Avie CAIHL draft report

Avie: patient-agency and patient-directed AI review

Research conducted 17 October 2026 PDT / 18 October 2026 UTC. Published 18 October 2026 UTC.

Status: Published AI-assisted draft. Mixed agency posture, clinic-directed and patient-facing. Unscored, with no numerical HugoScore assigned. Publication authorized by Hugo Campos, without a claim of comprehensive human verification.

Reviewed entity: Avie at avie.me, operated by AVIE TECNOLOGIA LTDA in Brazil. This is not a review of unrelated products sharing the Avie name.

Executive judgment

Avie belongs within HugoScore’s scope. Its documented design is patient-facing and clinic-directed. It could materially improve patients’ practical ability to arrange care, but the evidence does not establish patient-directed AI.

The strongest positive case is reduced administrative burden: patients can express scheduling preferences and complete or change transactions without waiting for reception to open. The consequential question is whether that convenience also gives them meaningful power to refuse, question, correct, or redirect the system. Those capabilities are much less clearly established than the clinic’s controls.

Five conclusions deserve attention:

• 1. Administrative does not mean inconsequential. Interpreting a request, deciding whether it reaches a person, selecting appointment options, and applying confirmation rules can change access to care. These are sufficient patient-impact pathways without assuming autonomous diagnosis.
• 2. Human oversight is well described for staff. Patient-initiated recourse is not equivalently documented. A clinic’s ability to take over does not establish a patient’s ability to demand a person or decline AI without disadvantage.
• 3. AI disclosure deserves a specific challenge. Public help describes human-like delivery pacing partly in terms of avoiding a robotic impression. The actual notice shown to patients, and whether clinics must retain it, remain unverified.
• 4. There is operational evidence, not established evidence of patient empowerment. Avie publishes small, vendor-run clinic measurements. These are more informative than unsupported performance slogans, but cannot establish patient understanding, equitable access, safety, or successful contestation.
• 5. Keep product boundaries explicit. The submitted receptionist should be the main unit of review. Optional payment/EHR workflows require labeled analysis. Avie Notes is a newly documented beta clinical-documentation feature and should receive a separate boundary and evidence assessment rather than being silently folded into reception.

Publication decision: In scope. Published as an unscored, clinic-controlled / mixed-agency qualitative review. Documented safeguards and practical choice receive credit. Unresolved features remain unverified, not absent or automatically harmful. This review is not an endorsement or a compliance verdict.

1. Method, evidence, and boundaries

This review read Avie’s product, pricing, integration, workflow, help, privacy, terms, deletion, security, and measurement materials, alongside HugoScore’s public method and scope test. Sources were accessed 18 October 2026 UTC. Page-update dates are recorded when available. Some pages were recoverable through search-index extraction when direct retrieval failed. Search captures can lag, so documented version differences are retained rather than flattened into one supposedly timeless product.

Four evidence labels are used:

• Published claim: What Avie says the product or policy does. This establishes a stated design or commitment, not successful implementation.
• Vendor observation: Measurements Avie reports from its own operation. These are not independent validation.
• Inference: The review’s reasoned assessment of an agency mechanism or risk.
• Unknown: A question the reviewed evidence does not resolve. This does not prove the capability is missing.

No account was created, conversation initiated, form submitted, booking attempted, or real patient information supplied. Public examples were treated as demonstrations, not audited patient outcomes. No source establishes that all clinics configure Avie identically.

Product units

Avie Atende receptionist

Review boundary: Main submitted deployment: conversational administrative responses and connected booking actions

Why it matters: Direct patient interaction and access to care

CRM outreach

Review boundary: Include the clinic-enabled reminders, follow-up, waitlist and reactivation attached to reception

Why it matters: AI can influence what patients are asked to do and when

Avie Cuida payments/EHR

Review boundary: Analyze only when enabled. Do not attribute all features to Atende

Why it matters: Financial confirmation rules, records, summaries and document transmission

Avie Notes beta

Review boundary: Separate supplementary assessment

Why it matters: AI-generated documentation can alter how the patient’s account enters the clinical record

Avie Cresce marketing services

Review boundary: Incentive/context evidence, not automatically an AI clinical feature

