FrontRx AI safe use and clinical verification

System boundary

In scope: audio transcription and draft-note generation; AI-assisted medication extraction, prescription drafts, interaction and laboratory suggestions; and OCR extraction of identity and insurance fields from RAMQ or OHIP documents.

Supporting but not independently AI: the physician web and mobile interfaces, authentication and storage APIs, note templates, RAMQ billing catalogues, ICD-10 reference data, establishment data, PDF creation, and fax transmission. The clinic electronic medical record remains the official record.

Intended users, uses, and required context

1. Use FrontRx only as an authorized healthcare professional or authorized clinic team member acting within your assigned role and clinic workspace.
2. Use the AI Scribe to transcribe a healthcare encounter or professional dictation and to prepare a draft note in supported English or French workflows.
3. Use Klio to prepare draft prescriptions, medication lists, pharmacy communications, and laboratory requisitions for review by the responsible licensed prescriber.
4. Use RAMQ/OHIP OCR only to assist with data entry from an authorized source document. Compare every extracted field with the source before saving it.
5. Confirm patient identity, the correct encounter and clinic context, and a lawful need to process the information before starting.
6. Obtain and record valid patient consent before recording or processing an encounter. In Quebec, do not process a patient under 14 without consent from the holder of parental authority or legal guardian. Resolve any uncertainty about capacity, authority, or withdrawal before continuing.
7. Use Quebec RAMQ billing features only in an authorized Quebec billing workflow. Use other jurisdiction-specific functions only where the clinic has confirmed they apply.

Known limitations and what to monitor

1. Speech recognition can mishear names, medications, allergies, doses, numbers, dates, negations, abbreviations, and speakers. Risk increases with noise, overlapping speech, poor microphones, accents, code-switching, or speech differences.
2. Generated notes can omit facts, add unsupported statements, place information in the wrong section, or give a misleading summary when the transcript or attached context is incomplete or conflicting.
3. Attached documents, OCR results, prior notes, and patient-list data may be incomplete, outdated, assigned to the wrong patient, or difficult to read. The current encounter and verified source documents take priority.
4. Medication interaction and laboratory suggestions are support tools and may be incomplete. They do not replace the prescriber, pharmacist, current product monograph, local formulary, or clinical guidelines.
5. A model or provider fallback can change wording, level of detail, or response behavior. Recheck the entire output after any regeneration or fallback.
6. FrontRx does not claim equal performance for every age group, specialty, accent, disability, language variant, or care setting. Apply additional scrutiny when the patient, audio, source document, or clinical context differs from the conditions described here.
7. If the service is unavailable, the input is inadequate, or an output cannot be verified from authoritative sources, stop the AI workflow and use the clinic-approved manual process.

Similar but not recommended and prohibited uses

1. Do not use FrontRx for autonomous diagnosis, triage, treatment selection, prescribing, medication changes, billing submission, or final charting.
2. Do not rely on it for emergencies, time-critical decisions, or situations where independent verification cannot occur before action.
3. Do not use it as direct medical advice for patients or as a patient self-service diagnostic or prescribing tool.
4. Do not use it outside an authorized professional role, clinic workspace, jurisdiction, or supported English/French workflow.
5. Do not record or process after consent is refused or withdrawn, when the required representative has not consented, or when authority to process the information is uncertain.
6. Do not infer missing identity, demographic, allergy, diagnosis, medication, dosage, result, or eligibility information. Leave uncertain fields blank and verify them manually.
7. Do not treat a RAMQ check-digit result as proof of identity, current coverage, eligibility, or entitlement to a service.
8. Do not use production patient information for development, testing, demonstrations, model training, or other non-production purposes.

Environmental footprint and usage estimate

FrontRx does not train or fine-tune its own foundation model. Its attributable AI development footprint is therefore limited mainly to API inference used for development, validation, and production. The upstream providers do not currently expose enough model, hardware, data-centre, and grid detail to allocate their training, embodied-hardware, cooling-water, and carbon footprints precisely to one FrontRx request.

Planning scenario for one typical scribe note: a 15-minute encounter at about 150 spoken words per minute produces about 2,250 transcript words, estimated as roughly 3,000 input tokens. Adding instructions and relevant context gives an estimated 4,000 to 6,000 input tokens; a 750 to 1,500-token draft note gives about 4,750 to 7,500 total language-model tokens. This excludes audio-transcription compute, OCR, retries, regenerations, fallback calls, and optional assistant actions, which must be counted separately.

At 100 notes per clinician per month, this scenario represents about 475,000 to 750,000 language-model tokens, plus about 25 hours of audio transcription. At 1,200 notes per year, it represents about 5.7 to 9.0 million language-model tokens, plus about 300 hours of audio. Actual use must be calculated from recorded request counts, input/output tokens, audio minutes, OCR pages, retries, and provider fallbacks.

Energy and carbon values are sensitivity estimates, not measurements of FrontRx providers. A 2025 infrastructure-aware study estimated 0.43 Wh for a short GPT-4o query and more than 33 Wh for some long reasoning prompts. Applying those published endpoints to one note-generation call gives a deliberately broad range of about 0.043 to 3.3 kWh for 100 calls per month, or 0.52 to 39.6 kWh for 1,200 calls per year. Carbon dioxide equivalent is electricity in kWh multiplied by the time- and location-specific grid emission factor. FrontRx does not publish a single CO2e value until provider region and measured energy data are available.

Method limits and sources: energy varies materially with model, input/output length, batching, hardware, software, region, and cooling. The International Energy Agency states that close monitoring and more systematic provider disclosure are required as energy per task and AI use change quickly. Sources: https://arxiv.org/abs/2505.09598 and https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary.

1. Track monthly AI calls, input tokens, output tokens, audio minutes, OCR pages, retries, and fallback calls by workflow without placing patient information in telemetry.
2. Report actual median and 90th-percentile usage per completed note, not only an average, because long encounters and regenerations can dominate consumption.
3. Minimize unnecessary context, duplicate OCR, regeneration, and provider fallback; process only when needed and retain information only for the approved period.
4. Review this estimate at least annually and whenever the model, provider, routing, typical note length, or measured use changes materially.

Mandatory clinician verification before integration

1. Match the output to the correct patient, encounter, source document, and intended recipient.
2. Compare the transcript with the encounter when any wording is uncertain. Verify names, medications, allergies, doses, routes, frequencies, numbers, units, dates, negations, abnormal results, diagnoses, and follow-up.
3. Confirm that no unsupported fact, inference, instruction, or omitted critical information changes the clinical meaning.
4. For prescriptions and laboratory requisitions, independently confirm the indication, medication or test, dose, route, frequency, duration, quantity, renewals, interactions, allergies, contraindications, monitoring, and destination.
5. For RAMQ/OHIP OCR, compare the name, identifier, date of birth, sex or gender field, version code, and medical-record number with the source. A valid format or check digit does not confirm identity or eligibility.
6. Correct the draft, sign or approve it as required, and transfer it to the clinic official record or approved transmission workflow.
7. If a potentially unsafe output is found, do not use it. Preserve only the minimum non-identifying information required for investigation and report it to [email protected] and the clinic privacy or safety contact.