Generative AI in healthcare: where it works in India today
Generative AI in healthcare is already useful in India, but mostly away from the diagnosis. The work that pays back today sits in the patient conversation (WhatsApp, chat and voice), in content and search visibility, in documentation support and in internal knowledge. Start where errors are cheap and reviewable, keep clinicians in charge of anything clinical, design for India’s languages and the DPDP Act from day one, and measure outcomes the CFO already tracks.
What generative AI in healthcare actually means
Most hospital AI until recently was predictive: a model scores a scan, flags a risky patient or forecasts bed demand. Generative AI is different. Large language models and their voice and image cousins produce new text, speech and summaries. That makes them good at conversation, drafting and search, and unreliable wherever a confident wrong answer could hurt someone.
Two newer terms matter. Conversational AI is generative AI wrapped in a chat or voice interface, the kind patients meet on WhatsApp or the phone. Agentic AI goes a step further: the system plans and takes actions, such as checking a doctor’s calendar and booking a slot. Each step up adds value and adds risk, so the governance has to step up with it.
I write about this from the digital, product and brand side of hospitals, not the clinical side. Everything below is about non-clinical use, or clinical-adjacent use where a clinician signs off.
Where generative AI works in Indian hospitals today
1. The patient front door: WhatsApp, chat and enquiries
Indian patients already message hospitals on WhatsApp, at all hours and in several languages. Generative AI can answer the logistical questions (timings, directions, documents, which doctor sees what), collect details and route the enquiry to the right desk. The measure that matters is resolution, not deflection; I set out the difference in chatbot containment vs resolution. For the plumbing, see the WhatsApp Business API for hospitals and WhatsApp as the hospital front door.
The hard rule: the system can route a medical question, it should not answer one. Using AI to triage enquiries without making clinical calls covers where that line sits in practice.
2. The contact centre and voice
Voice bots now handle multilingual calls well enough for appointment booking, report status and reminders. The economics depend on which calls you automate and how cleanly a caller reaches a human. Voice bots and contact-centre cost works through the numbers a CFO will ask for, and when automation should not answer the phone covers the calls to leave alone.
3. Content at scale, with medical accuracy
Generative AI drafts patient education pages, FAQs, video scripts and regional-language versions in minutes. The risk is fluent text that is subtly wrong. The workable model is AI drafting, a named doctor reviewing and a clear record of who approved what. I describe it in AI content at scale without wrecking medical accuracy and where the review line sits for doctor content. A shared prompt library keeps quality consistent across a team.
4. Being found by AI search
Generative AI also changes how patients find you. ChatGPT, Gemini, Perplexity and Google’s AI Overviews now answer many health and hospital questions directly. Hospitals that publish clear, reviewed, well-structured answers get cited; others disappear from the answer. Start with how hospitals get cited by AI search, SEO vs GEO for hospitals and structured data for hospitals.
5. Documentation support
Ambient documentation tools listen to a consultation and draft the note for the doctor to edit and sign. Done well, they give doctors time back and improve records. They also touch patient data and clinical records, so consent, accuracy checks and clinical ownership come first. Ambient documentation in the consultation room covers the questions to settle before a pilot.
6. Internal knowledge and operations
Some of the quietest wins are internal: searching policies and SOPs, summarising long vendor contracts, drafting board notes, answering staff questions about benefits or processes. These carry low patient risk and build the organisation’s confidence before anything faces a patient.
Where it should not go yet
- Diagnosis or treatment advice to patients from a general-purpose model, without a clinician in the loop.
- Unsupervised medical answers on chat, WhatsApp or voice. Route, do not advise.
- Anything that hides the human. Patients should know when they are talking to software and how to reach a person.
- Personal health data in tools you do not control. Check where data goes, who can see it and how long it is kept.
Putting AI in front of patients without an incident and a non-clinical risk framework for AI in hospital operations go deeper on guardrails.