Why it matters: Provider acquisition goals may shape the surrounding workflow

Atende is publicly priced at R$1,350 monthly, plus R$2,700 implementation, for up to 1,000 patients monthly. Cuida and Cresce are quoted separately. The purchasing relationship is with the clinic. There is no basis here to infer a separate patient subscription. Pricing

2. HugoScore scope and method

The public scope test has three gates: model judgment or estimation, an effect on health, health data, care, coverage, or a health decision, and an identifiable route to what a patient understands, decides, or can do. Institutional AI used on patients qualifies. Commercial or regulatory status is not an entry requirement. Scope test

Model judgment

Application to the reviewed deployment: Avie describes interpreting conversation, assembling context and generating or routing responses

Decision: Pass on documented product claims, architecture not independently verified

Health relevance

Application to the reviewed deployment: The output mediates access to clinical services and information

Decision: Pass

Patient-action pathway

Application to the reviewed deployment: Patient request → interpreted preferences or issue → options, booking/cancellation, or staff referral

Decision: Pass

A deterministic calendar alone would not qualify simply because a clinic uses it. The relevant unit is the AI-mediated interaction and the actions it influences. Notes is not needed to make the receptionist pass the test.

HugoScore’s published questions address who benefits and controls the tool, what patients can see, their ability to choose/question/correct, data use, and evidence. The axis is a direction from institutional control toward patient choice, not a safety or quality grade. Review method, About

The current BastionGPT report supplies useful methodological discipline: patient authority, critical capacity, and informed control are distinct lenses. Stronger privacy does not itself transfer authority. Unknowns do not automatically receive negative points. Its profile says draft v1.2 numerical anchors remain unadopted. That supports qualitative assessment here rather than a fabricated calculation. Report, Profile

The current unscored CareClinic profile offers a more directly relevant precedent for representing an evidence-linked qualitative posture without inventing a new numerical position. Unscored must not be rendered as zero. CareClinic profile

The receptionist passes all three scope gates. The public submission log records the scope decision and links to this AI-assisted draft.

3. Who holds authority?

“Patient-facing,” “patient-aligned,” and “patient-directed” answer different questions. A person can interact with an AI that primarily executes someone else’s policy. Conversely, institutionally purchased software can genuinely expand the person’s practical choices. Evaluate those effects rather than treating either ownership or friendliness as decisive. CAIHL

Select and purchase the system

Documented holder: Clinic

Patient-agency interpretation: Patients are not established as choosing the AI used for this encounter

Set factual information, conversational rules, assistant name and tone

Documented holder: Clinic, initially with Avie

Patient-agency interpretation: Institutional framing is explicit. Patient preferences operate within it

Choose among offered appointments or request alternatives

Documented holder: Patient, subject to clinic schedule and workflow

Patient-agency interpretation: A genuine but bounded action capability

Decide permitted clinical scope

Documented holder: Product-level limits, with clinic-defined routing rules

Patient-agency interpretation: Protective boundaries need not be patient-controlled to be beneficial

Review, rewrite, stop, or resume AI replies

Documented holder: Clinic staff

Patient-agency interpretation: Strong staff correctability, distinct from patient recourse

Enable outreach, payment requirements, and workflow options

Documented holder: Clinic

Patient-agency interpretation: Defaults can change access and pressure without a separate patient configuration interface

Correct or export derived patient context directly

Documented holder: Not established

Patient-agency interpretation: Do not confuse a chat history or legal request route with direct control of inferred records

Refuse AI and preserve equivalent access

Documented holder: Not established

Patient-agency interpretation: Central unresolved patient-authority question

Clinic control is unusually explicit in the help page: facts and conduct rules are editable, including when to persist, when to stop, cancellation policy, and the assistant’s chosen name. It recommends reusable rules rather than ad hoc system exceptions. This is evidence of who governs the workflow, not proof that clinicians cannot grant individual accommodations. Rules and tone

4. Evidence and claim matrix

All implementation statements below are published claims unless labeled otherwise. The right column records what remains to be demonstrated.

Scheduling preferences

Published evidence: Interprets requested dates/periods, searches real availability, offers alternatives, briefly holds a chosen slot. Scheduling help

Agency reading and unresolved point: Promising practical choice. Need evidence that constraints and omitted options are explainable and that misunderstood preferences can be corrected.

