Agentic AI in Healthcare: Use Cases for Indian Hospitals, and Where Not to Use It
Published · By Gaurav Phogat
Agentic AI in healthcare means software that does not just answer a question but carries out a task across several steps: it reads an enquiry, checks a calendar, books a slot, sends the confirmation and logs the record. In Indian hospitals the safe and useful ground for agents today is non-clinical work: enquiries and booking, reminders, insurance paperwork, international patient coordination, content operations, contact-centre QA and marketing operations. Each needs a named human checkpoint. Agents should stay out of triage, clinical advice, consent capture and anything NMC or DPDP sensitive.
I have spent the last few years putting AI assistants, CRM automation and contact-centre tooling in front of real patients and real hospital staff. The pattern I see now is that the word “agent” is being attached to everything, including tools that are chatbots with a new label. This guide is my working view of what an agent actually is, where it earns its place in an Indian hospital, where it must not go, and how to run a pilot that a CFO and a medical director can both sign. It is written from the digital, brand and patient-experience side, not as clinical guidance. For the wider picture of AI across Indian healthcare, start with the AI in healthcare in India pillar.
This article is part of the AI in healthcare operations series, best read in order from the pillar.
What agentic AI is, and how it differs from chatbots and copilots
The simplest test I use: who presses the button? A chatbot answers. A copilot drafts and a person acts. An agent acts, inside limits someone has set, and a person reviews either before or after. That one difference changes the risk, the governance and the business case.
| Chatbot | Copilot | Agent | |
|---|---|---|---|
| What it does | Answers questions from a script or a knowledge base | Drafts text, summaries or suggestions for a staff member | Plans and completes a multi-step task using tools and systems |
| Who acts | Nobody; it only replies | The staff member, after reading the draft | The software, within permissions and rules |
| Hospital example | Website bot giving OPD timings | Agent-assist that drafts a reply for a contact-centre executive | Reads a WhatsApp enquiry, checks doctor availability, books, confirms, updates the CRM |
| Systems touched | Knowledge base | Inbox, CRM screen | Calendar, HIS or booking engine, CRM, messaging, payments |
| Main failure | Wrong or vague answer | Staff approve a bad draft without reading | Wrong action taken at scale before anyone notices |
| Governance need | Content review | Training and spot checks | Permissions, audit trail, escalation rules, kill switch |
Two consequences follow. First, an agent is only as good as the systems it can reach; if your booking engine has no API, you do not have an agent, you have a chatbot that apologises. Second, the risk moves from “a wrong sentence” to “a wrong action”, which is why the human checkpoint matters more than the model choice. I covered the difference between a bot that closes a conversation and one that solves the problem in chatbot containment vs resolution; agents raise the stakes on that distinction.
Why agentic AI in healthcare is on the agenda in India now
Three things changed. Large health IT vendors now ship agents inside their products, so the question reaches hospitals through vendor roadmaps whether or not anyone asked for it. Analysts are openly warning about hype: Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls, and estimated that only about 130 of the thousands of vendors claiming agentic AI are genuine (it calls the rest “agent washing”). And India now has a policy frame: the IndiaAI Mission, the AI Centres of Excellence and, most importantly for patient-facing work, the DPDP Rules.
My read for a hospital group: the opportunity is real, but it sits in the unglamorous back office and the front desk, not in the diagnostic demos on conference stages.
Non-clinical use cases for Indian hospitals
These are the seven areas where I would put an agent first. The readiness column is my judgement of how far most Indian hospital groups are from a safe production version, based on the state of their systems and data, not on what vendors demo.
