Healthcare AI glossary: 34 terms every medical practice should know.
The AI voice category moved fast between 2024 and 2026. Vendor marketing moved faster. These are the terms that matter when a medical practice evaluates AI phone software, in plain language and without the spin.
Last updated: 2026
Every definition is written to stand alone. Copy any one of them into an email to your practice manager and it will make sense without the rest of the page. For how the terms fit together on a real call, see how AI phone agents work in a medical practice.
AI voice fundamentals
- AI phone agent
- An AI phone agent is software that answers inbound phone calls, understands what the caller wants through natural conversation, and completes the request without a human. In healthcare, an AI phone agent typically books appointments, reschedules, cancels, takes prescription refill requests, and transfers complex calls to staff.
- An AI phone agent differs from a phone tree because it does not require the caller to press numbers or navigate menus. The caller speaks normally and the system responds.
- AI receptionist
- An AI receptionist is an AI phone agent configured to handle the front desk phone workload at a specific business. In a medical practice, an AI receptionist answers calls, identifies the patient, checks provider availability, books or moves appointments, and routes anything outside its scope to human staff.
- The term is used interchangeably with AI phone agent. "AI receptionist" emphasizes the role being replaced. "AI phone agent" emphasizes the technology.
- Voice AI
- Voice AI is any system that processes spoken language as input, generates spoken language as output, or both. Voice AI in healthcare covers phone agents, clinical documentation tools, and patient-facing voice assistants.
- Voice AI is a broader category than AI phone agents. All AI phone agents are voice AI. Not all voice AI answers phones.
- Conversational AI
- Conversational AI is software that holds a multi-turn dialogue with a person, maintaining context across turns rather than treating each input independently. A conversational AI system can handle a caller who says "actually, make it Thursday instead" three turns into a booking conversation.
- Non-conversational systems require the caller to restart if they change their mind mid-request.
- Speech to text (STT)
- Speech to text, also called automatic speech recognition or ASR, converts spoken audio into written text. In an AI phone agent, STT runs continuously during the call so the system can interpret what the caller said.
- STT accuracy on proper nouns (patient names, medication names, provider names) is meaningfully lower than accuracy on general English. Vendors address this with keyword boosting, which pins expected terms so the model weights them higher.
- Text to speech (TTS)
- Text to speech converts written text into spoken audio. In an AI phone agent, TTS generates the agent's side of the conversation.
- TTS quality determines whether a caller perceives the agent as natural or robotic. Latency between the system deciding what to say and the audio playing is the other half of that perception.
- Latency
- Latency in voice AI is the delay between a caller finishing their sentence and the agent beginning its response. Human conversation has natural gaps of roughly 200 milliseconds. Delays above 800 milliseconds read as broken or laggy to most callers.
- Latency is the hardest constraint in voice AI because every step (speech recognition, intent detection, database lookup, response generation, speech synthesis) adds to it.
- Containment rate
- Containment rate is the percentage of inbound calls an AI phone agent resolves end to end without transferring to a human. A containment rate of 70 percent means 7 out of 10 callers get their request completed by the AI.
- Containment rate is the single most useful performance metric for an AI phone deployment. It directly translates to staff hours reclaimed. Vendors who cannot report containment rate on live deployments have not measured it.
- Transfer rate
- Transfer rate is the inverse of containment rate: the percentage of calls the AI phone agent routes to a human. A 30 percent transfer rate means 3 out of 10 calls reach staff.
- Some transfer volume is intentional and correct. Clinical questions, urgent triage, and anything outside configured scope should transfer. The transfer rate to watch is unintentional transfers caused by the AI failing to complete a task it should handle.
- Intent detection
- Intent detection is the step where an AI phone agent classifies what a caller wants from their first few sentences. Common healthcare intents include new appointment, reschedule, cancel, refill request, insurance question, billing question, and clinical question.
- Intent detection accuracy determines whether the rest of the call goes well. A misclassified intent sends the conversation down the wrong path immediately.
- Barge-in
- Barge-in is the ability of a caller to interrupt the AI agent mid-sentence and have the system stop talking and listen. Human receptionists handle interruption naturally. AI agents without barge-in support talk over callers, which reads as rude and frustrates people quickly.
- Noise suppression
- Noise suppression filters background audio (television, other conversations, road noise, office chatter) out of the caller's audio stream before speech recognition processes it. Without it, an AI phone agent can transcribe background speech as if the caller said it, producing nonsense responses.
- Noise suppression matters more in healthcare than in most industries because patients frequently call from cars, waiting rooms, and workplaces.
EHR and integration
- EHR (Electronic Health Record)
- An EHR is the software system a medical practice uses to store patient records, provider schedules, clinical notes, prescriptions, and billing information. Common outpatient EHRs in the US include athenahealth, eClinicalWorks, Epic, NextGen, ModMed, and Dentrix for dental.
