Responsible AI & Clinical Safety

01The core principle

Porvio assists with documentation. It does not practise medicine or dentistry. Every AI-generated clinical output, including a transcript, note, task, code, document, or evidence summary, is unverified assistive content until qualified human review.

The treating professional remains responsible for the encounter, medical or dental record, decisions, communications, coding, and care. Porvio is designed to make review explicit rather than hide AI involvement.

02Intended and excluded uses

Intended support

• transcribe clinician-controlled encounter audio;
• organize a structured, editable documentation draft;
• suggest follow-up tasks and possible codes for individual review; and
• synthesize a draft answer from retrieved PubMed records.

Not intended for

• emergency monitoring or communication;
• autonomous diagnosis, triage, treatment, prescribing, or ordering;
• automatic billing, claims, referrals, messages, or EMR write-back;
• deciding patient eligibility, coverage, employment, or legal rights; or
• use as the sole evidence or source of truth for care.

03How the AI workflow works

1. An authorized user selects the patient and attests to the complete, current patient AI notice, decision source, method, and consent status.
2. Audio is recorded or uploaded only after consent is marked granted.
3. Speech-processing AI produces an unverified transcript.
4. Clinical drafting AI receives the transcript, selected template, and limited patient context to produce a structured draft.
5. The clinician edits the note and accepts or dismisses suggested tasks and codes individually.
6. The note is locked only when the clinician deliberately finalizes it; connected source audio is then removed from active access and secure deletion is requested.
7. If the note remains unfinalized, source audio is removed from active access and secure deletion is requested 30 days after successful upload, unless a scoped legal hold or applicable law requires preservation.

04Known limitations

AI output can omit, invent, merge, or misattribute information. Review:

• negation and uncertainty;
• speaker identity and statements from another participant;
• medication, allergy, dose, unit, route, and frequency;
• dates, laterality, anatomy, dental notation, and follow-up timing;
• diagnostic and billing codes for the applicable jurisdiction;
• language, accent, audio quality, and accessibility-related errors; and
• whether a cited publication actually supports each statement.

A confidence score is not proof of clinical correctness. A fluent answer can still be wrong.

05Human review and product safeguards

• Generated and clinician-authored revisions remain distinguishable.
• Drafts remain editable before finalization.
• Tasks and codes require an individual accept or dismiss decision.
• Finalization requires a deliberate confirmation.
• The finalization confirmation explains the connected source-audio deletion before the clinician proceeds.
• Structured model responses are validated before being saved.
• Evidence citations must match PubMed identifiers returned by the search.
• Clinical content is rendered as text, not trusted executable HTML.

06Consent, refusal, and withdrawal

Do not record until the legally required consent is obtained and the patient or authorized substitute decision-maker understands the purpose of the AI-assisted workflow, the outside providers and possible processing outside Canada, the material privacy and reliability risks, the information retained, the right to refuse or withdraw, and the non-AI alternative. Obtain any required consent from every other audible participant. A clinician checkbox records an attestation; it is not patient consent by itself. Use the Patient AI Recording Notice as the minimum Porvio-specific layer within the clinic’s approved process.

Porvio stores the notice version, decision source, method, recording user, time, server-owned attestation, and append-only event history while the encounter is retained. If consent is withheld or withdrawn, stop using the recorder and use another documentation method without affecting care. Porvio detaches attached source audio, requests its deletion, and blocks later active AI results, but cleanup retries, backup lifecycles, and an external request already in flight may not be immediately retractable. Existing transcripts, drafts, finalized records, consent events, and audit history may remain under clinic record and retention duties. See the Privacy Notice.

07Privacy and model routing

Porvio’s current GPT-5.6 Luna text route is pinned to OpenRouter’s non-EU Azure provider, with fallback disabled, provider data collection denied, zero-data-retention routing and required parameters enforced, and a verified price ceiling. Release checks fail closed if those controls or the catalog signal drift.

MAI Transcribe 1.5 is release-verified against OpenRouter’s current catalog as Azure-only and zero-data-retention capable. Its current transcription request cannot carry the same per-request routing and privacy controls as the text route. Account settings, contracts, production verification, and continuing monitoring are therefore required. These technical controls are not an absolute provider non-retention or non-training promise, a compliance assurance, or a Canadian-residency guarantee. OpenRouter currently reports the named Azure routes as processing in the United States; see the Privacy Notice for cross-border handling.

08Performance and time-saving estimates

Any interactive time-saving calculator uses only the values selected by the visitor. It is an illustrative scenario, not a guarantee of Porvio’s performance, additional capacity, revenue, clinical outcome, or reduced workload. Results vary with specialty, workflow, encounter complexity, audio quality, review time, and service availability.

09Monitoring and accountability

Automated release checks confirm routing and application contracts; they do not prove clinical accuracy, fairness, safety, or regulatory status. Before a model or material model change is authorized for clinical use, named clinical, privacy, and security owners must document the intended use, evaluation evidence, acceptance thresholds, known subgroup and workflow limits, and rollback or pause criteria. Until that review is approved, the affected clinical use must remain unauthorized.

After authorization, the responsible owners must review material vendor changes, incidents, user feedback, and observed performance against those thresholds. A model or feature must be paused or restricted when its approval is missing or stale, a material change has not been assessed, or evidence falls outside the approved range. This page states the required governance standard; it is not evidence that a particular deployment has completed every review.

This approach is informed by the Ontario IPC’s guidance for AI scribes, but Porvio does not claim that reference alone establishes compliance.

10Report an unsafe or inaccurate output

Stop relying on an output if it may cause harm. Preserve only the minimum information needed to identify the problem and email support@porvio.app. Do not send patient names, recordings, transcripts, credentials, or full notes by ordinary email. Privacy concerns can be sent to support@porvio.app.