| Use case | What the agent does | Human checkpoint | Main risk | Readiness |
|---|---|---|---|---|
| Enquiry handling and appointment booking | Reads WhatsApp, web and call-transcript enquiries, finds the right department and doctor slot, books, confirms, writes to CRM | Any enquiry with symptoms, urgency words or a request for advice goes to a trained executive | Wrong doctor or wrong unit booked; symptom questions answered by software | High where the booking engine has an API |
| Follow-up and reminders | Sends appointment reminders, reschedule links, document checklists and post-visit feedback requests (no review requests) | Templates approved once by marketing, compliance and medical; changes need re-approval | Messaging without valid consent; over-messaging | High |
| Pre-authorisation and claims paperwork | Collects documents, checks them against the insurer or TPA checklist, flags gaps, drafts the cover note, tracks queries | TPA desk executive submits; the treating doctor alone writes the clinical justification | Incomplete file, wrong policy details, agent drafting medical necessity text | Medium |
| International patient coordination | Gathers reports, translates the logistics parts of messages, prepares the estimate request, schedules video calls, tracks visa letters | Coordinator reviews every outgoing message; doctors give every medical opinion | Translation of clinical content; promises on cost or outcome | Medium |
| Content operations | Turns approved doctor inputs into drafts, checks style, internal links, schema and metadata, flags NMC risk words | Medical reviewer signs off accuracy; marketing signs off claims | Inaccurate medical content published at scale | Medium |
| Contact-centre QA | Scores every call transcript for script adherence, missed bookings, consent wording and escalations, then queues coaching | QA lead samples the scores and owns any action on staff | Unfair scoring; privacy of recorded calls | High |
| Marketing operations | Builds campaign reports, reconciles leads to appointments, checks UTM and tracking, pauses broken ads, prepares weekly summaries | Marketing lead approves budget changes and any new creative | Spend moved on bad data; ad claims changed without review | Medium to high |
1. Enquiry handling and appointment booking
This is where most of the value sits, and most of the risk. The agent is not deciding what is wrong with the patient; it is routing a stated need (“knee replacement consultation, Saturday, near Banjara Hills”) to an available slot. The rule I hold to is that the moment a message carries symptoms, urgency or a request for advice, the agent stops and hands over to a person with the conversation summarised. I wrote the detailed version of that design in AI enquiry triage without clinical calls. Industry examples point the same way: an AI scheduling agent is a booking product, not a clinician.
2. Follow-up and reminders
Reminders are the easiest agent to justify because the work is rule-based and the failure is cheap. The design detail that matters is consent: each message needs a recorded purpose and an opt-out, and the agent must read the consent flag before it sends, not after. Feedback requests are fine; requests for public reviews are not, under the NMC 2026 rules covered in the NMC 2026 compliance hub.
3. Pre-authorisation and claims paperwork
Every TPA desk I have seen loses hours to missing documents and repeated queries. An agent that checks the file against the insurer checklist before submission and chases the gap is pure administrative help. The line it must not cross is writing the clinical justification or choosing the procedure code that drives the claim; those stay with the doctor and the coder.
4. International patient coordination
Medical travel enquiries arrive at odd hours, in several languages, with attachments. An agent can acknowledge within minutes, collect reports, schedule the video consultation and track the visa invitation letter. It should never give an opinion on a report, quote a final price or suggest an outcome. The coordinator owns every outgoing message during the pilot.
5. Content operations
Agents can run the production line around content: brief, draft from approved doctor inputs, check readability, internal links, schema, alt text and metadata, then queue for review. They cannot be the reviewer. I explain the review chain in AI medical content at scale and the marketing uses in AI in healthcare marketing.
6. Contact-centre QA
Most contact centres listen to a small sample of calls. An agent can read every transcript and flag missed bookings, wrong information, missing consent wording and calls that should have been escalated. Scores are a coaching input, never an automatic disciplinary trigger; HR and the QA lead own that step. The economics of voice automation are in voice bots in the hospital contact centre.
7. Marketing operations
Weekly reporting, lead to appointment reconciliation, broken tracking checks and pausing ads that point to dead pages are good agent work. Changing budgets, launching new creative or editing ad claims are not, because a wrong change costs money fast and can breach advertising rules.
Where not to use agents
This list matters more than the use-case list. In each case the reason is the same: the action is either a clinical decision, a legal act, or a public claim, and none of those can be delegated to software in a hospital.
- Triage decisions. Deciding urgency, which specialty a symptom belongs to, or whether someone should go to emergency is clinical. An agent can detect that a message needs a human and route it fast; it cannot rank the patient.
- Clinical advice of any kind. No dosing, no “this sounds like”, no preparation advice beyond what a doctor has approved as a fixed, standard instruction, and no interpretation of reports. If the patient asks, the agent says a doctor or nurse will respond and hands over.
- Consent capture. Consent for procedures, data sharing or research is a recorded act between the patient and the hospital. An agent can send the consent notice and record that a person clicked; it must not explain clinical consent, persuade, or mark consent on someone’s behalf.
- Anything NMC sensitive. Publishing ads, doctor posts, before and after material, or patient stories; editing claims; replying to public reviews. Agents can draft and flag; a named person publishes.
- Anything DPDP sensitive. Pulling patient lists for marketing, building audiences from health data, sharing records with vendors or other entities, or deciding retention. These need a lawful basis the hospital can show, and a human owner. See the DPDP Act guide for hospital marketers.
