An AI medical receptionist answers your practice's inbound calls, books and reschedules appointments directly in your EHR, runs intake, and escalates anything clinical to a human. The useful question is not whether the technology works, which it does, but which calls you are willing to hand over, what the real cost looks like at your call volume, and whether the vendor can legally touch patient data at all.

Most buyer guides on this topic are published by the vendors themselves, and the numbers in them are rarely sourced. This one cites every figure, and links to the primary source so you can check it.

Key Takeaways

• Per-minute AI pricing undercuts a salary at almost any realistic call volume. At 5,000 call minutes a month, a busy multi-provider clinic, published rates run $350 to $1,550 a month against a $3,168 median monthly receptionist salary. The gap is real, but per-minute rates exclude platform and integration costs.
  
  

• Your AI vendor is a HIPAA business associate, and HHS says so explicitly. HHS names a third-party AI chatbot handling appointment scheduling as a business associate. No signed BAA means no lawful disclosure of PHI.
  
  

• EHR integration is where most deployments actually fail. "Integrates with Epic" can mean real-time bidirectional booking or a nightly one-way export. Those are different products.
  
  

• The escalation path matters more than the voice quality. What the system does when it does not understand an emergency, an upset patient or a clinical question determines whether it is safe to deploy.

What an AI medical receptionist actually does

An AI medical receptionist is software that answers calls and messages without a human operator. It uses speech recognition and a language model to hold a conversation, then takes an action in your practice systems: booking a slot, sending an intake form, logging a refill request, or routing the call to staff.

What it handles well

The tasks that work reliably share a shape: high volume, repetitive, and verifiable against a system of record.

• Appointment booking, rescheduling and cancellation. The highest-volume reason patients call, and the easiest to check against live EHR availability.
  
  

• Appointment reminders and confirmations. Outbound, scripted, and measurable against no-show rates.
  
  

• New patient intake. Capturing demographics, reason for visit and insurance details before the visit.
  
  

• Prescription refill requests. Collected and routed, not approved. Refills sit alongside the other back-office tasks AI is quietly absorbing.
  
  

• Routine FAQs. Hours, location, parking, what to bring, accepted insurance.
  
  

• Overflow and after-hours coverage. Answering calls that currently ring out when every line is busy or the office is closed. The alternatives here are a human virtual receptionist or a traditional medical answering service.

Two of those are worth separating out. Overflow is usually where practices see the fastest return, because the baseline is not "a human did this well" but "nobody answered at all." A call that was previously abandoned has no cost to beat.

What it still cannot do

Be clear-eyed about the boundary, because vendors blur it:

• Clinical judgment. Deciding whether symptoms are urgent is triage, not reception. If a vendor markets both, confirm which product you are buying.
  
  

• Complex insurance and billing disputes. Verification of coverage is increasingly automated; arguing a denial is not.
  
  

• Anything requiring the patient to be de-escalated. Distressed and angry callers need a person, and the system should route them to one quickly.
  
  

• Judgement calls on exceptions. Double-booking a long-standing patient because the physician would want it is exactly the kind of decision a rules engine gets wrong.

The practical model is not replacement. U.S. Bureau of Labor Statistics projections show receptionist employment declining 2% between 2025 and 2035, a slow contraction across the whole economy, not a collapse. In practice, AI absorbs call volume so existing staff stop being interrupted.

What an AI medical receptionist costs

Start with the number you already pay. The median U.S. receptionist earns $38,010 a year, or $18.27 an hour, according to the Bureau of Labor Statistics (May 2025 data, across 947,500 jobs). That is $3,168 a month in salary alone, before payroll taxes, benefits, recruitment, or paying someone to cover holiday and sick leave.

Now compare it to usage-based AI pricing. Retell AI publishes its rates openly at 0.07–0.31 per minute for voice agents, a range driven mostly by which text-to-speech and language model you select.

The honest reading: at realistic call volumes, usage-based AI does not come close to a salary. Sully.ai's AI Receptionist is priced on the same usage model. Even at the top rate and 5,000 minutes a month, you are at roughly half the cost of one person's base pay, and that person cannot work nights.

The three pricing models

Per-minute (usage-based). You pay for connected conversation time. Best when call volume is spiky or seasonal, and it makes overflow coverage nearly free to pilot. Risk: a bad conversation design that leaves callers talking in circles costs you money directly.

Per-seat or per-provider (subscription). A flat monthly fee per clinician. Predictable, easy to budget, and usually the better deal for consistently high volume. Risk: you pay the same in a quiet month.

Enterprise contracts. Custom pricing with implementation support, dedicated infrastructure, SSO and negotiated compliance terms. Appropriate for health systems; disproportionate for a three-provider clinic.

The costs vendors do not quote

Most vendors in this category do not publish pricing at all. Zocdoc's Zo, CureMD and Luma Health all route you to sales. When you get a quote, make sure it includes:

• Telephony pass-through. Carrier minutes are often billed separately from the AI. Retell lists telephony at $0.015/minute on top of the agent rate.
  
