Healthcare Patient Triage AI
AI assistant handles patient inquiries, books appointments, and routes urgent cases to staff.
Client
Multi-specialty clinic chain
Industry
Healthcare
Region
South Asia
Duration
5 weeks build + 4 weeks clinical validation
A multi-specialty clinic chain was losing patients to competitor clinics because front-desk teams could not keep up - average response time was 6 hours. We deployed a clinical-grade triage AI on WhatsApp that gathers structured intake, applies clinical guidelines, books appointments, and routes urgent cases to on-call clinicians. Result: 24/7 coverage, response in under 2 minutes, 50% less admin load on clinical staff, and a 35% reduction in no-shows.
Headline results
The client operates 12 specialty clinics across three Indian cities - cardiology, orthopaedics, paediatrics, dermatology, and more. Inbound enquiry volume was 800-1,200 messages per day across phone, WhatsApp, and web. Front desks were processing roughly 30% of messages inside the hour, with the rest backlogging through the day and after-hours enquiries going unanswered until morning. Patients who could not get a fast response went to a competitor or, more dangerously, to an emergency room for cases that should have been a clinic visit.
What we built and shipped
Clinical-grade symptom-aware intake
An AI agent gathers structured intake (symptoms, duration, severity, history, medications, allergies) and applies clinical guidelines to triage cases into urgent, same-day, or non-urgent buckets.
- Specialty-aware question flows (cardiac, ortho, paeds, derm) tuned with clinical team input
- Red-flag detection that immediately routes to an on-call clinician
- Structured handoff data delivered to the doctor before the appointment
Multi-language conversational support
Built to handle English, Hindi, Tamil, and Marathi from day one, with the same triage logic across languages. Voice messages are transcribed and processed automatically.
- Native-quality conversation in 4 languages at launch
- Voice note transcription via Whisper + Indian-language fine-tune
- Language detection per message - patients can switch mid-conversation
Smart appointment booking
Books directly into the clinic's scheduling system with the right specialist, location, and slot type - accounting for travel time, doctor availability, and case complexity.
- Slot-type matching (consultation, follow-up, procedure) by case complexity
- Travel-time aware location suggestions across the 12-clinic network
- Automated reminder sequence at 24h, 4h, and 1h before appointment
Human-in-the-loop clinical safety
Anything flagged as urgent or unclear goes immediately to an on-call clinician. The agent never makes a final clinical decision - it routes intelligently and gathers structured information.
- All red-flag patterns trigger immediate clinician page
- Clinician dashboard with full conversation transcript and structured summary
- Periodic clinical audit of agent triage decisions for continuous improvement
How it actually works
Intake agent
Specialty-aware conversation agent gathering structured clinical intake. Per-specialty prompt sets and red-flag rules curated with the clinical team and audited quarterly.
Triage decision layer
Combines LLM reasoning with deterministic clinical rules - the LLM never overrides red-flag rules, so a chest-pain mention always pages the on-call clinician regardless of the wider conversation.
EMR integration
Read patient history if a returning patient identifies, write structured intake notes after the consultation, audit-grade event logging for every read/write.
PII and PHI safeguards
Field-level encryption for PHI, redaction at the model layer, no PHI used to train shared models, role-based access controls on the clinician dashboard.
Phased delivery timeline
Clinical co-design
Worked with the lead clinician of each specialty to draft question flows, red-flag patterns, and slot-type mapping. Built a 60-case evaluation set with sign-off from the clinical board.
Build & clinical shadow
Shipped intake + triage agent into shadow mode - every conversation was reviewed by a clinician before the patient saw a reply. Used the data to fix prompts and edge cases.
Clinical validation & go-live
4-week clinical validation phase comparing AI triage decisions vs. clinician-only decisions on the same cases. 95.6% concordance - clinical board approved go-live.
Monthly clinical audits + expansion
Monthly clinical audit of a random 50-case sample, with prompt updates per the audit findings. Expanded to handle post-appointment follow-up and medication adherence in months 4-6.
Before vs after
Same business, same team - measurably different operating model after the engagement.
- Average response time to patient6 hours
- Coverage hoursBusiness hours only
- Languages supported at quality1 (English)
- Clinical staff admin time30-40% of shift
- Appointment no-show rate22-26%
- ER redirects from clinic enquiriesCommon
- Average response time to patientUnder 2 minutes
- Coverage hours24 / 7
- Languages supported at quality4 native
- Clinical staff admin time12-18% of shift
- Appointment no-show rate14-16%
- ER redirects from clinic enquiriesEffectively zero from AI-handled cases
What changed and by how much
Operational and revenue metrics tracked from go-live, measured against the pre-engagement baseline.
