Conversational AI in healthcare is software that talks with patients by voice or text and acts on what it hears: it books the appointment, checks the coverage, flags the symptom that needs a clinician. Vendors increasingly sell it on one number: how often patients fail to notice it isn't human. That is the wrong goal. Patients don't need an agent that passes for a receptionist. They need one they can trust while knowing exactly what it is.
The gap is already visible. ElevenLabs' latest industry report presents it as proof of success that 85 percent of callers to EliseAI's voice agents didn't realise they were talking to AI. Regulators are moving the other way. California's AB 3030 requires a disclaimer on AI-generated clinical messages, and Texas has required providers to disclose AI in treatment since January 2026. Trust in a patient conversation doesn't come from the voice. It comes from what the agent shows, asks, and admits.
What is conversational AI in healthcare?
Conversational AI in healthcare is an AI system that holds a natural-language conversation with a patient, a clinician, or an insurer, by voice or text, and acts inside the systems behind it. That last part separates it from the healthcare chatbot most patients remember. A chatbot follows a decision tree and breaks the moment someone answers off-script. A conversational agent understands what the person means, keeps context across the exchange, and does something with it: writes to the schedule, reads the claim, routes the case.
The use cases with real evidence today share two traits: high volume and a clear outcome. Kaiser Permanente's Southern California group added an AI tool called Intelligent Navigator to its patient portal, where patients describe what they need in their own words and get routed to the right care. Across nearly 3 million encounters, it detected urgent medical issues with 97.7 percent accuracy, and patients completed a booking more than half the time, against an industry average of 30 percent. Cigna's generative AI assistant answers questions about benefits, claims, and care options, and in early results more than four in five members found it helpful. At UnitedHealth Group, AI chatbots handled over 65 million customer calls in 2024.
Healthcare shares most of its conversational AI use cases with banking, retail, and telecom: scheduling, intake, reminders, account questions. What changes is the cost of a wrong answer.
The models already handle the conversation. Whether patients trust it depends on everything around it: what the agent admits, what it already knows, what it shows on screen, and when it steps aside.
Kaiser's navigator cut the share of patients who left the portal without finishing their task from 30 percent to 3 percent.
Kaiser's navigator cut the share of patients who left the portal without finishing their task from 30 percent to 3 percent.
Sounding human is the wrong goal
Sounding human makes an agent no more accurate, faster, or safer. What it changes is what the patient thinks they agreed to. Someone who describes their symptoms to what they believe is a nurse, then learns it was software, has to reconsider everything they said and everything they were told.
Honesty has a price, and it is small. In a Duke Health study of more than 1,400 patients, learning that AI had helped draft a message lowered satisfaction by about 0.1 points on a five-point scale. Satisfaction stayed high either way, and the researchers still recommend telling patients.
Regulators are writing disclosure into law, and the rules read like interface specs. AB 3030 covers generative AI messages about a patient's clinical information. Each one must say AI was involved and tell the patient how to reach a human. The scope is narrow: scheduling and billing messages are excluded, and messages a licensed clinician has reviewed are exempt. Texas goes broader and requires providers to tell patients, in plain language, when AI is used in their treatment.
Character is not the problem. Norvana, the AI health companion we designed and prototyped, has a deliberate personality: curious, a little playful, willing to disagree. Everywhere it is introduced, it is an AI companion, and nothing in the product pretends to be a person. The design question is not how human the agent sounds. It is whether the patient always knows what they're talking to, and how to reach someone who isn't AI.
That is the standard AI-native healthcare products should hold themselves to. Patients who know what they're talking to can decide how far to trust it. Patients who were fooled can't.
California's AB 3030 even specifies where the disclosure goes: at the top of a written message, at the start and end of an audio message, and on screen for the whole of a chat or video.
California's AB 3030 even specifies where the disclosure goes: at the top of a written message, at the start and end of an audio message, and on screen for the whole of a chat or video.
Context before questions
Most patient conversations start by asking for things the system already knows. Date of birth. Insurance. The reason for a visit the patient booked last week. Each question tells the patient the same thing: nobody here has read your file.
Conversational AI can reverse that order. Norvana's voice onboarding opens with what it already has. If Apple HealthKit shows the user averages 15,000 steps a day, the first thing Norvana talks about is their walking, and only then does it ask about their routine. The opening proves the agent has been paying attention before it asks for anything personal. From there, the model works from a list of topics it needs to cover rather than a fixed script, and follows whichever thread the user starts.
