The honest answer is that we rely on intent signals rather than keywords alone. Early on, we made the mistake many teams make, in that we leaned on topic classification. Someone asking about a delayed critical-medicine delivery was routed and handled in exactly the same way as someone asking about an over-the-counter product. That was never a data problem; it was a design failure in how we had defined urgency in the first place.

What we have moved toward now is a composite signal that combines several inputs. We look at order criticality, distinguishing chronic medication from an over-the-counter purchase. We look at contact history, which tells us whether this is a first-time query or a repeat escalation. We factor in the time-sensitivity of the product, and sometimes even the tone of the message itself. A person who types in capitals and uses the word urgent three times is simply not in the same state of mind as someone calmly asking for a refund status, and the system needs to recognize that difference.

Genuine digital empathy does not come from a chatbot saying that it understands the situation must be difficult. It comes from the system actually doing something different, whether that is faster routing, a proactive callback, or simply not making someone repeat themselves to yet another agent. That is where AI has really moved the needle for us, and it has done so through action far more than through language.

I will be direct in saying that we are firmly in the middle of that spectrum, and I believe most healthcare platforms in India are still evolving through this maturity curve. We have moved well beyond scripted chatbot flows, and we now have AI that can handle a good percentage of customer queries without human involvement, covering areas such as refund status, order tracking, and prescription queries, all at an acceptable level of resolution quality. That part is real.

Cognitive CX that anticipates patient needs at scale is a different proposition altogether. It requires longitudinal data quality and behavioural consistency that most Indian healthcare platforms, including ours, are still in the process of building. The more interesting frontier for us is the shift from reactive resolution toward proactive intervention, where the system surfaces adherence-risk signals before a patient misses a dose, rather than after they call to complain about a gap in their refill cycle. We are running early experiments in that direction, though I would be careful not to present those experiments as business as usual just yet.

This is the question I think about most, and the uncomfortable truth is that the answer depends entirely on how you choose to implement it. If you deploy Agent Assist as a script-recommendation engine and measure success purely on handle time, you will end up with faster operators. They will close tickets more quickly, and your average-handle-time dashboard will look excellent. Your CSAT, however, will quietly erode, because patients can feel when someone is reading from a playbook rather than thinking about their problem.

The version that creates better caretakers is one where AI handles the cognitive overhead on the agent’s behalf. It surfaces order history, flags prior complaints, shares the customer profile, and pulls up the relevant and applicable policy, so that the agent’s mental bandwidth is freed for the human part of the conversation, which is listening, taking ownership, showing empathy, and exercising judgement. We have seen this distinction play out internally, and by using AI to help our agents think better, we are actually seeing improvements in first-contact resolution rather than just a reduction in handle time. The harder challenge was never deploying AI. It was redesigning workflows, retraining teams, and aligning operations around new customer expectations.

Let me be honest about NPS, which behaves rather like a postcard from the past, since by the time it arrives the customer has already decided how they feel about you. The leading indicators we have found to be predictive are somewhat different. The first is repeat-contact rate within 48 to 72 hours, which gives us a near-real-time signal that the first resolution has failed. The second is escalation velocity, meaning how quickly a new contact type moves up the queue. The third is the drop in unassisted containment on specific intent clusters, because if the AI resolution rate on a particular query type suddenly falls by 10 to 15 percent, it tells us that something has changed in the product, the policy, or customer expectations.

The most underrated signal we track is what I think of as silent dissatisfaction. These are orders where a customer has received their delivery but has shown no engagement for 30 or more days on a chronic medication that should be refilling. That silence is often far more telling than a complaint would be. Voice of Customer with AI, in our view, has little to do with sentiment scores and everything to do with building early-warning systems, so that we are working from the equivalent of a fever chart rather than a post-mortem report.

In healthcare CX, the real challenge has very little to do with access to data and almost everything to do with restraint. You will always face a tension between personalisation and intrusion, and it is not something you resolve once and forget. You manage it intelligently and empathetically, with intent, every single day. There is a clear line that guides us, which is that using patient history to ensure continuity of care creates real value, whereas using behavioural signals simply to push engagement risks breaking trust.

Trust is our real currency. Patients who trust you will share more, and patients who feel watched will pull back, and in healthcare that pullback is far more serious than ordinary churn, because it can mean missed care. It directly impacts retention, adherence, and long-term platform engagement. The mandate I work to is a simple one, which is to remove friction without ever creating pressure. If a particular use of data does not make the patient’s healthcare journey better, then it is just noise, and noise has no place in care.

The metric I care about is cost-per-resolution rather than cost-per-contact, and the two are different things. Deflecting a contact with a chatbot that does not actually resolve the issue may reduce your cost-per-contact, but it destroys your cost-per-resolution, because the customer comes back to you a second time, by now angrier and more expensive to serve. AI reduces cost in a good way only when it resolves the problem, not when it merely deflects it, and that distinction matters enormously for how you design your whole automation strategy.

Human intervention should be preserved as an intentional product feature in high-stakes moments, rather than treated as a fallback. I am thinking of first-time chronic-medication orders, prescription-upload complications, and post-surgery or post-discharge queries. These are not edge cases at all; they sit at the core value proposition of any healthcare platform that takes its responsibility seriously. If your AI strategy is built purely around reducing headcount, you will eventually make a mistake in a medical context that no efficiency gain will ever compensate for, and so the return-on-investment conversation has to include the cost of that risk.

The most important thing missing from most conversations about AI in healthcare is honesty about the stack underneath it all. AI does not fix broken data pipelines, undertrained agents, or unclear ownership of CX outcomes, and pretending otherwise helps no one.

The platforms that will get this right are not the ones with the most sophisticated models. They are the ones that did the unglamorous work of cleaning their data, defining their metrics clearly, and building organizations in which humans and AI are complementary rather than adversarial. That is the work that actually needs doing, and by comparison the technology is the easier part. The future of healthcare CX will be defined less by how much AI replaces humans and more by how intelligently organizations decide where humans matter most.