Composition of impact
Approximate breakdown of how this engagement contributed to the business outcome - the headline metric is a roll-up of these levers.
- Clinical time recovered for care34%
- Patient acquisition (no leakage)24%
- No-show & rescheduling reduction18%
- After-hours coverage value14%
- Multi-language access lift10%
What we built it with
Model layer
- GPT-4 with clinical-tuned prompts
- Whisper + Indic ASR for voice notes
- Custom rules engine for red flags
Integrations
- WhatsApp Business API
- Clinic EMR (HL7 / FHIR)
- Cliniko-style scheduling system
Safety & compliance
- PII / PHI redaction layer
- Field-level encryption
- Full audit log of every agent action
Monitoring
- Clinician audit dashboard
- Daily triage-accuracy reports
- Sentry + on-call paging
What we de-risked along the way
Clinical mis-triage causing harm
Mitigation: Deterministic red-flag rules that override LLM output, immediate clinician escalation on uncertainty, monthly clinical audit, and explicit positioning to patients that they are talking to an assistant - not a doctor.
PHI leakage or non-compliant data flow
Mitigation: Field-level encryption at rest and in transit, enterprise model tier with no training on patient data, full audit logs for read/write, DPDPA-aligned data residency.
Patient resistance to AI conversation
Mitigation: Transparent disclosure at conversation start, easy human-handover keyword, satisfaction tracked per conversation. Adoption climbed from 40% to 85% over the first 8 weeks.
What we'd carry into the next build
Clinicians co-design or it fails
The agent was only as good as the clinical question flows behind it. Three weeks of co-design with specialty leads got us to a 95.6% concordance that internal-team-only prompts never would have.
Red-flag rules are non-negotiable
We learned this in shadow mode - the LLM sometimes reasoned its way past a chest-pain mention into a 'sounds non-urgent' triage. Hard rules over red-flag patterns saved us. The LLM gets to reason; it does not get to override safety rules.
Voice notes are a game changer in India
60% of inbound messages on WhatsApp in our patient base were voice notes. Without first-class voice handling the agent would have served the wrong half of patients.
Clinical validation is the project, not a checkbox
We spent 4 weeks in clinical validation before go-live - running every AI triage against a clinician-only decision on the same case. That timeline was non-negotiable and earned the clinical board's trust.
ROI & payback
Investment
Mid-six-figures one-time build + mid-four-figures monthly run cost (model + WhatsApp + infra)
Payback period
Inside 6 months on patient-retention and clinical-time savings alone
Year-1 ROI
Estimated 4-6× ROI; the larger compounding value is in patient experience and brand reputation
“Our front desks have time to be hospitable again. Patients get answers immediately, the right cases get the right slots, and our clinicians spend their time on care, not coordination. The AI assistant has become part of the team - and the clinical board signed off on it because we did the validation work properly.”
Questions about this engagement
Is this AI making clinical decisions?+
No. The AI performs structured intake and triage routing - it decides what kind of appointment to book and when to escalate to a human clinician, but it never makes a final clinical decision. Anything flagged as urgent or unclear goes immediately to an on-call clinician with the full conversation context.
How was clinical safety validated?+
A 4-week clinical validation phase before go-live, where every AI triage decision was independently graded against a clinician-only decision on the same case. Concordance was 95.6%, with the remaining 4.4% predominantly cases where the AI was more conservative (escalated when it could have booked routine). Monthly random-sample audits continue in production.
What about HIPAA, DPDPA, and clinical compliance?+
The system uses enterprise model tiers with no training on patient data, field-level encryption for PHI, full audit logs of every read and write to the EMR, role-based access on the clinician dashboard, and DPDPA-aligned data residency in India. HIPAA-ready architecture for clients with US patient flows.
Which languages does the agent handle?+
English, Hindi, Tamil, and Marathi at native conversational quality, with voice note transcription via an Indic ASR fine-tune. Patients can switch mid-conversation and the agent maintains context.
How did the no-show reduction happen?+
Three things combined: (1) reminder sequences at 24h, 4h, and 1h before appointment via WhatsApp, (2) re-confirmation if the patient does not acknowledge the 4h reminder, and (3) easy WhatsApp-based rescheduling instead of forcing patients to call. No-shows dropped from 22-26% to 14-16% within the first quarter.
Let's buildReady to put AI to work in your business?
Book a free 30-minute strategy call. We'll map your highest-impact automation opportunities and give you a clear roadmap - no obligation.