- The EHR is the system of record. Any AI phone agent that does not write to the EHR is creating a second system your staff has to reconcile manually.
- EHR integration
- EHR integration means an external software system can read from and write to the EHR through an official interface. For an AI phone agent, real integration means the agent can look up a patient, check provider availability, create an appointment, and write a call summary to the chart, all through the EHR's API.
- Integration depth varies significantly between vendors. Some vendors describe integration as "we send an email to your front desk," which is not integration.
- API (Application Programming Interface)
- An API is a defined interface that lets one software system talk to another programmatically. EHR APIs expose specific operations: look up a patient, retrieve a schedule, create an appointment, submit a clinical task.
- API access is typically gated. EHR vendors require registration, review, and in some cases a formal partner agreement before granting production access.
- athenahealth Marketplace Partner
- An athenahealth Marketplace Partner is a vendor whose integration with the athenahealth EHR has been reviewed, tested in sandbox, and approved for publication by athenahealth. The review covers technical implementation, security posture, and HIPAA readiness.
- Marketplace Partner status is verifiable. Any vendor claiming athenahealth integration should be able to produce a Marketplace listing URL. Vendors without one are either using unofficial access or forwarding requests through a patient portal.
- Write-back
- Write-back is the ability of an external system to save data into the EHR, as opposed to only reading from it. An AI phone agent with write-back creates appointments, updates existing ones, and logs encounter notes directly in the patient chart.
- Read-only integrations can tell a caller what times are available but cannot book them. That still leaves the booking work with your staff.
- Reconciliation
- Reconciliation in scheduling is what happens when two systems try to book the same appointment slot at the same moment. If an AI phone agent offers a 3 PM Thursday slot to a caller while a front desk staff member books that slot for a walk-in, one of the two bookings has to fail gracefully.
- Reconciliation handling is a good technical litmus test for AI phone vendors. A vendor with a specific answer has built for production. A vendor with a vague answer has not hit the problem yet.
- Clinical inbox task
- A clinical inbox task is an item routed to a provider or clinical staff member inside the EHR for review and action. Prescription refill requests, lab result acknowledgments, and patient messages typically arrive as inbox tasks.
- An AI phone agent handling refill requests should create a proper inbox task, not send an email or leave a voicemail for staff to transcribe.
- Appointment type
- An appointment type is the category assigned to a scheduled visit, which determines its duration, the provider required, and sometimes the room or equipment needed. Family medicine practices commonly distinguish new patient visits, established patient follow-ups, annual physicals, and same-day sick visits.
- An AI phone agent must be configured with a practice's specific appointment types to book correctly. Generic configuration produces wrong-length appointments.
HIPAA and compliance
- HIPAA
- HIPAA (the Health Insurance Portability and Accountability Act) is the US federal law that governs how Protected Health Information is used, stored, and disclosed. HIPAA applies to healthcare providers, health plans, healthcare clearinghouses, and any vendor that handles PHI on their behalf.
- HIPAA compliance is a posture, not a certificate. There is no official government-issued "HIPAA certification." Any vendor claiming one is describing something else.
- PHI (Protected Health Information)
- Protected Health Information is any information that identifies a patient and relates to their health condition, treatment, or payment for care. PHI includes obvious identifiers like name and date of birth, and less obvious ones like voice recordings, appointment history, and phone numbers linked to a medical record.
- Voice recordings of patient calls are PHI. Any vendor treating call audio as something other than PHI is misreading the regulation.
- BAA (Business Associate Agreement)
- A Business Associate Agreement is a contract required by HIPAA between a covered entity (your practice) and any vendor that handles PHI on your behalf. The BAA defines permitted uses of PHI, required safeguards, breach notification obligations, subcontractor requirements, and what happens to PHI when the contract ends.
- A vendor without a BAA cannot lawfully receive PHI. If an AI phone vendor will not sign one before deployment, they are not a viable option.
- Business associate
- A business associate is any vendor or contractor that creates, receives, maintains, or transmits PHI on behalf of a healthcare provider. AI phone vendors are business associates. So are their cloud hosting providers, speech recognition providers, and any AI model provider processing patient audio.
- Each layer requires its own BAA. A vendor whose subprocessors lack BAAs has a compliance gap regardless of what their own BAA says.
- Subprocessor
- A subprocessor is a third-party service a vendor uses to deliver their product, which in the process handles your data. For an AI phone agent, subprocessors typically include the telephony provider, cloud hosting, speech recognition service, and language model provider.
- Every subprocessor touching PHI needs a BAA with the vendor. Ask for the subprocessor list. Vendors who will not disclose it are asking you to accept unverifiable compliance claims.