  

• Integration and implementation. One-off fees for connecting your EHR, and the staff hours to configure scheduling rules.
  
  

• Add-on features. PII redaction, safety guardrails, knowledge bases and QA tooling are frequently priced per minute on top of the base rate.
  
  

• Ongoing tuning. Someone has to review failed calls monthly and adjust. It is small, but it is not zero.

Choosing an AI receptionist for your practice setting

The right configuration depends far more on your setting than on the feature list.

Multi-provider clinics

Your constraint is scheduling complexity, not call volume. Multiple providers, varying appointment lengths, room and equipment constraints, and rules about who can see a new patient. Before anything else, test whether the system can honour your actual booking rules rather than a simplified version. Ask the vendor to configure your three most awkward scheduling rules during the trial, not after you sign.

Concurrency matters here too. If the AI handles calls one at a time, you have replaced a queue with a different queue.

Solo and small doctors' offices

Here the economics are starkest: you likely have one or two front-desk staff, and when they are at lunch or with a patient, calls go unanswered. Overflow and after-hours coverage are the whole case. Usage-based pricing suits you: volume is low enough that per-minute rates stay trivial, and you avoid committing to a per-seat subscription.

Be realistic about integration. If you run a smaller EHR, confirm a real integration exists before you evaluate anything else.

Specialty groups

Specialty practices have vocabulary problems. A system that mishears procedure names, laterality or medication names will generate booking errors that take longer to fix than taking the call would have. Test with your actual terminology, including the words patients get wrong.

Specialties with prep-dependent appointments, such as endoscopy, imaging or anything requiring fasting or a bowel prep, need the AI to deliver prep instructions reliably and confirm the patient understood. Ask how that is handled, and what happens if the patient asks a follow-up question about prep.

Wellness, integrative and alternative medicine practices

These practices often run on booking platforms rather than traditional EHRs, and their patients expect a more conversational front desk. Two things to check: whether the vendor integrates with your specific booking system, and whether the AI's voice and script can be tuned away from a clinical register. A system designed for a hospital call centre will sound wrong on your line.

Service menus here are also wider and change more often. Confirm you can update offerings yourself rather than filing a vendor ticket.

Telehealth and virtual-first practices

Your front desk is already a phone and a queue, which makes AI reception a more natural fit than for a bricks-and-mortar clinic, a case we make in full in why telehealth companies need an AI front desk. The differences are licensure and routing: appointments must be matched to clinicians licensed in the patient's state, and the AI needs to capture location reliably. Verify that the system can enforce state-based routing rules rather than simply booking the next available slot.

HIPAA: what compliance actually requires

This is the section vendors handle with a badge and a sentence. It deserves more, because getting it wrong is the one mistake that carries regulatory consequences.

Your AI vendor is a business associate

HHS is unusually direct here. Its business associates guidance lists, among its examples, a "third-party vendor Artificial Intelligence (AI) chatbot on a provider's patient portal that provides services involving the patient's PHI such as symptom assessment, medical reminders, and appointment scheduling."

That is this product category, described by the regulator. The consequence is simple: an AI receptionist handling PHI is a business associate, and per HHS, a covered entity may only disclose PHI to a business associate once it has satisfactory assurances in the form of a signed business associate agreement.

"HIPAA-compliant" on a marketing page is not that assurance. A countersigned BAA is, and our guide to what HIPAA compliance actually requires sets out what that agreement has to contain.

What the BAA must contain

HHS specifies that a BAA must contain the elements at 45 CFR 164.504(e). At minimum it must:

• Describe the permitted and required uses and disclosures of PHI by the business associate
  
  

• Provide that the business associate will not use or further disclose PHI other than as permitted by the agreement or required by law
  
  

• Where the business associate carries out your obligations under the Privacy Rule, require it to comply with the Rule in performing them

The Security Rule adds its own requirements where electronic PHI is involved, including that the business associate report any security incident it becomes aware of, which is the clause you will care about most if something goes wrong.

Subcontractors are where this quietly breaks

Here is the detail most buyers miss. Modern voice AI is assembled from parts: a speech-to-text provider, a language model, a text-to-speech voice, a telephony carrier. Several of those process the call content.

HHS is explicit that a business associate must establish a BAA with its subcontractor before disclosing PHI to it, and that all downstream subcontractors are themselves business associates. So the question to put to a vendor is not "are you HIPAA compliant" but:

Which subcontractors process call audio or transcripts, and do you hold a signed BAA with each one?

A vendor that cannot answer that quickly has not thought about it. Note also that HHS's conduit exception, the narrow carve-out for pure transmission services like the postal service, does not help here. It is limited to entities that only transmit PHI. A system that listens to a call and books an appointment is doing considerably more than transmitting.

EHR integration: the questions that separate real from marketing

Every vendor claims integration. The claims are not equivalent, and our ranked comparison of eight platforms sets out how they differ.

