醫療診所評估 AI 電話接待：購買前先建立損益平衡工作量台帳

AI Phone Reception for a Medical Practice: Build a Break-Even Ledger Before You Buy

A medical practice considering an AI phone receptionist should not begin with the vendor’s “calls answered” count. That number is a routing outcome, not a financial result. An answered call may still require a human to listen to the recording, correct a wrong appointment, return a duplicate message, explain a failed booking, or handle a clinical question that the automation should never resolve. The useful question is narrower: does the system remove a measurable, repeatable unit of work without shifting equal or greater work, risk, or cost elsewhere?

The practical answer is to run a small, controlled workload ledger. Measure representative calls before the pilot, classify what happens after each call, and compare the human minutes and operational costs that actually change. Separate released capacity from cash savings: time returned to a receptionist is not payroll saved unless staffing, overtime, or another real expense changes. Keep safety, accessibility, privacy, and service quality as gates, not as items that can be traded away for a positive spreadsheet.

This article gives a reusable model, a worked hypothetical example, a short pilot protocol, and a stop/go checklist. It is an evaluation method, not a claim that any particular supplier is effective. The example numbers below are invented solely to demonstrate the arithmetic; they are not market averages, measured clinic results, or a forecast.

The decision in one minute

Do not buy based on an answer-rate demo. First establish the specific call category, hours, and staff task you want to change. Then measure: incoming eligible calls; calls completed without human repair; human follow-up minutes; false or duplicate actions; integration, review, and governance effort; and any truly avoided cash cost. A “break-even” result is credible only if the same definitions are used before and during the pilot, the denominator is the eligible workload rather than every call, and service and safety gates pass independently.

There are three outcomes worth distinguishing. Cash-positive means a cash expense is actually avoided or an additional paid service is funded by a verifiable benefit. Capacity-positive means staff regain time that is used for named work, even though payroll does not fall. Not worthwhile means unresolved repairs, failed access, clinical-boundary events, or ongoing costs outweigh the demonstrated benefit. A practice can rationally run a capacity-positive pilot, but it should not describe that outcome as payroll savings.

Why the unit of analysis is a call disposition, not a call answer

An incoming call can have several endings: information delivered, appointment requested, appointment confirmed, cancellation recorded, callback queued, call transferred, voicemail left, or no safe completion. The same call can also move through more than one system: a phone agent, scheduling software, a practice-management system, a human inbox, and a patient callback. If the vendor counts “conversation ended” as success while the clinic later repairs the record, the metric hides rather than measures work.

Use the completed disposition as the unit. For every sampled call, record the caller’s starting request, the allowed endpoint, what the software actually did, what a human had to do next, and whether the caller received a clear next step. A call should be “completed without repair” only when the intended task is correct, the record is correct, the caller has been told the truthful status, and no unplanned follow-up was necessary. If one of these conditions is unknown, label it “unverified,” not “successful.”

This framing matters in healthcare because reception work is not merely a queue to shrink. A DREES study of French general practitioners’ schedules, based on a 2011 survey wave and published in March 2012, reported an average 57-hour working week and described administrative work as intertwined with consultations. That is historical context, not a current staffing benchmark and not evidence that a phone agent will reduce workload. Its useful lesson for a buyer is that work is heterogeneous and clinic-dependent; a local baseline is essential. DREES, Les emplois du temps des médecins généralistes

The boundary: administrative work versus clinical work

An administrative phone task can still become clinically sensitive mid-conversation. A caller may ask to move an appointment and then mention worsening pain, ask whether a symptom can wait, or request a medication change. The automation must not turn that context into a diagnosis, reassurance, or an invented urgency category. The practice needs an approved route for the caller and the staff member who receives the escalation.

Before calculating economics, define the eligible workload. A safe initial scope may include published opening hours, department routing, a request for a callback, or a narrowly permitted appointment action that is already supported by the practice’s calendar and identity rules. Exclude diagnosis, triage, treatment advice, test interpretation, prescription decisions, and any call type whose clinical route has not been designed by the responsible professionals. A call outside scope is not an automation failure if it is correctly handed to a person; it is an unsafe failure if the system pretends to resolve it.

