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] When Medical Speech AI Becomes Workflow Infrastructure: The SolutionHealth Case

    How a health system with more than 1,000 clinicians combined ambient documentation, active dictation, and Epic integration, reporting a 56% improvement in average documentation time.

    Conscious Engines

    Industry: Healthcare
    Organization: SolutionHealth
    Use case: Medical speech recognition and ambient documentation
    Evidence basis: Microsoft customer report, named SolutionHealth clinicians, and a December 2024 survey of 190 users
    Disclosure: This is an independent analysis by Conscious Engines based on public information. The primary source is a technology-vendor customer story, so results should be treated as organization-reported rather than independently audited.

    1. Outcome at a Glance

    SolutionHealth, a health system staffed by more than 1,000 clinicians, reported medical speech AI use across nearly 60,000 encounters. Average documentation time improved by 56%. The organization estimated that released capacity was equivalent to 2.5 additional appointments per clinician per day.

    Key Outcomes

    Nearly 60,000

    Clinical encounters

    Reported live-workflow volume across the combined speech and ambient documentation deployment.

    56%

    Documentation-time improvement

    Average improvement reported through the Microsoft customer case.

    2.5 per day

    Estimated appointment capacity

    Capacity equivalent, not evidence that every clinician booked additional appointments.

    190

    Survey respondents

    Microsoft survey across SolutionHealth facilities in December 2024.

    4 to 6 hours

    Weekly time saved by one physician

    Individual clinician report, not the deployment-wide average.

    MeasureReported resultInterpretation
    Health-system clinical workforceMore than 1,000Organization size, not a disclosed deployment count
    EncountersNearly 60,000Observed usage
    Improvement in average documentation time56%Operational performance
    Estimated capacity released2.5 appointments per clinician per dayCapacity equivalent, not proven booked volume
    Users saying patient experience improved90%Survey perception
    Users reporting more patient focus65%Survey perception
    Users reporting improvement in burnout or fatigue92%Survey perception
    Users reporting better work-life balance93%Survey perception
    Survey respondents190Microsoft survey across SolutionHealth facilities, December 2024

    One physician reported saving four to six hours per week. The result is commercially important because it shows value from a hybrid speech layer, not ambient capture alone. Clinicians could use ambient documentation during the encounter and active medical dictation when direct control was preferable.

    Gregory Kaupp, MD, a pediatrician at SolutionHealth, described the individual impact directly:

    “DAX has reduced my total documentation time per week by 4-6 hours.”

    The metrics need careful reading. The 2.5-appointment figure is an estimated capacity equivalent. It does not establish that every clinician saw 2.5 more patients, that revenue rose by the corresponding amount, or that staffing cost fell. The experience percentages came from a 190-user survey in December 2024, a useful but self-reported sample.

    What the public evidence establishes

    QuestionPublic answer
    Was the system used in live clinical workflows?Yes, across nearly 60,000 encounters; the number of distinct active clinicians is not disclosed
    Did it combine active and ambient documentation?Yes, Dragon Medical One and DAX Copilot were used together
    Was it integrated into the system of record?Yes, through Epic Haiku and Hyperdrive workflows
    Was the 56% result independently measured?No, it is reported through a vendor case study
    Is the denominator behind “56% improvement” public?No, the public story does not provide the baseline distribution or calculation method
    Did 2.5 more appointments occur per clinician per day?Not established; the figure is a capacity equivalent
    Are note accuracy and critical-error rates public?No

    This evidence is sufficient to show operational adoption and a credible time-saving signal. It is not sufficient to infer clinical safety, causal financial return, or equivalent performance across specialties.

    2. The Operational Problem

    Healthcare documentation varies by specialty, encounter type, clinician preference, and environment. A single capture mode is rarely sufficient.

    Ambient systems work well when a consultation is conversational and the output follows a repeatable note pattern. Active dictation remains valuable when the clinician needs to enter a precise impression, spell an uncommon term, amend an existing record, or work in a setting where continuous capture is inappropriate. Mobile access matters because physicians move between consultation rooms, wards, and administrative work.

    SolutionHealth also had a workflow constraint shared by most large providers: the AI could not become a separate destination. If a clinician must copy text between tools, search for the correct patient, or reconstruct context, much of the time saving disappears and the risk of wrong-record errors rises.

    The true problem was therefore broader than speech recognition accuracy. It was to create a clinical input layer that could:

    • recognize medical language and physician commands;
    • turn natural conversation into an editable structured note;
    • support direct dictation for controlled entry;
    • work on mobile and desktop surfaces;
    • pass drafts into Epic with the correct patient and encounter context;
    • preserve clinician review and accountability;
    • meet healthcare security and data-handling requirements.

    This framing changes the buying decision. The enterprise is not procuring “AI transcription.” It is redesigning how clinical information enters the record.