- Money and promises. Final estimates, discounts, refunds and anything that sounds like an outcome guarantee.
If you are putting any of this in front of patients, the incident patterns and the safeguards are covered in AI in front of patients without an incident.
Guardrails and governance
I grade every agent on an autonomy ladder before it goes live, and I make the level a line item in the approval note. Most hospital agents should live at levels 1 to 3 for their first year.
| Level | What the agent may do | Human role | Typical hospital use |
|---|---|---|---|
| 0. Suggest | Recommends an action on screen | Person decides and acts | Next-best-action prompt for a contact-centre executive |
| 1. Draft | Prepares a message, file or report | Person approves every item before it goes out | International patient replies, pre-authorisation cover notes |
| 2. Act with approval | Executes after a one-click approval | Person approves in batch or one by one | Rescheduling a cluster of appointments when a doctor is on leave |
| 3. Act and report | Executes within strict rules, logs everything | Person reviews samples and exceptions daily | Reminders, booking confirmed slots for stated needs, QA scoring |
| 4. Act autonomously | Executes and adapts with little review | Periodic audit only | Not recommended for patient-facing hospital work today |
Audit trail
Every action needs a record: what came in, what the agent decided, which system it touched, what it sent, and which rule allowed it. If a patient complains that they were booked with the wrong doctor, you should be able to reconstruct the decision in minutes. Ask vendors to show this log in the demo, not in a slide.
Escalation
Write the escalation list before the prompt. Mine always includes symptom words, urgency words, distress, complaints, legal or media mentions, minors, payment disputes, and any message the agent is less than sure about. Escalation must reach a staffed queue with a response-time target; an escalation to an unmonitored inbox is worse than no agent.
DPDP and data handling
The DPDP Rules, 2025 were notified on 14 November 2025 with an 18-month window for phased compliance; they require affected individuals to be told about a personal data breach without delay, and set up India-based consent managers. For agents this means: purpose-limited access to data, consent checks inside the workflow, a data processing agreement with every vendor, no training of vendor models on your patient data without explicit terms, and deletion rules that the agent itself follows. The ICMR ethical guidelines for AI in biomedical research and healthcare (2023) are also worth reading for governance and ethics review, even for non-clinical deployments.
A kill switch and an owner
Each agent needs one accountable owner by name, a documented way to switch it off within minutes, and a fallback process the team can run manually. Governance at group level is part of the AI strategy a hospital group can execute. Whether that ownership needs a dedicated executive is covered in chief AI officer in healthcare.
Real examples worth studying
I only use examples I can trace to a primary source. They are useful less for the products themselves than for how each one frames the human role.
- Epic’s named agents. In a March 2026 update, Epic describes Emmie (patient-facing: billing questions, scheduling and rescheduling, follow-up reminders), Art (clinician-facing) and Penny (revenue cycle: prior authorisation, coding-related denials and appeal letters), and says agent actions on its platform are traceable with local policies set by each organisation. The pattern to copy is the split by user and job, not the clinical features. I looked at the patient-facing side in what Epic Emmie means for the patient front door.
- Hippocratic AI. The company describes its agents as built for patient-facing non-diagnostic tasks, and its January 2025 announcement describes escalation to a nursing team when a remote-monitoring alert looks critical, with a healthcare professional validating each interaction. Whatever you think of the category, “non-diagnostic plus escalation” is the right shape.
- IndiaAI Mission. The Union Cabinet approved the IndiaAI Mission with an outlay of Rs 10,371.92 crore in March 2024, covering compute, datasets, skills and startups. For hospitals this mainly shows up as subsidised compute and Indian-language models over time.
- AI Centres of Excellence. In October 2024 the Ministry of Education announced three AI Centres of Excellence focused on healthcare, agriculture and sustainable cities, with Rs 990 crore for FY 2023-24 to FY 2027-28.
- Hospital AI pilots in general. The reasons most pilots stall (no owner, no integration budget, no production criteria) apply doubly to agents. See why hospital AI pilots never reach production.
