Estimated reading time: 8 minutes
Part of Selling AI to Indian Hospitals, a field guide.
Every patient-facing chatbot meets a question it cannot answer. The interesting part is what it does next. In a hospital the bad failure is not a bot that says “I don’t know”. It is a bot that answers with confidence and gets it wrong. A patient acts on a wrong answer about a report, a dose or a symptom, and the damage is done before any human sees it.
This piece is about that single moment, the fallback. It is also the question that decides a lot of hospital sales, so it helps to know what a good answer looks like before a buyer asks you.
The design, in one line each:
- When the bot is unsure, it says so. It never guesses.
- Every answer points to where it came from, so a human can check it.
- Anything clinical, urgent or personal leaves the bot and reaches a person.
- The handoff carries the conversation, so the patient does not repeat themselves.
- The questions the bot could not answer become the content you write next.
Why the failure case decides the sale
Hospitals do not buy a chatbot because it answers the easy questions. They buy it to take pressure off the front desk, which means it has to cope with the messy ones: a patient describing symptoms, someone asking about a bill, a family member worried about a report they cannot read.
A good demo answers ten clean questions. A good pilot shows what happens on the eleventh, the one the bot cannot handle. Buyers who have been burned before will test exactly that. It is the seventh entry in our list of objections you hear selling AI to Indian hospitals, and it is the one that separates a serious vendor from a demo.
The four outcomes a bot can produce
Any reply falls into one of four buckets. Good design makes the bucket explicit, and picks the safe one when the bot is unsure.
- Answer, with a source. The bot knows the answer and can point at the page or record it came from. Fine for opening hours, departments, directions, appointment policy.
- Answer, with a caveat. The bot is fairly sure but the topic is sensitive, so it answers briefly and offers a human. Fine for general information about a procedure, not for a decision about one patient.
- Hand to a human. The bot does not know, or the patient asks for a person. The conversation moves to staff with the context attached.
- Route to urgent care. The message carries a red flag. The bot stops answering and points to emergency care or the clinic, straight away.
The mistake most tools make is collapsing the middle two. A bot that can only answer or apologise leaves the front desk with the hard cases and no warning that they are coming.
What “I don’t know” should look like
A refusal is a feature, as long as it is useful. Three things make it useful:
- It is short and plain. “I don’t have a reliable answer for that. Let me connect you to the hospital team.” Not a paragraph of apology.
- It offers the next step in the same message, so the patient never ends up at a dead end.
- It records the question, so the hospital learns what patients actually ask and can add the answer later.
That last point is worth more than it sounds. The unanswered questions are the cheapest content research a hospital will ever get, and they are specific to that hospital’s own patients. A vendor who brings that list to a monthly review is showing the buyer something no competitor’s demo can.
Human handoff that actually works
A handoff is not a link to the contact page. It is a transition, and the patient should feel it happen. Five things make the difference:
- A named destination. The conversation goes to a person or a queue, not into a form nobody monitors.
- Context that travels. The staff member sees what the patient already asked, so the patient does not start again.
- Hours that match reality. If the desk is closed, the bot says when it reopens and offers a callback or a next step now.
- A wait the patient can see. Tell them roughly how long, and tell them if it changes.
- A human reply, not a survey. The first message after a handoff should answer the question.
After hours is where most deployments quietly fail. The bot answers well at 3pm and abandons people at 11pm, which is exactly when a worried family searches. Decide what happens then before the pilot starts, not after the first complaint.
Red flags do not go through the bot
Some messages should stop the conversation. Severe pain, heavy bleeding, breathlessness, chest pain, a patient who says they are fainting: these go to a human or to emergency care immediately, and the bot does not try to be helpful first.
There is a harder version of this. A patient who says they feel hopeless, or mentions hurting themselves, is not asking a product question. The bot should not counsel, reassure or change the subject. It should say plainly that a person will help, pass the message to a human with the same urgency as a clinical red flag, and let the hospital’s own crisis protocol answer. Test that path before launch, because the first time you find out it does not work is the worst possible time.
The rule that matters is simple. Never reassure. A bot that says “that is probably nothing to worry about” is doing clinical triage, badly, without a licence. If your tool touches anything clinical, get a clinician to sign off the exact wording of those replies, and keep that wording under version control so a copy edit cannot quietly weaken it. This is the same point we make in why healthcare AI has a trust problem.
Answers you can trace
If a hospital asks why the bot said something, the answer has to be a page, not a shrug. That means:
- Each answer is grounded in hospital content, not in the model’s general knowledge.
- The content is dated, so a stale price or a moved department can be spotted and fixed.
- Changes to the content are logged, so a wrong answer can be traced to the version that caused it.
- A patient can ask for a person at any point, not only when the bot fails.
What to measure in the pilot
Fallback quality is measurable, which is good news, because it means the pilot can judge it rather than argue about it. Watch five numbers:
- The share of conversations that end in a handoff. Rising is not automatically bad. It can mean the bot is catching more.
- How long a patient waits for a human after the handoff.
- The share of handoffs a human actually answers.
- The top unanswered questions, in the patient’s words.
- Reported wrong answers, and how long each took to correct.
Put those five in the pilot agreement next to the commercial metric. A pilot that only measures leads captured will happily hide a bot that guesses. The general shape of a pilot that converts is in the Indian hospital pilot that actually converts.
Three questions to ask a vendor
Buyers can test all of this in one conversation, before any contract:
- Show me a question your bot cannot answer. What does the patient see, and what do we see?
- When a patient asks for a person at 11pm, what happens, exactly?
- Can you show me where a specific answer came from, and who last changed it?
A vendor who answers those three without hesitating has built this before. One who talks about accuracy percentages instead has not understood the question. The same test is on the one-page hospital AI vendor readiness checklist, which also covers the data and compliance questions that come up in the same meeting.
Where this fits in a deployment
Fallback design is one seam in a bigger integration. The bot sits on top of the hospital’s appointment system and its own content, and it has to respect both. If you are planning the whole picture, the HIS, HMIS and WhatsApp integration guide covers where the bot plugs in, the WhatsApp guide covers the channel where most of these conversations actually happen, and the DPDP data questions cover what you may store once a conversation turns into a patient record.
One honest limit. This is not a specification for clinical triage, and no chatbot should be one. The moment a message could describe a medical emergency, the right answer is a person, and the design work is about getting there quickly and reliably. Everything else is a useful assistant.
Related reading
- The objections you hear selling AI to an Indian hospital
- Why healthcare AI has a trust problem
- The hospital AI vendor readiness checklist
- The full field guide: selling AI to Indian hospitals
Next in the field guide: the one-page readiness checklist, which pulls the whole series into a single page. Or see all the guides.
Get the monthly letter
Building BeyondChats in public: the numbers, the decisions, and what broke. One email a month. No spam, unsubscribe anytime.