WhatsApp and AI in Cancer Care: Logistics Yes, Clinical Answers No

At 1:14 AM, a message arrives on the practice WhatsApp. "Papa ko chemo ke baad bukhaar hai. Kya karein?" My father has fever after chemo. What should we do?

What happens next is the whole point of this post.

If an AI assistant replies with general advice from the internet, that is dangerous. If nobody replies until 10 AM, that is dangerous too. The only right answer is a system that recognises the message, gets it to a human who can help, and tells the family where to go right now if it is urgent.

AI and automation can take an enormous load off an oncology team. But in cancer care, the boundary between logistics and clinical questions isn't a nice-to-have. It is the first thing you build.

What should an oncology practice automate?

Plenty. Most of the messages your team sends every day are logistics, and logistics are exactly what automation does well.

Every one of these saves your navigator time and makes sure nothing is forgotten on a busy Saturday.

What must never be automated?

Here is the hard rule I would write at the top of your SOP: the AI books the appointment, the doctor gives the advice, never the other way round. In oncology, add: any message about how the patient is feeling goes to a human, every time.

That means the assistant never answers:

• Questions about symptoms, side effects, fever, pain or anything the patient is feeling.
• Questions about medicines, doses or whether to take something.
• Questions about results, scans or what a report means.
• Questions about treatment choices, prognosis or "should we continue".
• Anything emotional, angry or distressed.

For all of these, the assistant's only job is to say, honestly, "I am the practice's assistant and I can't answer medical questions. I am passing your message to our nurse now," and then actually do it.

How do you build the hand-off?

The hand-off is the most important part of the system. Map it on paper before anyone touches software:

1. Recognise the message. Set up keyword and topic triggers in English, Hindi and your local language, including common spellings like "bukhar", "dard", "ulti" and "saans". Keep the list broad. It is far better to send a logistics question to a nurse by mistake than the other way round.
2. Route it instantly. During working hours, to the nurse on duty, with an alert on their phone, not just a message in a shared inbox.
3. Set a response time. Agree how fast the nurse or doctor responds during working hours, and measure it.
4. After hours, give a clear route. The assistant replies with the practice's on-call number and a plain line: "If this is urgent, please call ___ now or go to the nearest emergency department." Then it alerts the on-call person.
5. Close the loop. Every clinical message is marked as handled, with who handled it and when. Review open ones every morning.

What guardrails should the assistant have?

From the way I describe building AI carefully in Rich Doctors Walk Away, with oncology's own additions:

• It only uses approved information: timings, address, fee sheet, document lists and appointment slots.
• It always says it is an assistant, never pretends to be a person.
• It never promises results or gives any clinical information.
• It never quotes a cost that isn't on the fee sheet.
• It hands over to a human the moment a conversation needs one, or someone asks for a person.
• It only messages families who have agreed to hear from you.

Any educational or promotional content made with AI should be clearly labelled, and never used to create patients, stories or results.

How do you test it before going live?

Have every team member send the assistant twenty messages as the most frightened relative they have ever dealt with. Mix languages. Spell things wrongly. Combine a logistics question with a symptom: "Cycle kal hai na? Aur unko kal raat se ulti ho rahi hai." Write down every answer that was wrong, vague or slow to hand over, and fix it. Then do it again.

Go live in phases: reminders and document collection first, then appointment changes, then anything else. Review a sample of real conversations every week.

How do you protect patient privacy?

Use a business WhatsApp setup with a shared inbox, not personal phones. Limit who can see clinical conversations. Don't send reports or results over automated messages. Keep the records inside your practice system. Tell families, in plain words, how their messages are handled.

What should you measure once it is live?

Measure:

• Share of cycle reminders confirmed.
• Missed appointments, compared with the months before automation.
• Time for a clinical message to reach a nurse, and to get a response.
• Number of clinical messages the assistant tried to answer itself. The target is zero.

If the last number is ever above zero, pause and fix before anything else.

This is the kind of setup we build in our CRM, WhatsApp and AI work for oncology practices, alongside the conversion systems that sit behind it.

The short version

• Automate logistics: reminders, documents, appointment changes, directions and insurance requests.
• Never let AI answer anything about symptoms, medicines, results or treatment.
• Build the hand-off first: recognise, route, respond, and a clear after-hours route.
• Label the assistant honestly, use approved information only and protect privacy.
• Test it like a frightened relative would, and track any clinical message it answered. The target is zero.