How to pilot an agent in 90 days
Pick one use case from the table above, one unit, one channel. The goal of the pilot is not to prove the AI is clever; it is to prove the process is safe and the numbers move.
| Weeks | What happens | Gate to pass |
|---|---|---|
| 1 to 2: Scope | Choose the task, write the escalation list, set the autonomy level, name the owner, agree the baseline metrics | Medical, legal and IT sign the one-page scope |
| 3 to 4: Build | Connect the systems (booking, CRM, messaging), load approved content, write the audit log, set the kill switch | End-to-end test with dummy data; every action appears in the log |
| 5 to 8: Shadow | Agent runs at level 1: drafts every action, staff approve or correct each one | Correction rate falls steadily; zero clinical or consent breaches in drafts reaching patients |
| 9 to 12: Controlled live | Move the safest sub-tasks to level 3 for part of the traffic, keep the rest at level 1 | Agreed metrics beat baseline; escalations answered on time; owner recommends scale or stop |
Two things I always budget for: integration work, which is usually larger than the licence, and the data clean-up nobody plans for, covered in the data work nobody budgets for. Write the stop criteria on day one. A pilot that cannot fail is not a pilot.
Measuring value
Measure the task, not the AI. The numbers I put in front of a CFO are:
- Resolution rate: share of tasks completed correctly with no human rework, not the share of chats the bot “handled”.
- Time to first response and time to booking for enquiries, by hour of day.
- Enquiry to honoured appointment: did the booked patient actually arrive?
- Rework and correction rate from the shadow phase onwards.
- Escalation accuracy: how many messages that needed a human reached one, and how fast.
- Staff hours released and where they were redeployed, which is the real saving.
- Incidents: complaints, wrong bookings, consent misses, each logged and reviewed.
The longer method, including how to avoid vanity numbers, is in measuring non-clinical AI projects.
Frequently asked questions
Agentic AI in healthcare is software that completes a multi-step task, such as reading an enquiry, finding a slot, booking it and updating the CRM, within limits set by the hospital. Unlike a chatbot it takes actions in other systems, so it needs permissions, an audit trail and a human checkpoint.
A chatbot answers questions. A copilot drafts something and a staff member decides. An agent carries out the action itself inside set rules, and a person reviews before or after. The difference is who acts, which changes the risk and the governance.
Non-clinical examples include booking appointments for stated needs, sending reminders, checking pre-authorisation files against insurer checklists, coordinating international patient logistics, running content production checks, scoring contact-centre calls for QA and reconciling marketing leads to appointments.
Start with one non-clinical, high-volume task where the systems already have APIs, usually enquiry to appointment booking or reminders. Run a 90-day pilot with a named owner, a written escalation list and stop criteria, and decide on scale only after the controlled live phase.
Ask for resolution rate, time to booking, enquiry to honoured appointment, rework rate and staff hours released, all against a baseline. Expect integration costs to exceed the licence in year one, and treat any case that rests only on chats handled as incomplete.
It should not make triage decisions. It can detect that a message contains symptoms or urgency and route it quickly to a trained person with a summary. Deciding urgency or specialty is a clinical judgement that stays with clinicians.
Yes, if the hospital has a lawful basis for each use of personal data, checks consent inside the workflow, limits the agent to the data it needs, has processing agreements with vendors and can report breaches. The DPDP Rules were notified in November 2025 with an 18-month phased compliance window.
Ask which systems the agent writes to and how, how permissions are scoped, where data is stored and processed, whether patient data trains their models, how the audit log looks, how fast the agent can be switched off and what the fallback process is.
Yes, for scope and content. A medical lead should approve the escalation list, any standard instructions the agent sends and the boundary between administrative and clinical messages. Clinicians do not need to approve each booking.
Agents can draft and flag content, but publishing ads, doctor posts or replies to public reviews needs a named person. The NMC 2026 rules bar testimonials, review requests, superlatives and cure claims, so agents should be configured to flag these words, not to publish.
Clear role changes for staff whose tasks move to the agent, training on reviewing agent output, and a rule that QA scores from an agent are coaching inputs, not automatic disciplinary triggers. Redeployment of released hours should be planned before go-live.
An agency can build content pipelines, reporting agents and campaign checks, but the hospital should own the accounts, data, prompts and logs. Contracts should cover data processing, model training restrictions and handover of all assets.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 because of rising costs, unclear value or weak risk controls. In hospitals the usual causes are no named owner, missing integrations, poor data and no agreed production criteria.
Sources
- Gartner, press release, 25 June 2025: over 40% of agentic AI projects predicted to be cancelled by end of 2027
- Epic, “Real Results, Right Now”, 10 March 2026
- Hippocratic AI announcement, Business Wire, January 2025
- PM India, Cabinet approves IndiaAI Mission, 7 March 2024
- Ministry of Education (PIB), three AI Centres of Excellence, October 2024
- PIB, DPDP Rules, 2025 explainer, November 2025
- ICMR, Ethical Guidelines for Application of AI in Biomedical Research and Healthcare, 2023
Read my takes first in Google Search