Connected actions

Published evidence: Google Calendar and Feegow supported, availability checked again before confirmation, conflicts referred to staff. Other direct management-system integrations are not claimed. Integrations

Agency reading and unresolved point: Transaction checks reduce a plausible error route. No observed reliability rate, recovery time, or patient-visible action ledger was established.

Price visibility

Published evidence: Clinic-written explanation includes price before registration and chosen-slot confirmation. Pricing

Agency reading and unresolved point: Positive sequencing: cost is available before collecting identity details. Test whether a patient asking only for price can obtain a concise answer without repeated sales framing.

Cancellation

Published evidence: WhatsApp cancellation is described. The clinic website also supports confirmation/cancellation links. Physician product page, Pricing

Agency reading and unresolved point: Potentially important exit power. Actual clicks/turns, confirmation receipts, fees and refunds were not tested.

Staff takeover

Published evidence: Manual staff response pauses AI until staff release it. Approval mode permits editing or rejecting drafts. Sensitive/invalid actions still require review in automatic mode. Review and takeover

Agency reading and unresolved point: Credit staff control. A patient saying “I want a person” is a different capability and needs its own evidence.

Clinical boundaries

Published evidence: Help describes fixed prohibitions on diagnoses, medication guidance and symptom-severity assessment. Complaints, emergencies, sensitive topics and commercial exceptions route to staff. Limits

Agency reading and unresolved point: Appropriate stated separation of roles. Recognizing which message requires escalation is still safety-relevant classification.

After-hours handling

Published evidence: Staff-dependent cases wait for opening, with the patient told not to expect immediate human return. After-hours help

Agency reading and unresolved point: Acknowledgment is not resolution. Need coverage rules and visibility of when a person actually accepts the case.

Outreach controls

Published evidence: Clinic-enabled programs include reactivation and waitlist offers. Sending limits, quiet hours and stop conditions are described. CRM

Agency reading and unresolved point: Useful anti-spam controls. Patient choice by message category, withdrawal propagation and consequences of declining outreach remain unclear.

Patient stop request

Published evidence: Homepage examples exclude people who asked not to receive certain messages. Homepage

Agency reading and unresolved point: This is positive published evidence. Do not say there is no opt-out. It does not establish a universal AI opt-out or granular consent management.

Payment-dependent confirmation

Published evidence: Optional “paid, confirmed” workflow does not accept a verbal confirmation as sufficient. Unpaid cases go to staff to retain/release the appointment. Refunds occur through the clinic’s provider. Payments

Agency reading and unresolved point: Access can become conditional on payment. Do not describe slot release as an autonomous AI decision when staff decide. Need hardship, delayed-payment and dispute routes.

Provider incentives

Published evidence: Reports track new contacts, booking conversion, absences, cancellations and payment information. Reports

Agency reading and unresolved point: Measures center clinic operations. No corresponding patient-defined success metrics were established. That is an evidence gap, not proof of improper intent.

Sensitive registration

Published evidence: A physician-facing page describes collecting name, birth date, CPF and address before confirmation. Physician product page

Agency reading and unresolved point: Ask which fields are necessary for which service and what happens if a patient cannot or does not want to supply them through chat.

Patient context

Published evidence: Help describes stored conversation-derived profiles, response-construction records and audio transcription. Its newer October 3 version also describes AI image descriptions/text extraction, while clinical interpretation is excluded and PDFs remain staff attachments. Data help

Agency reading and unresolved point: Context is being created about patients, not merely scheduling metadata. Patient inspection, provenance, correction propagation and onward export require verification.

Clinic records

Published evidence: Signed entries retain the original with subsequent rectification. Named access/change logs and separate user accounts are described. WhatsApp summaries sit outside the consultation note. EHR

Agency reading and unresolved point: Credit provenance and traceability. Staff audit access does not establish patient access to logs or to disputed AI-derived summaries.

Data roles and model providers

Published evidence: Current security page identifies the clinic as controller, Avie as processor, and OpenAI/Google as AI providers. Security

Agency reading and unresolved point: More specific than “we use AI.” Exact service tiers, models, data locations, subprocessors and contractual processing restrictions remain important.

Privacy rights

Published evidence: Published policy allows requests for access, correction, portability, information and deletion, subject to limits. Privacy

Agency reading and unresolved point: There is a rights route. It is not evidence of self-service export, a defined response deadline, or patient control over every inference.