Research on medical reception also cautions against treating “booking” as a mechanical form fill. A conversation-analysis study examined 18 recorded calls in one New Zealand university primary-care practice and found that receptionists’ appointment interactions involved the booking system and clinical components, including attention to possible urgency. This small, local qualitative study is not evidence about AI performance or the prevalence of any call type. It is evidence that the human role can include judgment and interactional work that a break-even model must not silently assume away. Weatherall and Grattan, A Conversation Analytic Study of Calls to Medical Reception for Doctor’s Appointments

The workload ledger

The following ledger is the reusable asset. Keep one row per sampled call or a privacy-preserving call identifier. Do not put free-form clinical narratives into a finance sheet. Store only the minimum operational fields needed for evaluation, under the clinic’s approved data-handling process.

Do not collapse “appointment requested,” “slot held,” and “appointment confirmed.” A request is an intention; a hold may be temporary; confirmation requires authoritative evidence from the scheduling system. When the integration times out, the correct state is “uncertain—verify,” not a confident confirmation. A caller should not be told the appointment is booked merely because a language model generated a plausible sentence.

Formulas that separate capacity from cash

Let:

N = number of eligible calls in the measurement window
M0 = baseline human minutes per eligible call, including downstream handling
M1 = pilot human minutes per eligible call, including review and repair
F = loaded staff cost per hour used only for a capacity-equivalent calculation
V = vendor subscription and usage cost for the same window
I = one-time implementation and integration cost
G = governance, training, monitoring, and incident-review cost
S = cash expense demonstrably avoided during the window

The estimated capacity returned is N × (M0 − M1) / 60 hours. Its wage-equivalent value is that number multiplied by F, but it is not cash savings unless a real cash cost is avoided. A simple cash contribution for a steady-state window is S − V − G. For a first-window decision that includes setup, use S − V − G − I. If implementation cost is amortized, state the chosen horizon and the amortization method; do not compare a one-time setup invoice with a single day of benefits.

For an economic comparison between the current process and the pilot, do not assume every minute of baseline time can be removed. Let A be the share of returned minutes that can actually be redeployed or converted into a cash change. Then a capacity-equivalent net is N × (M0 − M1) / 60 × F × A − V − G, with A shown as an explicit assumption. If staff remain fully paid and the time is spent on other valuable work, report “capacity released” and show a scenario range for A; do not relabel it as cost reduction.

When a clinic has an actual avoidable expense, such as paid overtime or a temporary answering service that can be reduced, calculate only that observed expense as S. Name the line item, the period, and the evidence that it changed. Avoid a second benefit for the same time: if saved receptionist minutes are already valued through reduced overtime, do not also count all those minutes as full productivity value.

A worked example with clearly hypothetical inputs

Suppose a practice samples 120 eligible calls over four weeks. In a baseline period, complete handling averages 5.5 human minutes per call. During a constrained pilot, the same definition includes a mean 2.2 minutes of review, repair, and follow-up per eligible call. The difference is 3.3 minutes. The arithmetic suggests 6.6 hours of capacity returned across the sample: 120 × 3.3 / 60.

Now suppose, purely for illustration, the practice enters a loaded staff cost of $32 per hour, a vendor and usage cost of $240 for the four-week period, governance and monitoring time valued at $80, and implementation cost of $500. If the practice makes no cash change, then demonstrable cash savings S are $0. The first-window cash contribution is 0 − 240 − 80 − 500 = −$820. The capacity-equivalent value before costs is $211.20; that does not make the pilot cash-positive. In steady state, after setup is no longer charged to the same window, the capacity-equivalent net before any redeployment factor is $211.20 − 240 − 80 = −$108.80.

The example is intentionally not a promise and not a market price. It shows why “minutes saved × hourly rate” is not enough. A buyer needs to identify whether those minutes are real, whether they were moved rather than erased, whether the time can be used, and what recurring costs remain. If the vendor cost was $100 instead, or the practice had a verified reduction in overtime, the result would change. The decision should be made with the clinic’s own invoices and observations, not with the illustrative values here.

