When my co-founder and I talk to nurses about what their shifts actually look like, we hear about interruptions. Lots of them. A 2018 study in the Journal of Emergency Nursing followed thirty-eight ED nurse shifts and counted an average of eighty-five interruptions per shift, which works out to roughly one every seven minutes.
Many of those interruptions are patient or family questions that don’t actually need a nurse’s clinical judgment to answer. A patient might want to know whether a particular sensation in their incision is normal, or what the surgeon meant when she said “soft diet.” Both questions matter, especially to the patient. But answering them for the tenth time on a single shift, while three other patients down the hall are deteriorating, is not what nursing school prepared anyone for.
We started Suvi Health because we thought patients deserved a place to get those questions answered without paging anyone. And care teams deserve support from AI that actually reduces burden instead of creating a more dangerous one: a hallucinating chatbot giving inaccurate guidance to a patient.
The first problem is obvious. The second is the real work.
Most consumer AI products manage hallucinations statistically rather than eliminating them architecturally. There are well-known engineering techniques that reduce hallucination rates — retrieval augmentation, grounding, constrained generation, verifier models, fine-tuning, uncertainty calibration — but none reduce the rate to zero. That’s because modern language models are fundamentally probabilistic systems trained to predict likely continuations, not deterministic systems designed to guarantee truth. For most consumer applications, that tradeoff is acceptable. The downside of a confidently wrong restaurant recommendation is usually a mediocre dinner. The downside in healthcare can be much larger.
In a clinical setting, the downside is different. A patient who’s told a side effect is normal when it isn’t can sit at home until something becomes irreversible. The acceptable rate of confidently wrong answers in that setting is zero, and “statistically reduced” is not zero.
So when we designed how the Suvi Health voice agent and chatbot would handle questions from patients and caregivers, we started from a different premise. The model wouldn’t be allowed to answer from its own training data. Instead, it responds exclusively using information drawn from prior conversations with the care team and trusted, clinically approved materials.
The approved sources include the patient’s conversations with their care team, along with discharge instructions, assigned educational materials, and a broader set of clinical references approved for the populations that organization serves.
What’s intentionally excluded is everything else. The model does not pull from the open web or rely on its generalized training corpus when answering patient questions, eliminating the primary pathway through which hallucinations typically occur.
When a patient asks something, the system retrieves from the approved sources and constructs an answer in plain language. Every claim in that answer is tied back to where it came from. If the cardiologist said “no driving for two weeks” at the bedside on Tuesday, the answer cites that statement. Both the patient and the care team see the citation, and the audit trail captures it. Attribution isn’t a logging feature we added at the end. The system won’t produce an answer without one.
Here’s the part we spent the most time on. What happens when a patient asks something the approved sources don’t cover?
A consumer chatbot, in that situation, will usually attempt an answer anyway. That’s the design intent of those products: be helpful, fill the silence above all else. In our setting, the instinct is exactly wrong. A patient asking whether they can take an over-the-counter sleep aid that nobody on their care team has discussed with them does not need a confident-sounding paragraph generated from statistical regularities about sleep aids in general. They need the question routed to someone who can actually decide.
So Suvi Health, in those cases, doesn’t try. It tells the patient that this is a question for their care team, and explains that it has sent the question over. The phrasing is plain and the next step is concrete: a specific clinician on their care team will see the question with the conversational context attached.
We’ve been describing this internally as knowing when to ask, and the framing matters. The patient experience isn’t “the AI refused me,” which would feel like getting bounced from a customer support chat. It’s closer to what happens when a junior team member says “I want to get this right, let me check with my senior.” Trust goes up, not down.
The questions that get handed off go into what we call the parking lot inside the Suvi Health care team dashboard, where each one is attached to the patient who asked it and the context that produced it. From there, a clinician can do one of a couple of things. They can answer asynchronously through Suvi Health between cases (which is how most of these tend to get cleared, often faster than a return phone call would manage). Or, if the question is better addressed in person, they can bring it up at the next round and use the parking lot as a kind of pre-loaded rounding agenda.
Either way, Suvi Health closes the loop with the patient and attributes the answer to the clinician who gave it. Nothing gets lost, and nothing that needed a human gets handled by a model.
The realistic alternative to Suvi Health in most hospitals today is not “the patient asks no one.” It’s the patient pulling out their phone at midnight and asking a general-purpose chatbot. We didn’t fully appreciate how common this was until we started doing user research. It turns out to be pervasive, and the privacy implications are uncomfortable.
A consumer chatbot has no legal obligation under HIPAA to protect what a patient types into it. Conversations are typically logged for later use, including model training, and under some terms of service they get shared with third parties most patients have never reviewed. When a patient pastes their symptoms and current medications into one of those chatbots, they’re handing protected health information to a system with none of the protections that information would have if they were emailing their hospital’s portal.
Suvi Health operates inside a HIPAA-compliant boundary. The conversations and the questions are treated as protected health information from the moment they enter the system, with the same legal protections the rest of the patient’s record gets. Both Suvi Health and a consumer chatbot will give the patient an answer of some kind. Only one of them is bound by law to protect what was asked.
Eighty-five interruptions per shift is just the start of the story. The cost isn’t only the time the interruptions take. Every interruption pulls a nurse’s attention out of whatever clinical reasoning she was doing before it, and the cognitive cost of switching back is substantial. The research on attention-switching in clinical settings is unambiguous on this point, and it tracks with what every clinician we’ve worked with has told us in their own words.
Stack enough of those interruptions on top of the documentation burden and the moral weight of caring for sick people, and you get the thing the field has been calling burnout for years. Nurses leave the profession over it. The 2026 NSI National Health Care Retention & RN Staffing Report put the average cost of replacing a single staff RN at $60,090, and estimated that the average hospital loses roughly $5.19 million a year to RN churn. Those are the financial numbers, the kind an admin can put on a spreadsheet. The numbers nobody can put on a spreadsheet are the patients who got worse care during the months a unit was running short, and the experienced nurses whose institutional knowledge walked out the door with them.
What we hear from the clinicians is that the burnout-driving interruptions are disproportionately the well-bounded ones. These are the questions a patient asks because nobody has explained something clearly enough, or because they need reassurance at 2 a.m. and don’t want to bother a nurse over it. Suvi Health is built for those questions specifically. The parking lot surfaces the rest, sorted and with context, so the care team can work through them when they have the bandwidth.
Lookup work goes to the AI. Clinical judgment work goes to the clinicians. That’s what practicing at the top of your license is supposed to look like.
From hospitalization through recovery at home — we carry clarity forward.