Deletion

Published evidence: Public process may involve the clinic, promises a reasonable period with legal/security/audit exceptions. Deletion

Agency reading and unresolved point: Not an unconditional erase button. Request completion, backups, derived data and legal holds need a precise account.

AI disclosure

Published evidence: Vendor-facing pages label the product AI. Patient dialogue examples do not establish a mandatory notice in every deployment. Homepage

Agency reading and unresolved point: Notice must be checked where the patient actually encounters the system.

5. The most consequential patient-agency issues

5.1 Administrative access is substantive agency

Reducing dependence on office hours could matter greatly to someone with an inflexible job, caregiving responsibilities, or difficulty making phone calls. An understandable cancellation path can be more empowering than a sophisticated explanation that leaves a person stuck with an unwanted appointment.

The right outcome is not simply that a transaction completed. Ask whether the patient achieved their chosen outcome with less burden: booking, postponing, declining, obtaining information, disputing an error, or reaching a person. A system that improves booking while obstructing cancellation would have an asymmetric agency effect.

This is a causal hypothesis grounded in the workflow, not a measured Avie outcome. It should be tested with patients rather than inferred from appointment volume.

5.2 “Human in the loop” is incomplete unless patients can use it

Staff review is a credible safeguard against mistakes. It does not answer who has power when the patient disagrees. A usable recourse pathway needs:

• an explicit patient-triggered route, including when the AI considers the issue routine.
• a receipt explaining whether the request was queued or accepted.
• realistic hours and response expectations.
• a fallback when the person never responds.
• authority for that person to correct the consequential action.
• confirmation that the correction reached the calendar, CRM, reminders and relevant record.

No inference of a guaranteed human response should be drawn from a pending item appearing in a staff dashboard. Nor should after-hours availability be advertised as equivalent to around-the-clock clinical support.

5.3 Disclosure and anthropomorphic presentation

The September 2 after-hours help explains configurable human-style typing partly by saying an immediate long reply reveals a robot and changes the perception of care. This creates a documented transparency concern, not proof of deception in every live clinic. After-hours help

Friendly tone and readable pacing are not inherently problematic. The issue is whether people can accurately understand whom they are interacting with, who set the rules, and where the system’s authority ends. A product openly marketed as AI to purchasers can still be opaque to the people subjected to it.

Useful evidence would include a reliably presented patient-facing AI notice that clinics cannot remove or make misleading, a clear answer to “Are you a person?”, and a reliable route to a human. Ask whether disclosure survives a clinic changing the assistant’s name or script.

5.4 Revenue goals can align with access, or conflict with it

Filling a canceled slot can help both patient and clinic. A reminder can prevent an unwanted missed visit. Neither benefit requires assuming that maximizing appointment conversion is the patient’s objective.

Potential conflict appears when a patient wants to defer, compare providers, decline outreach, challenge a charge, or describe a constraint the standard workflow does not accommodate. Review those cases specifically. Do not infer actual upselling of unnecessary treatment: the reviewed evidence does not establish it.

For waitlists, “in order,” matching a slot, timed offers and first acceptance can produce different allocations. The public descriptions do not sufficiently resolve the operative selection and hold rules across configurations. A fair review should ask who gets overlooked and whether urgent or accessibility needs are handled through an accountable human process.

5.5 Confirmation and cancellation can become gatekeeping

A missed message is not necessarily a changed intention. Connectivity loss, shared devices, literacy difficulties, work shifts or disability may prevent a rapid reply. Payment processing can also fail after a person acts in good faith.

Evaluate the entire chain: advance notice → deadline → alternative response method → exception request → staff decision → appeal → restoration or refund. A technical payment-status error must not silently become a judgment about the patient’s willingness to attend. The payments page acknowledges staff manual reconciliation and late-payment intervention. The test is whether patients can reliably reach that remedy.

5.6 Clinical routing remains safety-relevant

The receptionist does not need to diagnose to affect clinical access. A false “routine” classification could delay a request, while an inappropriate escalation might disclose information more widely or create unnecessary distress.

Avie’s mental-health workflow describes clinic-specific handling of distress, including urgent staff alerts, emergency redirection when an emergency is happening, and avoiding promises of immediate clinician response. That is richer than a generic “human escalation” claim, but one configured workflow cannot establish general emergency performance. Mental-health workflow

Needed evidence includes missed-escalation cases, false escalations, coverage after hours, acknowledgment and resolution times, and whether patients are told an alert has merely been sent. This review does not validate the described emergency process or give clinical advice.

