We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] 338,000 Calls a Month and 8.8x ROI: Inova Health's Voice AI Front Door

    How a nonprofit health system integrated voice agents with Epic, CRM, and telephony to automate appointment work and release 4,272 staff hours each month.

    Conscious Engines

    Industry: Healthcare
    Organization: Inova Health
    Use case: Patient-access voice agents for appointments, routing, FAQs, and prescription workflows
    Evidence basis: Hyro customer case with named Inova executives, plus Inova's live description of Navigator functions
    Disclosure: This is an independent analysis by Conscious Engines. The source is published by the voice-agent vendor and does not disclose the full ROI calculation.

    1. Outcome at a Glance

    Inova Health deployed voice agents across its patient support operation and reported an 8.8x return on AI investment within six months. The system covered approximately 338,000 calls per month and released an average of 4,272 staff hours per month.

    Key Outcomes

    50%

    Appointment-management automation

    Calls successfully resolved by AI reported in the cited case.

    100% within six months

    Patient-access call coverage

    Coverage, not full automation reported in the cited case.

    79% in first 30 days

    Smart routing

    Intent identified and routed reported in the cited case.

    4,272 hours per month

    Staff capacity released

    Operating estimate reported in the cited case.

    8.8x

    Return on AI investment

    Vendor and customer calculation, formula not published reported in the cited case.

    MeasureReported resultEvidence interpretation
    Automated callsAbout 338,000 per monthVendor-reported volume
    Appointment-management automation50%Calls successfully resolved by AI
    Patient-access call coverage100% within six monthsCoverage, not full automation
    Smart routing79% in first 30 daysIntent identified and routed
    Staff capacity released4,272 hours per monthOperating estimate
    Return on AI investment8.8xVendor and customer calculation, formula not published

    Matthew Kull, Inova's chief information and digital officer, said, “We've saved around 4,000 hours per month through efficient call coverage.”

    Inova serves more than one million unique patients and more than four million visits annually. The scale makes patient access a material operating system, not a secondary call-center project.

    The public numbers describe three different levels of success that should not be collapsed into one metric:

    LevelReported resultWhat it means
    Coverage100% of patient-access call volume within six monthsThe system could participate in or route the full call population
    Smart routing79% in the first 30 daysThe system identified intent and directed calls to the appropriate destination
    Appointment automation50%Half of appointment-management calls were reported as resolved by AI

    Inova's current patient website independently shows that Inova Navigator can confirm and cancel appointments, provide hours and directions, and send a text related to the reason for a call. It corroborates a live patient-access surface. It does not independently validate the 8.8x return or 4,272-hour estimate.

    2. The Operational Problem

    Appointment management was the largest reason patients called. Long wait times caused some callers to abandon requests to schedule, verify, change, or cancel appointments. That can create no-shows, unused capacity, repeated calls, and poor access.

    The call-center infrastructure was also fragmented. A useful voice agent needed to understand why the person called, find the correct provider or location, interact with scheduling, and transfer complicated cases to staff.

    Healthcare creates additional constraints:

    • the system must verify identity before exposing protected information;
    • emergency or clinical symptoms require immediate escalation;
    • dates, names, prescriptions, and appointment details require high field accuracy;
    • callers interrupt, correct themselves, and use varied language;
    • a successful call must create the correct transaction in the system of record.

    3. What Was Built

    The Hyro voice agents were integrated with Epic, Inova's Cheers CRM, and NICE CXone telephony.

    System at a Glance

    Speech-to-text

    Convert patient speech and preserve critical details.

    Intent routing

    Identify scheduling, refill, FAQ, location, or provider needs.

    Epic connection

    Read or change eligible appointment information.

    CRM context

    Preserve patient-service and contact state.

    Telephony integration

    Cover inbound calls and transfer with context.

    Voice response

    Explain choices and confirm the transaction.

    CapabilityWorkflow role
    Speech-to-textConvert patient speech and preserve critical details
    Intent routingIdentify scheduling, refill, FAQ, location, or provider needs
    Epic connectionRead or change eligible appointment information
    CRM contextPreserve patient-service and contact state
    Telephony integrationCover inbound calls and transfer with context
    Voice responseExplain choices and confirm the transaction
    Safety and handoffEscalate clinical, uncertain, or sensitive cases

    The difference between coverage and automation is essential. The system could touch 100% of patient-access calls while autonomously resolving 50% of appointment-management calls. The remainder could be classified and transferred.

    A bespoke voice system can route predictable tasks to small intent and entity models, use deterministic Epic APIs for transactions, and reserve a larger language model for explanation. This keeps cost and risk proportional to task complexity.

    The production boundary

    A healthcare voice agent should use probabilistic models for language and deterministic systems for authority and transactions:

    caller speech -> domain ASR -> intent and entity extraction -> identity and policy checks -> Epic or CRM transaction -> spoken confirmation -> audit record

    The model may infer that a caller wants to cancel Tuesday's cardiology appointment. It should not decide which patient record is authorized, manufacture an appointment identifier, or report success before the scheduling system confirms the transaction.

    Failure modeRelease metric
    Wrong intentIntent accuracy by call type and confusion matrix
    Wrong date, provider, location, or medicationField-level precision and critical-entity error rate
    Authentication failureFalse-accept and false-reject rates
    Failed transaction reported as successfulConfirmed transaction success and false-confirmation rate
    Unsafe clinical request handled administrativelySafety-routing recall and time to human connection
    Caller repeats the same requestSeven-day repeat-contact rate
    Transfer loses contextWarm-transfer completion and agent rework time

    4. How It Reached Production

    Inova's six-month result suggests a workflow-first rollout.

    Start with high-volume administrative intents. Scheduling, modification, verification, cancellation, location, provider search, and FAQs have measurable completion criteria.

    Integrate before optimizing conversation. A pleasant voice that cannot complete the Epic transaction creates another handoff.

    Confirm critical fields. Dates, locations, providers, and patient identifiers should be repeated and validated before action.

    Measure capacity honestly. Released hours become financial value only when staffing, overtime, access, or appointment utilization changes.

    Audit ROI. An 8.8x return should specify cost, avoided work, revenue, no-show reduction, appointment conversion, time period, and attribution.

    Health systems should also report first-call resolution, rapid repeat calls, abandonment, wrong routing, patient satisfaction, transfer completion, and safety escalations.

    What remains undisclosed

    Evidence neededWhy it matters
    Calls offered, answered, contained, transferred, and abandonedSeparates traffic coverage from successful automation
    Average handle time before and after deploymentTests the 4,272-hour capacity calculation
    Cost of software, integration, telephony, support, and retained staffEstablishes the denominator for 8.8x return
    Appointment conversion and no-show changesTests the claimed revenue and access mechanism
    Accuracy by intent, language, and caller populationIdentifies cohorts hidden by an aggregate automation rate
    Patient satisfaction and complaint ratesConfirms that efficiency did not reduce access quality

    5. What Healthcare Leaders Should Take Away

    Inova's case demonstrates a commercial path for healthcare voice AI beyond clinical dictation. The voice agent sits at the patient front door, where repetitive administrative work and high call volume create measurable value.

    A production-ready patient-access system combines domain ASR, multilingual TTS, constrained dialogue, Epic connectors, private RAG, identity controls, and human escalation. The health system owns the workflow rules, test calls, and outcome data.

    The target metric is cost per correctly completed patient-access task, with safety, satisfaction, and repeat contact as release gates. The reported 8.8x ROI is a strong proof point, but prospective buyers should request the calculation before using it as a benchmark.

    Sources