who work it out on their own do not call back.

Where we would start, in order

Most of the value in a local phone system does not need AI at all, and starting with the phone is starting at the most expensive end.

1. Missed-call text back first. If a call is missed, an automatic text goes out within seconds asking how the office can help. We set this up in GoHighLevel and it is the fastest payback of anything on this list.
2. Then the after-hours voicemail-to-text and booking link. Still no AI, still recovers calls.
3. Then web chat, with an AI drafting and a human sending. Low stakes, written not spoken, and you can read every answer before it goes out.
4. Then an AI voice agent on the after-hours forward only. One number, one window, nothing that touches daytime calls.
5. Only then daytime overflow, and only if the after-hours logs have been clean for a couple of months.

Measure it on booked appointments, not on “calls handled”. A system that handles two hundred calls and books nothing has done nothing.

It is also worth noticing that the company behind most of this spent the weekend in public turmoil over its own leadership. Nobody knows how that ends. It is a decent argument for building your front desk on plumbing you could point at a different vendor, rather than on one company’s roadmap.

What we are telling clients this month

• Turn on missed-call text back this week if it is not already running. It needs no AI and no budget conversation.
• Pull the call logs for the last month and count the calls that went unanswered, by hour. That is the actual size of the problem.
• Write the six questions your front desk answers most, with the correct answers, and keep them as a document. Any system you buy later needs it, and so does the new hire.
• If you want to test an AI answering agent, point it at the after-hours forward only, and listen to every recording for two weeks.
• Get a business associate agreement in place before any tool touches a patient call, and check your state’s rules on recording consent.
• Keep a person on daytime calls. The front desk is still your best salesperson.

We are building and testing these systems now rather than selling them on the strength of a demo video, and what we have learned so far lives on our AI agents page. For practices, the honest starting point is usually the boring automation around the phone rather than the phone itself, which is where most of the recovered appointments in our dental accounts have come from this year.