5.7 Equity and accessibility need deployment evidence

Audio input followed by written response may help some people and fail others. Portuguese-language fit is valuable for the intended market, but no multilingual equivalence or accessibility evaluation was established. WhatsApp familiarity is not the same as universal access.

Tests should include screen-reader use, low literacy, voice transcription errors, slow replies, unstable connectivity, nonstandard Portuguese, shared phones, caregiver permissions, and a person who cannot use WhatsApp. Compare successful actions and abandonment across these situations. Do not assume that a clinic booking website is an equivalent human alternative.

6. Evidence of performance: what the data can support

Índice Avie v3, updated October 17, 2026, reports medians for the first reply from AI or staff during AI-enabled operation: 49 seconds in 283 pediatric conversations and 58 seconds in 104 adult mental-health conversations, over roughly three weeks. A pre-AI mental-health sample reports 2.8 hours in 185 conversations. These are vendor-run, nonrandomized observations.

Its method defines conversations using six-hour inactivity and includes first responses within 48 hours. Imported prior history is partial and final-day follow-up shorter. These restrictions affect unanswered/late cases and comparison. Generation time differs from patient-visible reply time. Older speed figures reflect earlier measurements. Índice Avie v3

Assessment: Suggestive evidence of faster initial contact in those settings. It does not establish causal improvement in access, correct resolution, emergency safety, comprehension, equitable treatment, or patient autonomy. Ask for denominators covering every eligible request, unresolved cases, and patient-reported outcomes. No independent validation was established in the primary materials reviewed.

7. Privacy, context, and control

7.1 What deserves credit

The October 17 security description names controller/processor roles, individual role-based access, clinic-separated databases, and EHR audit trails. It describes protection for cached application screens and says medical records are retained for the applicable long-term recordkeeping period, identified there as 20 years. These are specific vendor assertions, not an independent security audit. Security

The data help says patient content is not sold, published or shared for third-party commercial purposes. That is a meaningful stated restriction. It is not equivalent to an explicit prohibition covering every form of model training, evaluation, fine-tuning, retention or product improvement. Data help

7.2 What remains unresolved

• Processing map: Which content reaches which provider for reception, transcription, image processing, summaries and Notes? Which infrastructure and regions are used?
• Use restrictions: What contracts govern training, evaluation, human review and secondary use? Do the rules cover Avie itself, subprocessors, backups and de-identified derivatives?
• Retention by data class: Chat text, attachments, ordinary voice notes, transcripts, inferred preferences, prompts, logs and signed records are different objects. A rule for one must not be generalized to all.
• Patient correction: Can patients see and challenge inferred facts? Is a correction propagated into future prompts and downstream workflows, or merely appended somewhere staff might not consult?
• Portability: What can the patient obtain directly, in what format, how quickly, at what cost, and with what provenance? Clinic migration features do not establish patient portability.
• Identity and delegation: How are shared phones, children, caregivers and confidential communications handled? How are wrong-recipient links or attachments revoked?
• Incident response: No tested detection, notification, recovery or patient-remedy performance was established.

The privacy policy’s general retention language is need-based rather than a content-specific schedule. Its Instagram section gives more explicit limits for that separate channel. Those must not be assumed to govern WhatsApp or medical records. The terms place substantial content and access responsibilities on the clinic and acknowledge dependence on third-party services. Privacy, Terms

7.3 WhatsApp is one boundary in a longer chain

Using WhatsApp does not establish end-to-end confidentiality through Avie’s downstream systems and model providers. WhatsApp itself distinguishes personal messaging protections from business/cloud-storage arrangements and directs users to businesses’ own privacy practices. The exact configuration used by an Avie clinic was not tested. Do not claim that all WhatsApp traffic is unencrypted or that a messaging encryption label governs every subsequent processing step. WhatsApp privacy explanation

7.4 Rights are important, but rights and usable controls differ

A legal entitlement, a contractual commitment, an administrator setting and a patient-facing capability are four different kinds of evidence. A patient may be entitled to correction yet still face a burdensome process to make the correction matter before a consequential action.