• Is it real-time and bidirectional? Can the AI read live availability and write a booking that appears immediately, or does it queue a request that staff confirm later? Both are legitimate products. Only one removes work.
  
  

• Which specific version and edition? Integrations are often built against one configuration. Name yours.
  
  

• Who maintains it when the EHR updates? Get the answer in writing.
  
  

• What happens on failure? If the integration drops mid-call, does the AI tell the patient someone will call back, or does it book into a void?
  
  

• What does it write? A booking is one thing; intake data, insurance details and call summaries written into the chart are another.

Ask for a live demonstration against a sandbox of your own EHR. A recorded demo against the vendor's instance tells you nothing about your configuration.

What happens when the AI fails

No competitor guide covers this properly, and it is the part that determines whether deployment is safe.

Emergency and clinical escalation

Define the escalation triggers before go-live, and test them. A patient describing chest pain must reach a human or emergency services immediately, with no conversational detour. Ask the vendor to show you the intercept and what happens outside business hours, when there may be no human to transfer to. If the answer is "it takes a message," that is a clinical risk, not a feature gap.

Loops and misunderstanding

The failure mode patients hate most is an AI that cannot understand them and will not let them out. Require a hard rule: after a set number of failed attempts, the system transfers or takes a callback number. Some vendors ship this as a named feature; with others you must configure it.

Downtime

Ask what happens when the vendor's service is unavailable. The acceptable answer is automatic failover to your existing phone tree or voicemail. Ask for an uptime commitment in the contract, and what the remedy is when it is missed.

Liability

Decide in advance who is responsible when the AI mis-books, mis-records insurance details or mishandles an escalation. Read the limitation-of-liability clause. Most vendor agreements cap liability at fees paid, which is a rational commercial position and also a reason to keep humans reviewing exceptions.

A buyer's checklist

Run a two-week pilot on overflow calls only, the ones currently going unanswered. The baseline is zero, so the risk is minimal and the data is real. Sully.ai's AI Receptionist pilots in two to three weeks.

Frequently Asked Questions

Will a medical receptionist be replaced by AI?

Not wholesale. BLS projects receptionist employment will decline 2% from 2025 to 2035: gradual contraction, not elimination. In practice, practices deploy AI against unanswered and after-hours calls first, which adds coverage rather than removing a person. The role shifts toward exception handling and in-person patient experience.

How much does an AI medical receptionist cost?

Published usage-based rates run 0.07–0.31 per minute for voice agents (Retell AI), putting a busy clinic at roughly 350–1,550 a month for 5,000 call minutes. Subscription vendors typically quote per provider. Most vendors in healthcare do not publish pricing and require a sales conversation, so insist on a quote modelled on your real call volume including telephony and add-ons.

Is there a free AI receptionist?

Several platforms offer free trial credits, and Retell starts with $10 in credits and no contract, but there is no sustainable free tier for a practice handling PHI. Free consumer tools generally will not sign a BAA, which makes them unusable for patient calls regardless of price.

Does an AI receptionist need a business associate agreement?

Yes, if it touches PHI. HHS explicitly lists a third-party AI chatbot providing appointment scheduling among its examples of business associates, and a covered entity may only disclose PHI once it has satisfactory assurances in a signed BAA. Confirm the vendor's subcontractors are covered too.

How long does EHR integration take?

It depends entirely on whether a pre-built integration exists for your system and version. Where one does, configuration is typically measured in days to a few weeks. Where one does not, you are commissioning development, so ask for a written timeline and who owns maintenance when the EHR updates.

What happens if a patient has an emergency?

A correctly configured system intercepts emergency language immediately and transfers to a human or directs the caller to emergency services, without continuing the booking conversation. This is the single most important thing to test before go-live, including after-hours behaviour when no human is available to receive a transfer.

Can patients tell they are talking to an AI?

Usually yes, and most vendors now recommend disclosing it at the start of the call. Patient tolerance is highest when the alternative was a voicemail box, and lowest when the AI blocks access to a person. Always offer an explicit route to a human.

Sources

1. U.S. Bureau of Labor Statistics: Occupational Outlook Handbook: Receptionists. Median pay 38,010/year(18.27/hour), 947,500 jobs, projected −2% employment change 2025–2035. May 2025 wage data.
   
   

2. U.S. Department of Health & Human Services: Business Associates guidance. Business associate definition and examples, BAA requirements at 45 CFR 164.504(e), subcontractor obligations, conduit exception.
   
   

3. Retell AI: Pricing. Published pay-as-you-go voice agent rates of 0.07–0.31/minute, telephony at $0.015/minute, $10 starting credits.
   
   

4. U.S. Department of Health & Human Services: Sample Business Associate Agreement Provisions.
   
   

5. Zocdoc: Zo AI phone assistant. Referenced as an example of a healthcare AI reception vendor that does not publish pricing.