How to choose a baseline that does not flatter the pilot

Use matched windows where possible: comparable weekdays, service hours, staffing levels, seasonal conditions, and appointment availability. If the pilot runs only on quiet evenings but the baseline includes Monday morning, the comparison is biased. If staffing changes, note it. If the clinic closes for a holiday, either match the period or exclude the days with a written rule decided before analysis.

Measure a representative sample of the workload, not only successful demonstrations. If call volume is too large to review fully, choose a sampling plan before the pilot: for example, a fixed random sample per day plus every critical failure, with risk events reported separately. The sample size should be decided with the clinic’s analyst or quality owner based on expected variability and consequences. A small sample can identify failure modes; it cannot establish a precise population effect. Never describe a 20-call convenience sample as an industry benchmark.

Record caller drop-offs and retries when the system makes that data available and permitted. An “answered call” metric can increase while the caller repeats the task later, uses another channel, or waits for a human. Those are not necessarily visible in the telephony log. The practice can sample callback records, appointment change logs, and staff notes to identify such work, while minimizing personal data and following its legal obligations.

Time is not the only cost

Include recurring vendor fees, usage overages, phone-number or telephony costs, calendar connector charges, after-hours routing, transcription or storage options, maintenance, support tiers, training, staff review, and the clinic’s own time spent on policy updates. Include migration, integration, data mapping, testing, security review, procurement, and exit costs when evaluating the full lifecycle. Ask vendors to identify the unit behind every fee: per minute, call, concurrent session, appointment, user, connector, or monthly minimum.

Cost is not only money. A failure can consume trust, create rework, delay access, expose information, or create clinical confusion. A workload model should therefore have two layers: a financial comparison and non-negotiable service/safety gates. A positive arithmetic result cannot compensate for a failure to route a concerning call, unauthorized disclosure, or misleading appointment confirmation. Record such events as stop conditions with an owner, investigation path, and restart decision. A provider may describe service features and typical use cases in its own material—for example, the medical-practice voice-AI overview—but that page is a primary source only for what the provider says, not independent proof of savings, adoption, safety, or patient acceptance.

For technology and governance, PwC’s September 2026 discussion of operationalizing trust in AI agents emphasizes clear authorization, monitoring, escalation, and human accountability, and argues that oversight needs to continue as systems and operating conditions change. This is a consulting perspective, not a controlled study of clinic reception or a numerical ROI estimate. It supports a practical budgeting choice: reserve time and cost for monitoring and escalation rather than treating go-live as the end of the work. PwC, As AI agents take on more work, trust should become operational

For a security baseline, CISA’s Secure by Design guidance is useful for asking whether the provider makes secure configuration and safe defaults part of its product responsibilities. It does not certify a medical phone workflow and does not replace the clinic’s privacy, procurement, or regulatory review. Use it as a procurement prompt, not as a healthcare compliance certificate. CISA, Secure by Design

A decision model with hard gates

The economic result is only meaningful after minimum gates pass. Set these with the responsible practice owners before the pilot. At minimum, specify: the allowed call types; the human route for out-of-scope or uncertain requests; what information may be collected; who can see the record; what constitutes confirmed booking; and how a failed or partial system write is resolved.

Then set a quality gate. Track wrong appointments, duplicate appointments, missing callbacks, unverified “confirmed” statements, failed human transfers, privacy events, clinical-boundary events, and caller confusion. Define which event forces a pause or shutdown before seeing vendor results. A severe event should trigger investigation, not be hidden by a high average completion rate. This is consistent with the general risk principle that frequency and consequence must be reported separately: one serious failure may matter more than many successful routine calls.

Only after gates pass, compare the baseline and pilot ledger. Separate results by call type and time window. Report both the median and tail repair time when the distribution is skewed; a few difficult calls can dominate staff burden. Keep unresolved calls as a category. Do not remove them because the model classified them “out of scope” after deployment unless the original scope definition already did so.

The 4-week pilot protocol

The following schedule is a proposed evaluation design, not a proven universal duration. Extend it if the clinic has too little volume to obtain a representative sample or if the workflow is seasonal. A short calendar duration does not guarantee a valid sample.