Brazilian legal context should be treated as a baseline for verification, not a substitute for testing. This report makes no finding that Avie or any clinic complies with or violates applicable law. Public policies were reviewed, but clinic-specific executed agreements, patient notices, legal-basis assessments and actual deployment practices were not obtained.

Two current primary sources are especially relevant:

• LGPD: Health information has sensitive-data protection. Access, correction and qualified portability/deletion rights matter. Consent is not the only possible lawful basis, and cross-border processing is not automatically prohibited. Ask for purpose-specific bases and safeguards rather than inferring them from “LGPD” branding. Consolidated LGPD
• Brazil’s April 2026 patient-rights statute: Law 15.378 addresses active participation, accessible information, consent/withdrawal, confidentiality, second opinions, and record access/copy/rectification. Its duties concern healthcare providers and related entities. The software supplier’s legal position cannot simply be assumed identical. These are relevant clinic-workflow benchmarks, not automatic product credits or evidence of violation. Official statute

Recordkeeping also needs precision: Brazil’s record law addresses a minimum period measured from the last entry, with qualifications. It does not establish a universal deletion date for every chat, prompt or attachment. Law 13.787, Article 6

8. Separate boundary: Avie Notes beta

The October 17 Notes page describes recording/upload/text input, transcription, AI note drafting and clinician signature. Recording requires the professional to attest that the patient or responsible adult was notified. Audio is stated to last at most six hours. Clinicians can edit transcripts or discard generated material. It also describes clinician-controlled writing-rule adaptation from corrected notes and template extraction from prior notes. This is application-level adaptation, not proof of foundation-model training. Avie Notes

Agency assessment: This affects how a patient’s narrative becomes an institutional record. Notification is not the same as a demonstrated choice to refuse recording without losing care. Clinician review is valuable but does not establish patient review or control over the representation of their account. Short audio retention reduces persistence while potentially narrowing later verification opportunities.

Keep open: informed refusal/withdrawal, patient access to drafts and sources, role attribution, correction of summaries, whether disputed statements remain identifiable, retention of transcripts and drafts, uploaded-audio notice rules, and validation across accents, specialties and multi-person encounters. Do not attribute this beta to every receptionist deployment.

9. Implications for patient-controlled context

Patient-controlled context would preserve source material, separate patient statements and AI interpretation, record corrections, and keep the patient’s question and judgment central. The following design implications do not assume an existing Avie integration or an implemented patient-controlled context system.

Avie addresses the clinic’s operational side of an encounter. Patient-controlled context addresses the person’s continuity of understanding across encounters and tools. They could be complementary, but only if the institutional side can receive patient-authored context without silently converting it into an unchallengeable “fact.”

Five useful design lessons follow:

• 1. Preserve authorship. A patient statement, a caregiver statement, a clinician conclusion and an AI inference should remain distinguishable.
• 2. Make correction consequential. A corrected preference or disputed fact must affect subsequent use, not just remain in an isolated note.
• 3. Carry scope and permission. A patient-provided context packet should say what it may be used for, who may receive it, and what needs confirmation.
• 4. Keep a receipt of action. Patients should be able to see what was booked, changed, forwarded or inferred, and on whose authority.
• 5. Support exit and continuity. A useful patient copy must survive a clinic change or vendor change without depending on continued subscription by the institution.

The key interoperability question is therefore not “Can Avie store a patient summary?” It is: Can the patient inspect, correct, authorize and carry the context that Avie and subsequent clinic AI actually use? This research does not establish that capability, and a personal context document alone would not enforce it.

10. Open verification questions

Priority A: needed to substantiate patient authority

• 1. Please provide the exact first-contact AI notice patients receive, including a clinic-configured example. Is any AI disclosure mandatory and uneditable? What happens if a patient asks whether they are talking to a person?
• 2. Can a patient request a human at any time, decline further AI processing, and still receive equivalent service? Show the workflow, hours, expected wait and fallback if nobody accepts the request.
• 3. Show a complete cancellation and rescheduling journey, including the action receipt, fee disclosure, missed confirmation, payment delay and refund/dispute path. Which steps are automated, and which decisions require staff?
• 4. Can patients inspect and correct conversation-derived facts, preferences and summaries? Show how an accepted correction changes the next AI interaction and linked reminders/records.
• 5. What data can a patient export directly, including source messages, AI-generated summaries and corrections? In which formats and timeframe? What remains only available to the clinic?
• 6. What choices can a patient make separately for reminders, reactivation, birthdays, waitlist offers and AI interaction? How is withdrawal applied across programs and channels?

