01 - Real-time voice agent architecture

High call volume strained the front office.

A proposed voice agent for orthopedic scheduling and call routing. The concept focuses on a limited set of tasks, clear operating rules, and staff review rather than a broad virtual receptionist.

Concept study based on public sources. It is not a claim of client work, endorsement, partnership, production deployment, or access to private company systems.
About the AI AssessmentView Concept Studies

Runtime

Real-time voice system

Designed for live conversation and reviewable call records.

Voice stack

Speech recognition, language model, and generated voice

The concept uses these components for a responsive voice conversation.

System boundary

Practice scheduling system

The concept considers scheduling access, not insurance eligibility decisions.

Delivery scope

Proposed scheduling and routing pilot

A smaller proposed task than handling every kind of call.

02 - Business Context

900 calls. 200 callers. Patients redialing.

Summit Orthopedics ended its outside call-center service after ongoing patient complaints, problems answering insurance questions, and scheduling errors. The problem was how patients could reach the practice and get scheduled while staff handled heavy call volume.

Calls each day

~900

Peak call volume described in the public information used for this concept.

Different callers

~200

The gap suggests that some people called more than once.

Calls about scheduling

50%

The largest group of calls considered for limited automation in this concept.

Billing calls

~33%

Billing calls would be transferred to a separate billing team.

Operating environment

The practice was receiving more calls than its front office could handle.

The practice had ended its contract with an outside call center after patient complaints, problems answering insurance questions, and scheduling errors. The remaining staff faced a high volume of calls.

Scheduling system support

+
Patient lookup by name and date of birth
+
Provider schedule and availability reads
+
Appointment creation
+
Basic patient demographics

Not available

-
Check whether insurance covers a visit
-
Apply complex insurance rules
-
Track referrals

What repeat calls may mean

Repeat calls may mean callers were unable to complete what they needed by phone.

03 - Why Scope Matters

The key question is whether Summit could trust a voice agent with part of its phone service while the organization was changing its operations.

The proposal is limited: start with appointment scheduling and call routing, clear exclusions, staff review, and a gradual rollout. Add more tasks only if the system performs reliably.

04 - Proposed Scope

A focused scheduling agent with clear limits.

A focused scheduling agent is easier to test and review than a general AI receptionist.

Why workers' compensation is excluded

These appointments require authorization details, claim numbers, and insurance coordination that the proposed system cannot check. A wrong booking could create legal and day-to-day problems.

Why insurance questions wait

Staff knowledge about insurance is not yet written down in a form the system can check. Until those rules are documented and tested, the agent should not claim to know whether a patient is eligible.

Included in an initial pilot

+
Answer calls and identify what the caller needs
+
New and existing patient scheduling
+
Look up a patient by name and birth date in the scheduling system
+
Provider availability lookup
+
Appointment creation after confirmation
+
Route callers based on referral and body part
+
Transfer billing calls to the billing team
+
Workers compensation hard transfer
+
Record medical requests and pass them to staff
+
Staff review screen

Outside the initial pilot

-
Medical advice of any kind
-
Complex insurance interpretation
-
Check whether insurance covers a visit
-
Confirm referral details
-
Workers compensation scheduling
-
Surgery scheduling
-
Post-operative triage
-
Prior authorization handling
-
Payment collection

05 - Architecture

The voice system connects calls, conversation, and review.

This concept separates the phone service from the system that talks with callers. Each call would have its own session.

The room model matters because callers interrupt, pause, search for insurance cards, ask side questions, and become frustrated. A request-response chatbot architecture would feel brittle.

Who is on the call

Call participants

+
Patient calling by phone
+
Voice agent greeting, understanding the request, scheduling, or passing the call to staff
+
Staff member joining for a transfer or review

Information saved for each call

+
Call number
+
Phone number
+
System version
+
Practice location
+
Patient lookup status
+
Appointment status
+
Reasons a task may be blocked
+
Actions the system may take
+
Whether the call was passed to staff
+
How sure the system was about the caller's words

Call flow

+
Patient calls the practice
+
Phone service routes the call to the voice system
+
The system starts a dedicated call session
+
The agent follows the scheduling or routing workflow
+
Staff can join or receive a transfer when needed

06 - Technical Design

Every layer earns its place.

The challenge is to hold a natural phone conversation while keeping important actions, such as booking an appointment, under clear rules.
+
Live conversation. A system manages each call and keeps a record of the agent’s actions for review.
+
Phone service. The phone system sends calls to the voice agent and allows transfers to staff.
+
Speech recognition. Software turns spoken words into text, including practice and provider names.
+
Conversation timing. Audio detection helps the system notice pauses and when a caller starts speaking.
+
AI language model. AI handles limited conversation tasks. Written rules control actions that could affect a patient’s appointment.
+
Call review. Calls can be checked later for rule-following and answer quality.
+
Generated voice. The system speaks during the conversation at a calm pace and can stop when interrupted.
+
Fixed workflow. Clear steps and rules control what the system can do; the AI cannot freely take actions.

Speech recognition checks

+
Check that provider names are recognized and confirmed correctly.
+
Check common words patients use for body parts.
+
Ask patients to confirm their birth date before looking them up.
+
Check medical words separately from how the AI responds.

Voice style

+
Slightly slower than default speaking speed
+
One question at a time
+
Calm and clear, not clinical
+
No sales tone or exaggerated warmth
+
Pause after confirmations

Routing rules

The AI interprets what the caller says. A fixed list selects the provider group; the AI cannot choose a provider on its own.