Before week one, write the eligible-intent list, the excluded intents, the baseline measurement definitions, privacy minimization plan, human escalation route, outage fallback, owner, stop conditions, and vendor pricing units. Save the policy version. Review the vendor’s data flow and contractual arrangements with the appropriate clinic staff. If health data is hosted or processed, determine applicable requirements for the jurisdiction and service design; do not infer that a vendor’s badge settles the question.

Baseline period: observe the current process without changing staffing or call routing. For each sampled call, measure end-to-end human minutes, including callbacks, corrections, and duplicated handling. Capture whether callers get a clear next step. Record the same call categories that will be used in the pilot. Do not copy unnecessary clinical details into the economic dataset.

Constrained pilot: enable only the narrow eligible category, with a transparent caller notice if required by the clinic’s policy or applicable law. Keep a named human fallback available. Review all critical events and a predefined sample of routine events. Verify appointment changes in the authoritative scheduling system. If a connector cannot establish whether a write succeeded, do not announce success; reconcile the state and inform the caller through an approved route.

Review and decision: compare like with like, list changed operating conditions, calculate capacity separately from cash, and show low/base/high scenarios only for assumptions that are genuinely uncertain. The practice owner, receptionist lead, clinician responsible for safety, and technical owner should each sign their part of the decision. A vendor should not be the sole party defining success.

Sample review sheet

The short template below is a working artifact. Use the same row definitions across both periods; do not treat the example as actual observations.

period,call_id,intent_class,eligible,expected_endpoint,actual_endpoint,
record_verified,caller_next_step_clear,human_minutes,repair_minutes,
review_minutes,failure_class,critical_event,policy_version,notes_without_clinical_detail
baseline,random-001,appointment_request,yes,callback_queued,callback_queued,
yes,yes,4.0,0.0,0.0,none,no,v1,illustrative-only
pilot,random-002,appointment_request,yes,slot_confirmed,uncertain,
no,no,1.0,2.5,1.0,partial_write,yes,v2,illustrative-only

For an actual spreadsheet, calculate the eligible call denominator from a locked rule. Count duplicate follow-up contacts as additional work linked to the original random ID, not as a new “successful” call. Keep identifiers pseudonymous, access restricted, and retention limited according to the clinic’s approved plan.

Reusable scenario matrix

Run scripted simulations and, where approved, safe shadow observation for the following cases. These are test scenarios, not statistics. The list is deliberately designed to expose hidden costs and make success falsifiable.

For each scenario, record the expected disposition, the actual system state, the caller’s wording of the next step, the human minutes, and any failure. Re-run a test after changing the calendar connector, clinic policy, phone routing, model, or hours. A one-time acceptance test says nothing about later configuration drift.

The break-even table

How to prevent double counting

Make a one-page benefit register before computing a return. Each benefit needs a source, owner, measurement period, and a note on whether it is cash, capacity, quality, access, or a non-monetized outcome. If the practice counts minutes saved as improved capacity, it should not also count the same minutes as a cash saving. If the automation’s callback is counted as completed work, the human callback minutes still belong in the handling total. If an avoided no-show is included, verify it was caused by this workflow and not by a simultaneous reminder change.

Use a counterfactual that the practice can explain. “What would have happened without the pilot?” may be estimated from a matched baseline, but it is still an inference. State it as such. If the clinic changes opening hours, adds a receptionist, changes its booking rules, or deploys a new online scheduler during the same period, the result is confounded. Either isolate the change or report that attribution is uncertain.

What current research can and cannot tell you

The evidence for medical voice systems is not a single pool. A study of speech-to-text documentation, a chatbot trial, an AI scribe evaluation, and an automated phone-reception workflow answer different questions. Do not borrow an effect size from documentation to predict call-handling savings. A 2025 systematic review of AI-powered voice-to-text documentation in primary and outpatient care is an adjacent evidence base; it concerns documentation, not practice phone reception, and its results should not be treated as a direct estimate of booking performance. PubMed, systematic review of AI-powered voice-to-text clinical documentation

Likewise, a study of human receptionist calls helps identify work and patient burden, but it cannot certify an AI’s safety, reliability, or financial benefit. Vendor case studies may help generate hypotheses, but they are not a substitute for the clinic’s own denominator, repair minutes, and system-of-record checks. The most honest conclusion may be “the test was too small to decide.” That is a useful answer: it prevents a premature expansion based on noisy evidence.

