AI agents running an eye hospital behind the scenes. Ask the agent below anything you like.
Ranabir Bhattacharya, Chief Financial Officer, Disha Eye Hospitals, Kolkata.
Not a doctor. Not a programmer.
This is one of the agents. It knows what we built, what it cost, what broke, and how you would build the same thing. It will tell you when it does not know.
It answers only from what we have actually done. It will not give medical advice, and it cannot quote you a price.
This is not something you can buy. It is something you learn, and it is faster if someone has already made the expensive mistakes. We have started doing this outside Disha, so if any of it is useful to you, leave your number.
Your number is used to call you about this and nothing else. It is not added to any list and it is not shared.
The short version. Ask the agent above if you want any of it in more depth, it will go as deep as you like.
A chatbot answers a question. An agent acts on it. It opens the system, runs the report, reads it, decides whether anything matters, does something about it, and leaves a record. Then it does the same thing again tomorrow without being asked.
You give it the goal, not the steps.
A laptop, nothing special. A Claude subscription with Claude Code, which is the thing that writes and runs everything else. One person with a couple of hours a week and permission to try.
Later, a small cloud server at about fourteen dollars a month, so it keeps working while you sleep. You do not need a developer, a vendor, an integration project, or a capital approval.
No, and this is the part people get wrong. Every hospital IT project dies at that sentence.
An agent logs in to the same portal your staff use and reads the same screen. No export button, no API, no data warehouse. The only real prerequisite is that the number exists on a screen a person can open, rather than in a paper register or in somebody's head.
The most boring report you have. The one somebody assembles by hand every evening, where being late is annoying but nobody is harmed.
Write one paragraph saying what it should contain, who it goes to, and when. That paragraph is your entire specification. You type it in English and it builds the thing.
Your first useful agent is an evening's work.
Making it reliable enough that you stop checking takes a few weeks. That gap is the whole job, and it is where most attempts quietly die.
Logins expire, usually at three in the morning. Portals change without warning. Networks drop.
So in the same breath as the report, ask for retries, automatic re-login, and a watchdog that checks afterwards that the job really happened and tells you when it did not. Automation that needs a human to restart it is just a person with extra steps.
The server that runs all of it is about fourteen dollars a month. The AI subscription that writes and supervises the agents is separate and is not free, and that is the real cost, not the server.
Together, still less than one salary. But the biggest cost is neither: it is attention. Someone has to own it.
No. Everything at Disha was built by a finance person describing outcomes in plain English. Nobody was hired for it.
What you do need is somebody stubborn enough to keep going when the third attempt does not work.
Anything clinical. Anything that spends money. Anything that messages a patient without a signed consent trail.
Start where being wrong is cheap. Nobody is harmed by a late number. Every agent we run began life as a report.
At Disha it removed paperwork nobody wanted to do in the first place. No agent makes a clinical decision, spends money, or does anything with a real consequence without a person saying yes.
The aim is that people stop doing machine work, not that they leave.
Give it an owner whose actual job is to read what it produces and switch it off when it stops being useful. Put the output somewhere another person receives it, not somewhere only you look.
An agent with no owner becomes noise, then gets ignored, then becomes a liability.
Fifty-seven scheduled jobs across seventeen always-on services, clocking in about two thousand times a day.
They send 122 doctors their own numbers every night and answer their follow-up questions in about two minutes, watch how long patients are waiting branch by branch, message patients after their visit in their own language, read the punch-in machine, and check every two minutes that nothing has broken.