07 - Safety Layer

The AI suggests an action. Rules decide whether it is allowed.

Appointment creation changes the practice schedule and can trigger follow-up work. It should not be treated as an unchecked suggestion from an AI model.

Core tool signatures

+
Look up a patient by name and birth date
+
Check a provider's available appointment times
+
Create an appointment only when the rules allow it
+
Transfer a call and record why
+
Send a task to staff
+
flag_for_review(call_id, reason, payload)
+
log_patient_statement(payload)

Allowed

Book only after the caller confirms

The agent can request an appointment only after the caller confirms the time and the request follows the approved rules.

Excluded

Do not book workers' compensation visits

For these calls, the proposed system transfers the caller to staff because it cannot check authorization and claim details.

Excluded

Do not give medical advice

If callers ask medical questions, the system records the request and passes it to staff. It cannot generate medical guidance.

Staff approval during pilot

The agent prepares an appointment request for staff review and confirmation instead of writing directly to the practice schedule.

08 - Handling a Conversation

Let callers interrupt, and give them time to finish.

Callers should be able to interrupt the agent. The agent should also wait long enough for someone who speaks slowly or pauses to finish.

Patience prompts

+
"No problem, take your time."
+
"I heard part of that, but I want to make sure I get it right."
+
"I may have misheard the provider name. Did you say Dr. Chen or Dr. Cohen?"

Confirm when speech is unclear

If speech recognition is unsure about a provider name, birth date, or appointment date, the agent asks the caller to confirm before taking action.

Normal scheduling conversation

700-900ms

Elderly or slow caller

1000-1400ms

Yes / no confirmation

500-700ms

Caller searching for insurance card

1500-2500ms

Noisy audio

Conservative plus clarification

09 - Staff Review

Staff need to see and review what the agent did.

Staff need to see what the agent heard and what it plans to do. They can then correct errors and decide whether an appointment is ready to confirm.

What to measure first

How often staff need to correct an appointment.

Staff review

Call details

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Live written transcript

Shown to staff or reviewers so they can check what happened on the call.

Staff review

When the caller or agent spoke

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Actions the agent tried to take

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Response times

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Replay a call

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Test unusual situations

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Calls needing staff review

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Test staff transfers

Shown to staff or reviewers so they can check what happened on the call.

Staff review

Why staff corrected a booking

Shown to staff or reviewers so they can check what happened on the call.

10 - Staged Rollout

Earn scope through performance.

The agent would not start by booking appointments on its own. Each step would be reviewed before allowing it to take on more work.

Rollback triggers

!
Staff often need to correct the selected provider
!
Workers compensation booked by mistake
!
The system gives medical advice
!
A duplicate or nonexistent appointment is created
!
Calls fail to transfer to staff
!
Patient complaints increase

Phase 0

Write down the rules

Record which providers and appointment types are allowed, where calls should go, how billing calls are handled, and which medical requests go to staff.

Phase 1

Test without booking

The agent listens to practice calls or recordings and suggests actions, but does not make appointments.

Phase 2

Staff approve each booking

The agent handles a limited set of calls and sends suggested bookings to staff to confirm.

Phase 3

Allow limited booking

Let the agent book only approved appointment types, with staff checking the results each day and a way to turn it off.

Phase 4

Consider a wider rollout

Cover more calls only after quality checks show the agent reduces staff work and makes few mistakes.

11 - Acceptance Gates

Proposed test targets, not results from a live system.

These are proposed targets to check with simulated calls. They are not measured results from a live production system.

Identify why the caller is calling

>=92%

Design target on simulated calls after one follow-up question.

Record birth date correctly

>=96%

Required before looking up a patient or preparing an appointment.

Typical reply time

<700ms

Target time from when the caller finishes to when the agent starts speaking.

Slowest replies

<1.5s

Target for the slowest five percent of replies that need the system to check appointment details.

Rule 1

Never book without confirmation

The caller must confirm the appointment time before the system can request a booking.

Rule 2

Transfer workers' compensation calls

The system cannot book these visits. It must transfer the call to staff.

Rule 3

Pass medical questions to staff

The system records the question and sends it to staff. It cannot generate medical advice.

12 - Proposed Pilot

How a pilot could be evaluated.

This describes a proposed system with records staff could review. It does not mean a live phone system is active.

What a demo would show

A sample call, how the system understands the request, and whether it schedules, transfers, or sends the caller to staff.

What staff could review

The written call transcript, actions the system took, response times, rule checks, and a recording of the call steps.

Situations to test

Background noise, a caller who speaks slowly, unclear provider names, a scheduling system outage, no open appointments, and a workers' compensation call.

13 - Recommendation

Start with a focused voice agent and expand carefully.

The proposal prioritizes reducing pressure on staff. Any future system should earn more responsibility only after it demonstrates safe scheduling, reliable transfers, and a net reduction in staff work.

14 - Study Status

Serious proposal. Not a deployment claim.

This concept study explores a focused response to call and scheduling pressure, with defined limits, safeguards, and a path for further review. It is not completed or deployed work.

What this case study is

A proposal for how a voice agent, scheduling rules, staff review, and a staged rollout could work in this setting.

What it is not

A claim that this exact system was deployed for Summit Health, or that private systems, private data, or live production metrics were available.

Start Here

Start with the business boundary, then choose the AI.

Talk with Life Tap Labs about your call-handling needs and the safeguards required before considering a voice agent.

About the AI AssessmentView Concept Studies