- Encryption at rest
- Encryption at rest protects data stored on disk. The current standard is AES-256. For an AI phone agent, encryption at rest covers stored call recordings, transcripts, and any cached patient data.
- Ask vendors to name the standard. "Everything is encrypted" without naming AES-256 suggests they have not thought carefully about it.
- Encryption in transit
- Encryption in transit protects data moving between systems over a network. The current standard is TLS 1.3. For an AI phone agent, this covers the connection between the caller, the vendor's infrastructure, and the EHR API.
- Audit log
- An audit log is a permanent record of every access to and action on PHI, including who did it, what they did, and when. HIPAA requires covered entities to be able to produce audit logs during an investigation.
- If your AI phone vendor cannot export audit logs for your practice on demand, you cannot demonstrate your own compliance for the calls they handled.
- Minimum necessary standard
- The minimum necessary standard is a HIPAA requirement that covered entities and business associates access only the PHI needed to accomplish a specific purpose. An AI phone agent booking an appointment needs the patient's identity and schedule, not their full clinical history.
- Vendors pulling more data than the task requires are creating unnecessary exposure.
Practice operations metrics
- Call abandonment rate
- Call abandonment rate is the percentage of inbound calls that ring out, go to voicemail, or are hung up before reaching a person. Busy medical practices commonly run 15 to 25 percent abandonment.
- Most practices have never measured this. Most modern phone systems can report it if you ask.
- Patient lifetime value (PLV)
- Patient lifetime value is the total revenue a practice expects from one patient over the full course of their relationship. It is the figure that turns a missed call rate into a revenue number, and it is the input most worth being conservative about.
- US primary care averages about $3,000 per patient (Physicians Weekly, 2021). Specialty and procedure heavy practices run substantially higher. The MedPhone calculator defaults to that figure and then discounts it twice before reporting anything, once for the share of inquiries that would have booked and once for the share of unreached callers who never ring back.
- Fully loaded cost
- Fully loaded cost is the true annual expense of an employee, including base salary plus payroll taxes, benefits, paid time off, equipment, training, and amortized recruitment and turnover cost. For a US medical front desk role, fully loaded cost runs roughly 45 to 55 percent above base salary.
- A $35,000 base salary front desk hire costs a practice approximately $52,000 per year all in.
- No-show rate
- No-show rate is the percentage of scheduled appointments where the patient does not arrive and did not cancel in advance. US outpatient practices commonly see 10 to 20 percent.
- After-hours reschedule capability reduces no-shows because patients who realize at 8 PM that they cannot make tomorrow's appointment can move it instead of simply not showing.
- After-hours volume
- After-hours volume is the share of inbound call volume arriving outside normal business hours, including evenings, weekends, and holidays. Family medicine practices typically see 15 to 20 percent of total call volume after hours.
- Without AI or an answering service, this volume defaults to voicemail. Voicemail conversion to booked appointment is substantially lower than live conversation conversion.
Two of these terms have their own pages. Containment and transfer rate are set out with real numbers in the AI phone agent case study at a family medicine practice, and what the athenahealth Marketplace Partner status means in practice is on the athenahealth AI receptionist page. The compliance terms are covered on the HIPAA compliant AI voice agent page. For the rules themselves rather than our summary of them, Health and Human Services publishes the HIPAA Rules.
Healthcare AI glossary FAQ
The questions that come up once the vocabulary is settled.
Still have questions?
Can't find the answer you're looking for? Reach out to our team and we'll get back to you shortly.
A phone tree requires the caller to press numbers to navigate a menu and routes them to a queue or voicemail. An AI receptionist holds a natural conversation, understands what the caller wants in their own words, and completes the request directly.
Yes. Any voice recording that identifies a patient and relates to their health, treatment, or payment qualifies as Protected Health Information under HIPAA. Voice recordings must be encrypted and access logged, and the vendor handling them has to be under contract to protect them.
Well-tuned deployments in primary care typically reach 65 to 80 percent containment within a few months of go-live. Early deployments often start lower, around 50 to 60 percent, and improve as call flows are refined for the specific practice.
No. Integration depth varies widely. Some vendors write appointments directly to the EHR through an official API. Others forward requests to a patient portal or an email inbox that staff must review manually. Ask specifically whether the vendor writes appointments and call summaries to the chart.
Typical deployments run two to four weeks from contract signature to first live call. Most of that time is spent configuring call flows for the practice's specific appointment types and transfer rules, not on the technical integration itself.
Named encryption standards for data at rest and in transit, a written account of which vendors in the chain can touch PHI and under what obligations, a written commitment that PHI is not used for model training, and the ability to export audit logs on demand.
Now hear one take a real call.
Twenty minutes. Your EHR, a real call, and plain answers to anything on this page.
Book a demo