For broader AI measurement context, Stanford HAI’s 2026 AI Index explains that its purpose is to collate and visualize data about AI progress and impact, while noting concerns about transparency and the need for independent measurement. It is not a study of clinic phone reception. That distinction reinforces the local-test rule: macro adoption or capability trends cannot substitute for a clinic’s own cost and outcome evidence. Neither the AI Index nor CISA’s guidance provides a ready-made financial model for a medical phone queue. Adapt their broad principles to a concrete workflow and have qualified local professionals validate healthcare obligations; do not present a general AI framework as healthcare approval.

Privacy and procurement costs

The workload ledger can become a shadow patient file if it copies names, symptoms, transcripts, or appointment details without necessity. Before collecting evaluation data, map the data path: telephony, model provider, transcription, storage, analytics, scheduling system, and support access. Ask where each field goes, who can access it, how long it persists, whether it is used to train a model, and what happens at contract end. Make the practice’s privacy lead or designated professional part of the review.

Cost the controls, not just the monthly licence. Staff training, policy maintenance, access reviews, vendor changes, incident investigation, audit exports, and periodic scenario retests require real owner time. Ask for an exportable event log that can answer: what the system was allowed to do, what it attempted, what the scheduling system recorded, where a human intervened, and which version of the routing policy was active. If a supplier cannot show evidence, list the gap as an operational dependency rather than assuming it is free.

Failure modes that turn an ROI claim into a cost transfer

Answer-rate substitution. The dashboard counts calls answered, but callers repeat their request or staff spend longer repairing it. Mitigation: track the linked disposition and downstream minutes.

Human review omitted. A manager checks summaries off the clock or the receptionist corrects calendars later. Mitigation: include review and repair time in the pilot ledger.

Uncertain write reported as success. The call ends with a booking statement while the scheduling connector has timed out. Mitigation: distinguish requested, held, confirmed, and unknown; require system-of-record verification.

Hard calls disappear from the denominator. The system labels difficult calls “out of scope” after the fact. Mitigation: freeze eligibility rules before the pilot and report exclusions.

Free capacity treated as cash. Staff time is valued at an hourly rate even though payroll and overtime do not change. Mitigation: report capacity separately and name any actual expense reduction.

Scope drift. A routine-hours system gradually handles symptoms, prescription questions, or sensitive identifiers. Mitigation: version the policy and make out-of-scope escalation a release gate.

Repeat calls counted as separate successes. A caller contacts the practice again because the first answer was ambiguous. Mitigation: sample callbacks and link permitted records through a random study ID.

Vendor-specific assumptions become permanent. A model, integration, fee schedule, or support limit changes after the pilot. Mitigation: identify change-notification duties, review windows, and an exit path in procurement.

How to make the result auditable

Keep a compact proof pack: the approved scope; the baseline and pilot definitions; the sampling plan and dates; the tariff and invoice inputs; the time observations; the exception and incident log; the policy versions; the appointment-system verification rule; the written assumptions; and the signed decision. Every chart should be reproducible from this pack. Do not publish a percentage without its numerator, denominator, period, and exclusions.

Separate facts, inferences, and recommendations. A fact might be: “the scheduling system showed 42 confirmed actions in the reviewed sample.” An inference might be: “the pilot likely reduced manual entry for this call category.” A recommendation might be: “extend the test to another matched window.” Keeping those labels visible helps a practice avoid turning an attractive hypothesis into a claimed outcome.

If the vendor reports a high completion percentage, request the underlying definition and raw event categories. Ask whether unanswered, abandoned, transferred, uncertain, duplicate, and corrected calls are included. Request a reconciliation sample that the clinic independently selects. Do not accept a metric if the provider controls both the test set and the success label without independent verification.