    3. What Was Built

    The public case describes a combined stack using Microsoft Dragon Medical One, DAX Copilot, and integration with Epic, including Epic Haiku and Hyperdrive workflows.

    System at a Glance

    Dragon Medical One

    Active medical dictation and voice commands.

    DAX Copilot

    Ambient capture and draft clinical documentation.

    Epic integration

    Patient context, workflow continuity, review, and record entry.

    Mobile and desktop access

    Capture at the point of care and complete work across settings.

    Clinician review

    Validate, edit, and sign the final note.

    ComponentRole
    Dragon Medical OneActive medical dictation and voice commands
    DAX CopilotAmbient capture and draft clinical documentation
    Epic integrationPatient context, workflow continuity, review, and record entry
    Mobile and desktop accessCapture at the point of care and complete work across settings
    Clinician reviewValidate, edit, and sign the final note

    This is a useful reference architecture for bespoke deployments. The speech model should be separated from the note-composition model. Each can then be evaluated independently. Speech quality can be measured with word error rate and a clinically weighted term error rate. Note quality can be measured with fact coverage, unsupported statement rate, negation errors, section placement, edit distance, and clinician acceptance.

    Evaluation layerMinimum production measures
    Medical ASROverall WER, critical-term error rate, medication and dosage accuracy, speaker attribution, performance by accent and specialty
    Note generationRequired-fact recall, unsupported-fact rate, negation errors, section correctness, material edit rate
    WorkflowTime to signed note, after-hours EHR time, failed insertion rate, wrong-patient or wrong-encounter events
    AdoptionEligible encounters, activation rate, repeat-user rate, use by specialty, abandonment after first use
    SafetyCritical omissions, clinically material additions, escalation rate, complaints, incidents after model changes
    EconomicsCost per generated note, cost per clinician-approved note, support cost, external-model fallback rate

    A private healthcare stack may add retrieval over local templates and policy, deterministic checks for medication names and units, specialty routing, speaker diarization, and a structured output schema. It may also route easy encounters to a smaller model while escalating ambiguous notes to a stronger model. This makes cost proportional to difficulty.

    The most valuable customization usually sits in four enterprise-owned assets: the vocabulary, the note policy, the integration logic, and the evaluation set. These encode how the institution actually works.

    4. How It Reached Production

    The deployment demonstrates five production disciplines.

    Offer more than one interaction mode. Clinicians could choose ambient documentation or active dictation. Adoption improves when AI adapts to the task instead of forcing every task into one interface.

    Integrate with the record. Epic integration reduced context switching. Any healthcare buyer should test patient matching, encounter selection, note destination, role-based access, and failure recovery before broad rollout.

    Measure both speed and experience. The reported 56% documentation-time improvement is more useful when paired with clinician and patient-experience measures. Efficiency that degrades trust is not a successful healthcare outcome.

    Preserve review. Drafts must remain visibly provisional until accepted by the clinician. High-risk terms, unsupported diagnoses, dosages, allergies, and negation should receive special treatment in the evaluation suite and user interface.

    Build a model-change process. Every update to ASR, prompts, templates, or summarization models can alter clinical behavior. A release gate should compare the candidate against a frozen local test set and report regressions by specialty.

    For a similar rollout, a practical 90-day program would begin with two specialties, 20 to 40 clinicians, and a benchmark set of de-identified encounters. The enterprise would measure baseline documentation time and after-hours work, run a shadow evaluation, introduce supervised use, and expand only after clinical safety and adoption thresholds are met.

    The release decision should be made at the specialty level. A model that performs well for routine primary-care follow-ups may not be ready for behavioral health, complex oncology, emergency medicine, or heavily interpreted encounters. One enterprise-wide average can hide clinically important failure clusters.

    5. What Healthcare Leaders Should Take Away

    SolutionHealth's strongest lesson is architectural: clinical speech AI becomes valuable when it is a flexible input layer inside the workflow. Ambient capture, direct dictation, mobile access, and record integration solve different parts of the documentation burden.

    The business case should remain conservative. A provider can model value as:

    annual value = encounters × minutes released × loaded clinician cost + incremental contribution from capacity actually used + avoided documentation rework

    The model should apply separate realization rates. Time released after clinic may improve retention and wellbeing without creating booked revenue. Capacity during scheduled hours may support more encounters, but only if demand, rooms, nursing support, and payer rules allow it.

    A production-ready clinical model system combines healthcare-tuned speech-to-text, configurable note generation, private enterprise RAG, deterministic validation, Epic or EHR workflow integration, and an evaluation layer owned by the provider. The target metric is not tokens generated. It is cost per clinician-approved note at an agreed safety level.

    The reported results establish a credible upper range for value: a 56% documentation-time improvement, strong clinician sentiment, and nearly 60,000 documented encounters inside a health system with more than 1,000 clinicians. The next step for any enterprise is to reproduce those measures on its own specialties, accents, templates, and record workflows before assuming the same return.

    Sources