The India context
Three things make Indian deployments different. First, language: patients switch between English, Hindi and regional languages, often in one message, so test in the languages your patients actually use. Second, channel: WhatsApp and the phone carry far more of the patient conversation than web forms or apps. Third, regulation: the Digital Personal Data Protection Act, 2023 and the DPDP Rules, 2025 shape consent, notice and retention for any system that handles patient data. Professional conduct rules for doctors and advertising guidelines still apply to anything the AI says in your name. None of this is legal advice; involve your legal and data protection teams early.
A 90-day way to start
- Weeks 1 to 2: pick one problem with a number attached, such as unanswered WhatsApp enquiries after 8 pm or report-status calls. Write down today’s baseline.
- Weeks 3 to 6: build a narrow pilot with a clear handover to humans, reviewed scripts and a log of every conversation. Get legal and clinical sign-off on scope.
- Weeks 7 to 10: run it on real traffic for a defined segment. Review a sample of conversations every week.
- Weeks 11 to 13: compare against the baseline, decide to scale, fix or stop, and write down why.
Most pilots fail after this point, not during it. Why hospital AI pilots never reach production and the data work nobody budgets for explain why, and an AI strategy a hospital group can execute shows how to sequence the next projects.
How to measure it
Measure what the business already tracks: enquiries answered within a set time, appointments booked from those enquiries, calls handled without a callback, hours of doctor time returned, content published with review completed. Avoid vanity measures such as messages sent or tokens used. Measuring an AI project when nothing clinical is being claimed gives a template.
Where generative AI sits in the patient journey
If you map the journey from first search to follow-up, generative AI fits naturally at the search, enquiry, booking, preparation and follow-up stages, and much less at the clinical core. Where AI belongs in the patient journey maps it stage by stage, and healthcare marketing automation in India shows where rules-based automation is still the better tool.
Questions people ask
Generative AI uses large language models and similar systems to produce text, speech and summaries. In healthcare it is used for patient conversations, content, documentation support and internal knowledge. It differs from predictive AI, which scores or classifies data such as scans.
Mostly outside the diagnosis: answering and routing WhatsApp and phone enquiries, voice bots for bookings and report status, drafting patient education content for doctor review, improving visibility in AI search, documentation support and internal knowledge search.
Not without a clinician in the loop. A general-purpose model can give fluent but wrong answers. The safer pattern is to let the system answer logistical questions and route medical ones to a doctor or nurse, with emergency signposting.
Conversational AI talks with people through chat or voice. Agentic AI also takes actions on its own, such as checking availability and booking an appointment. Agentic systems need stronger controls because their mistakes become actions, not just words.
The Digital Personal Data Protection Act, 2023 and the DPDP Rules, 2025 govern consent, notice, purpose and retention for personal data. Any AI system handling patient data needs a clear legal basis, a privacy notice and controls on where data goes. Check specifics with your legal team; this is not legal advice.
One with a clear baseline and low clinical risk, such as after-hours WhatsApp enquiries or report-status calls. It shows value quickly and builds confidence before anything more sensitive.
About 90 days is realistic: two weeks to define the problem and baseline, four weeks to build and get sign-off, four weeks of live use, and a few weeks to review and decide.
Use measures the business already tracks: enquiries answered on time, appointments booked, calls resolved without callbacks, doctor hours returned and content published with review complete. Compare against a baseline taken before the pilot.
It can draft it. A named doctor should review and approve every page, and the page should show who reviewed it and when. Content must also follow professional conduct rules and advertising guidelines.
Increasingly well for Hindi and major regional languages, and less evenly for mixed-language messages and smaller languages. Test with real patient messages in the languages your patients use before going live.
Usually not because the model is weak. Common causes are no clear owner, messy or inaccessible data, no budget for integration and change management, and no agreed measure of success.
Rarely. Most hospitals do better configuring proven platforms with their own content, guardrails and integrations, and spending their effort on data, workflow and review rather than model training.
Patients increasingly get answers from AI assistants and AI Overviews. Hospitals that publish clear, reviewed, well-structured answers with good structured data are more likely to be cited in those answers.
Read my takes first in Google Search
