How do you handle DPDP Act 2023 requirements for patient data?
We treat patient records as sensitive personal data by default. Consent notices go out before the first interaction on WhatsApp or IVR. Data lives in your VPC (AWS Mumbai, Azure India, or on-prem) and never crosses borders unless you approve it. For LLM calls we prefer self-hosted Llama or Mistral for anything touching PHI, and use zero-data-retention agreements with OpenAI or Anthropic for less sensitive tasks. Right to erasure, purpose limitation, and grievance officer workflows are wired in.
Do the agents work with ABDM and ABHA?
Yes. We integrate with ABDM Sandbox and production APIs, register facilities as HFR entities, and link patient records to ABHA IDs where the patient has consented. Agents can fetch prior health records via the ABDM consent framework and push discharge summaries or lab reports back to the patient locker. Useful for clinic chains that want to differentiate on continuity of care.
Can agents integrate with our HIS, EMR, or LIS?
Yes. We integrate with Insta HMS, MediXcel, HealthPlix, Practo, Bahmni, Attune, Suvarna, and custom hospital systems via REST APIs, HL7 v2 messages, FHIR R4 resources, or direct database adapters where APIs are missing. Lab integrations include SRL, Metropolis, Thyrocare, and custom LIMS. If no API exists, we build a thin middleware layer.
On-premise or cloud, which should we pick?
Depends on your data policy and volume. Small clinics and healthtech startups run fine on AWS Mumbai or Azure India with self-hosted open-source models for anything sensitive. Larger hospitals with strict IT policy or NABH audit exposure often prefer on-prem with a GPU box (a single RTX 6000 Ada or L40S handles most workloads). We deploy the same agent stack in both modes so you can migrate later.
What about NMC advertising rules and clinical claims?
Our agents are strictly non-diagnostic. They coordinate, schedule, remind, and route. They do not offer clinical advice, do not suggest diagnoses, and do not recommend treatments. Any symptom collection during intake is passed as-is to a clinician for review. This keeps you inside NMC and state medical council advertising and practice rules.
What is the realistic ROI on an appointment agent?
For most clinic chains, we see 30-50 percent no-show reduction from reminder plus reschedule flows, 40 percent front-desk time saved on scheduling calls, and payback in 3-6 months for a mid-size clinic doing 200-500 appointments a day. Larger hospitals see the ROI faster because a single filled OT slot from better scheduling can cover a month of agent operating cost.
Which models do you use and why?
Router pattern. Cheap, fast open models (Llama 3.1, Mistral, Qwen) for classification, intent detection, and structured extraction. Larger models (GPT-4o class, Claude Sonnet) only for complex reasoning steps such as multi-turn triage routing or drafting a TPA pre-auth letter. Keeps unit economics workable at Indian patient volumes.
What does it cost end to end?
Pilot with one agent lands at 1.5-3 lakh. A full multi-agent build with 3-5 agents, HIS integration, and orchestration is 5-15 lakh depending on scope. Managed operations start at 40 thousand per month and cover model costs, prompt tuning, integration monitoring, and audit logs. Most of the ongoing cost is people time on your escalation queue, not model tokens.