Priority B: data governance and deployment evidence

• 7. Provide a feature-by-feature data-flow map and subprocessor list, including exact OpenAI/Google services, regions, retention and contractual prohibitions on training/evaluation/secondary use.
• 8. Provide retention and deletion schedules for messages, attachments, voice notes, transcripts, model inputs/outputs, inferred context, audit logs, backups and records. Identify legal exceptions and how patients are informed when deletion is denied.
• 9. How are identity, caregivers, minors and shared phones handled? Can a patient request confidential routing or stop sensitive documents appearing in a shared chat?
• 10. Which security claims have independent evidence, with scope and dates? What are the incident-notification, recovery and correction processes? We are not requesting credentials or patient data.
• 11. For Notes, show the notice/refusal/withdrawal workflow, whether uploaded audio has the same gate, and what patients can inspect or correct. Clarify transcript/draft retention separately from audio and describe safeguards for cross-patient writing-rule adaptation.

Priority C: outcomes, inclusion and configuration

• 12. Provide counts of unresolved requests, missed/incorrect escalations, mistaken bookings, failed cancellations and time to human resolution, with denominators and configuration/model versions. What patient-reported outcomes have been measured?
• 13. How are waitlist matches, order, deadlines and competing responses handled? Which exceptions protect people with accessibility or connectivity constraints?
• 14. What accessibility, Portuguese-variant, transcription and shared-device testing has been done? What non-WhatsApp or non-AI route is available?
• 15. Which safeguards are nonconfigurable, which are defaults, and which depend on clinic choices? How are materially changed models/rules tested and communicated?
• 16. Have patients participated in design or governance? Can a clinic measure success as successful patient refusal, correction or cancellation as well as conversion and filled slots?

Request synthetic demonstrations and redacted evidence. No real patient data is needed to answer these questions.

11. A practical verification plan

These are reviewer-proposed operational tests, not tests that were performed or adopted HugoScore pass/fail rules. Any future demonstration should use a vendor-authorized sandbox, fictional identities and clearly synthetic clinical scenarios.

First encounter

Patient’s goal: Understand that AI is involved

Proposed verification criteria, not an adopted scoring rule: Clear notice before consequential disclosure, truthful identity answer

Human request

Patient’s goal: Talk to a person without explaining a clinical reason

Proposed verification criteria, not an adopted scoring rule: Reliable transfer, receipt, expectations and fallback

Refusal

Patient’s goal: Continue care without AI where feasible

Proposed verification criteria, not an adopted scoring rule: Documented alternative and no hidden penalty

Preference repair

Patient’s goal: Correct an inferred date, modality or constraint

Proposed verification criteria, not an adopted scoring rule: Corrected options and no stale preference later

Cancellation

Patient’s goal: Cancel without persuasion loops

Proposed verification criteria, not an adopted scoring rule: Timely cancellation receipt, explicit fee/refund status

Payment mismatch

Patient’s goal: Preserve access when payment status is wrong

Proposed verification criteria, not an adopted scoring rule: Accountable review, reconciliation and restoration/remedy

No response

Patient’s goal: Avoid accidental loss of a slot

Proposed verification criteria, not an adopted scoring rule: Clear rules, proportionate notice, accessible exception route

Stop outreach

Patient’s goal: End optional contact while keeping essential logistics

Proposed verification criteria, not an adopted scoring rule: Preference persists across applicable programs

Sensitive handoff

Patient’s goal: Reach appropriate support without false promises

Proposed verification criteria, not an adopted scoring rule: Boundary-respecting acknowledgment and traceable staff response

Wrong derived fact

Patient’s goal: Correct a profile or summary

Proposed verification criteria, not an adopted scoring rule: Visible provenance, dispute record and downstream correction

Portable copy

Patient’s goal: Leave with usable information

Proposed verification criteria, not an adopted scoring rule: Export with source/AI/authorship distinctions and corrections

Notes refusal

Patient’s goal: Decline recording without derailing the visit

Proposed verification criteria, not an adopted scoring rule: Recorded refusal and alternative documentation workflow

Measure task success, number of turns, elapsed time, patient comprehension, unmet requests and correction persistence. Repeat after configuration/model changes. Use diverse participants to test accessibility and burden. Do not equate one successful synthetic demonstration with validation.