Decision tree

flowchart TD
    A[Define one eligible call category] --> B{Clinical or uncertain request?}
    B -- Yes --> C[Use approved human route; exclude from automated completion]
    B -- No --> D[Measure matched baseline including callbacks and repairs]
    D --> E[Run constrained pilot with caller-safe fallback]
    E --> F{Critical service or privacy gate failed?}
    F -- Yes --> G[Pause, investigate, correct, retest or stop]
    F -- No --> H[Verify system-of-record dispositions and human minutes]
    H --> I{Cash expense demonstrably changed?}
    I -- Yes --> J[Compare verified cash change with recurring and setup cost]
    I -- No --> K[Report capacity returned separately from cash]
    J --> L[Decide with uncertainty and exit cost visible]
    K --> L

Go, extend, or stop

Go only when the permitted workflow is narrow, the human fallback works, the record is correct, no critical gate is open, and the benefit can be reproduced from the evidence pack. A “go” can still be limited to specific hours or call types. Avoid expanding scope merely because the first pilot worked on routine cases.

Extend when no critical issue appears but the sample is too small, the cost inputs are incomplete, or results vary substantially by call category. State what additional evidence is needed and the maximum duration or spend authorized for that extension. A pilot that cannot produce a decision after repeated extensions should be stopped or redesigned.

Stop or constrain when the agent makes misleading confirmations, mishandles a caller seeking human help, collects information the practice did not authorize, creates unverified system writes, or generates more repair work than the clinic can safely manage. Preserve the evidence, disable the unsafe action, notify the accountable owner, and decide whether a corrected release deserves a new test. Do not quietly relabel a failure as an edge case after deployment.

Limitations of this model

This model does not estimate clinical outcomes, patient satisfaction, national savings, or legal compliance. It does not prescribe one sample size or one acceptable error rate. Those depend on the workflow, patient population, jurisdiction, service promise, clinical risk, and available fallback. The worked arithmetic uses hypothetical inputs only. The DREES material is historical, the receptionist studies are small or context-specific, and governance recommendations from consultancies are not causal evidence that a specific product works.

The method is deliberately conservative about cash. Some practices may reasonably value capacity even when payroll is unchanged, particularly when returned time supports care coordination, patient access, or staff wellbeing. Those benefits may be important, but they should be described in their own terms and measured with suitable indicators. They should not be forced into a false dollar amount just to justify a vendor purchase.

Checklist before signing

• [ ] One bounded call category and an explicit exclusion list are written down.
• [ ] “Answered,” “completed,” “confirmed,” “repaired,” and “uncertain” have separate definitions.
• [ ] Baseline and pilot periods are comparable, with known staffing and calendar conditions.
• [ ] Human follow-up, review, callback, correction, and duplicate work are counted.
• [ ] The system-of-record is the source of truth for appointment changes.
• [ ] Clinical questions and uncertainty route to a real, staffed, approved path.
• [ ] Privacy and data-flow review covers recording, transcription, storage, connectors, support, and deletion.
• [ ] One-time, recurring, usage, training, monitoring, and exit costs are included.
• [ ] Released capacity is not described as cash savings unless cash actually changes.
• [ ] The model avoids double-counting and reports exclusions and uncertainty.
• [ ] Stop conditions, incident ownership, version changes, and restart criteria are agreed.
• [ ] The decision is signed by the practice owner, frontline staff representative, clinical lead, and technical owner.

Source and method note

The original asset here is the call-disposition workload ledger, break-even formula, and scenario matrix. They are a proposed evaluation tool, not a dataset or experimental result. No clinic calls were collected for this article; no vendor was benchmarked; the illustrative numbers are fictional; and no performance rate or ranking claim is made. Sources support only the nearby claims: DREES describes a historical French general-practice work schedule; the receptionist studies describe context-specific human work; CISA, Stanford HAI, and PwC inform security, measurement, and governance questions; and the vendor page is a primary source for its own stated service, not independent proof of outcomes. Recheck program, legal, and vendor details before applying the method in a live practice.

The answer

An AI phone receptionist is worth considering only when the practice can show which work it removes, which work it transfers, and what it costs to keep the system safe and current. Start with the disposition ledger, count the human minutes after the call, verify the schedule rather than the transcript, and keep service gates separate from ROI. The most trustworthy outcome may be that the tool is useful for one narrow queue, only after-hours, or not at all. That is still a successful evaluation: it gives the practice evidence to choose without confusing a fluent answer with a finished task.