Norvana's onboarding runs 15 to 30 minutes, about as long as a short doctor's visit. Voice is what makes that length feel like a conversation rather than a form.
Norvana's onboarding runs 15 to 30 minutes, about as long as a short doctor's visit. Voice is what makes that length feel like a conversation rather than a form.
The same principle holds after onboarding. Norvana keeps the last three days of activity, sleep, and meals, plus any biomarkers outside normal ranges, available in every conversation. When a user says they've been tired lately, the agent can connect that to low iron in their last blood test without asking the user for it. Context only counts if it reaches the model. The quiet failure is an agent that answers as if it read the blood test when the data never arrived.
For a provider's voice agent, the equivalent is simple. Open a post-op follow-up call with the procedure, the surgeon, and the date the check-up is due. Ask only for what the record can't tell you. Every question the agent could have answered itself spends trust it hasn't earned yet.
The voice speaks, the screen confirms
Voice is the fastest way to collect information and the worst way to verify it. A patient can't re-read a dosage that was spoken once. They can't check a time slot they half-heard while driving. In healthcare, anything the patient will act on needs to land somewhere they can see it, check it, and change it.
Norvana splits the work this way. Conversation gathers context, and the screen carries anything that needs confirming. When a user scans a meal and the model isn't sure what it sees, it asks one question, chicken or turkey, as a tappable choice rather than a text prompt. When it is confident, it skips the question entirely. Whether to ask or to answer depends on the context the agent has, not on a designer mapping every state in advance.
The same logic shapes what the agent shows. Norvana composes its responses from a library of components: metric cards, biomarker visualisations, nutritional breakdowns. If a scanned meal affects a biomarker that is already out of range, the agent adds a card that connects the two. Health status uses plain words, such as "Needs attention" and "On track", instead of scores that turn health into a game.
For a booking agent, the pattern translates directly. The call confirms the slot by voice, then the slot appears as a card the patient can open, check, and cancel. The voice handles the conversation. The screen holds the commitment.
A number invites people to chase it. A word tells them where they stand. That is why Norvana replaced health scores with plain status labels.
A number invites people to chase it. A word tells them where they stand. That is why Norvana replaced health scores with plain status labels.
Escalation is a feature
Vendors report containment: the share of calls the agent finishes without a human. It is a useful operations metric and a dangerous design goal. An agent optimised to keep the conversation will keep it a turn too long. In healthcare, knowing when to stop is the core safety feature, and it needs the same design attention as the conversation itself.
The strongest deployments treat handover as part of the product. Kaiser's Intelligent Navigator, the triage tool behind the numbers above, doesn't try to handle severe complaints such as chest tightness. It connects those patients to a clinician right away. Cigna's assistant hands members to a human advocate when they need more help. Norvana asks clarifying questions when a user describes a worrying symptom, and when the answers point to something critical, it tells them to see a doctor. Blood test results get the same treatment: critical values surface as alerts and stand apart from the rest of the summary.
Escalation also means disagreeing. Early versions of Norvana agreed with everything. Tell it you planned a marathon after three strength sessions and it called that a great idea. One line in the system prompt, "you have your own opinion", turned it into an agent that pushed back and raised recovery first. An agent that never says no to a patient is not friendly. It is unsafe.
"You have your own opinion." One line in Norvana's system prompt turned an agreeable assistant into a companion that pushes back.
"You have your own opinion." One line in Norvana's system prompt turned an agreeable assistant into a companion that pushes back.
Language is an escalation trigger too. Many platforms now speak dozens of languages, but holding a conversation in a language and holding a clinical conversation in it are different things. About 28.5 million people in the US have limited English proficiency. In a pilot study of US hospitals, 49.1 percent of adverse events among these patients involved physical harm, compared with 29.5 percent for English speakers. An agent that can't hold a clinical conversation in the patient's language should recognise that early and hand over to an interpreter or a bilingual human, along with everything it has already learned.
That last part matters for every handover. When the agent steps aside, the clinician should receive the conversation, not a ticket. A patient who has to repeat their symptoms to a human after explaining them to a machine learns that talking to the machine was a waste of time.
The industry will keep shipping voices that sound more human. The agents patients return to will be the ones that say what they are, show what they did, and know when to stop.
How would your patients experience this?
Discuss with your AI.
A Product Strategist with over 13 years of experience in marketing, product strategy, and branding. His love for analytics, funnels, and a structured approach ensures that the digital products we craft aren't just functional—they impress.