12. Publication assessment

Scope: In scope, based on the receptionist’s own care-access pathway.

Posture: Clinic-directed, patient-facing, practical agency gains plausible, consequential patient control incompletely established.

Score: Unscored. No adopted numerical calculation was applied.

Overall research confidence: Medium: adequate official documentation, limited independent evidence. There is strong documentary support for what the pages state. Implementation, reliability, patient experience and net agency effects remain substantially less certain.

Publication status: Published as an evidence-linked AI-assisted draft with explicit caveats. A stronger assertion about disclosure, refusal, correction or real-world benefit needs additional evidence.

Avie’s described booking changes and cancellation functions can already confer genuine task-level agency. Clinic control can coexist with patient authority over individual goals. The unresolved questions concern the breadth, durability and contestability of that authority: whether patients can refuse, correct and redirect consequential uses, and carry their context forward. Friendly conversation, rapid replies and staff oversight alone cannot settle those questions.

Source register

All accessed 18 October 2026. Links above point to the claims they support. This register records notable dates and evidence limitations.

• HugoScore scope, method, about: public method pages consulted during the research. The Avie scope decision is recorded with this publication.
• CAIHL: conceptual source for separating audience, institutional deployment, alignment and agency, not an Avie validation.
• Avie company: Company, vendor identity, founding and early-clinic description. Not independently audited corporate history.
• Avie pricing, homepage, CRM, payments, reports, physician page: current product claims. Illustrative interface/dialogue is not live patient evidence.
• Integrations: page updated 3 October 2026.
• Help pages on scheduling, takeover, tone, after-hours and limits: dated 2 September 2026, some recovered through search-index extraction. Older metrics must not replace the October 17 measurement update.
• Privacy: updated 9 October 2026. Terms and deletion: updated 18 June 2026.
• Data help: newer opened text dated 3 October includes AI processing of images. Older search copies dated September 2 are less specific.
• Security and Notes: updated 17 October 2026, full relevant text recoverable through search-index extraction despite direct-fetch errors.
• EHR: current retrieved product text includes Notes beta. An older search capture says the clinician writes notes without AI. Treat this as version lag requiring careful dating, not proof of misconduct.
• Índice Avie v3: updated 17 October 2026, operational period through October 16. Vendor observations, not independent evaluation.
• Mental-health workflow: published 3 October 2026, one clinic configuration and older v2 metrics, not a universal protocol.
• WhatsApp privacy explanation: official messaging-platform context, does not independently establish Avie’s technical configuration.

Discrepancies that need precise handling

• Recording retention: The current EHR FAQ describes audio deletion after transcription, while its body describes a review window up to six hours. The detailed Notes page uses the latter account. These could be compatible shorthand, but the exact trigger, maximum duration and treatment of other copies need confirmation. EHR, Notes
• Images: Broad wording on the physician-facing page about not reading or evaluating photos should not be used to infer that no image reaches a model. The more specific October 3 data help describes image description/text extraction without clinical evaluation. Scope and version should be clarified. Physician page, Data help
• Waiting-list and slot holds: Public descriptions refer to ordering, matching, first acceptance and short reservations. After-hours help describes a hold that can span turns overnight. These may be configuration-specific. Do not assert a universal allocation rule or hold duration without verification. Homepage, Scheduling help, After-hours help

No medical, legal, security-certification or purchasing recommendation is made by this report.

Review provenance

criteriaSource: CAIHL-derived HugoScore framework and current qualitative review method. Draft numerical rubric anchors are not adopted.

reviewer: AI-assisted public-source research, adapted for public HugoScore publication

model: OpenAI GPT-6 Astra (gpt-6-astra)

humanReview: Hugo Campos authorized publication of the completed review. Publication authorization is not comprehensive human verification.

reviewDate: 2026-10-18

limitations: Public-source documentary review. Some sources were recovered through search-index extraction. No account, conversation, booking, health-data upload, clinical evaluation, security audit or vendor interview. Vendor observations are not independent validation. Configuration-specific behavior and patient controls remain unverified.

Publication adaptation: Private addressee/context framing was replaced with public language. Tables were converted to readable findings, and the approved publication decision and model provenance were added. The source-linked findings, limitations and qualitative assessment are retained.