What the AI Listens for During Your Visit

A medical conversation contains a lot of words. Not all of them carry the same weight. A physician asking "how are you feeling today?" is clinically different from the same physician saying "I am adjusting your metformin to 1000mg twice daily." The challenge Kin has to solve is distinguishing the clinically important speech from the rest, in real time, without knowing in advance what the appointment will cover.

Understanding how Kin does this helps you interpret your summaries more accurately and gives you a clearer picture of what the technology can and cannot reliably do.

The four categories Kin focuses on

When we designed the structure of Kin's output, we started by asking what a patient actually needs to have in writing after leaving an appointment. That question produced four categories that appear consistently across almost every type of visit.

The first is clinical findings and diagnoses: what the doctor said about your condition, test results, or health status. The second is medications: anything prescribed, adjusted, stopped, or mentioned as a possibility. The third is action items: things the patient or caregiver is supposed to do after the visit. The fourth is conditional instructions: warning signs to watch for and what to do if you notice them.

These four categories capture what a patient needs to carry forward from a conversation. Kin's processing is organized around extracting these four types of information from the spoken record of the visit.

How Kin identifies medical language in natural speech

Medical conversations do not follow a script. Physicians shift topics, return to earlier points, hedge their statements, and use varying levels of technical terminology depending on who they are speaking with. A doctor might say "your HbA1c is 7.4, which is slightly elevated" in one visit and "your blood sugar control is a little higher than we want" in the next, both meaning the same thing clinically.

Kin uses a speech model trained specifically on medical vocabulary to handle this variation. The underlying model understands the relationship between clinical terms and their plain-language equivalents, which allows it to group related expressions together in the summary rather than treating each phrasing as a separate item. It also understands the contextual signals that mark a statement as clinically significant: phrases like "I am prescribing," "I would like you to," "your results show," or "if you notice" are linguistic markers that typically precede the kinds of statements patients most need to retain.

The treatment of medications specifically

Medications require particularly careful handling. The accuracy of medication information in a summary is not just a quality issue; it is a patient safety issue. A summary that lists the wrong dose, attributes a medication to the wrong person, or conflates a current prescription with one that was stopped could cause real harm if acted on without verification.

We have spent significant effort on medication accuracy. The model distinguishes between a medication being prescribed, adjusted, discontinued, discussed as an option, or mentioned historically. It captures dosage details and timing instructions when they are spoken. When the recording quality or the phrasing is ambiguous, the summary reflects that ambiguity rather than filling in a guess. You may see a medication listed without a dose, for example, if the dose was not clearly stated in the recorded conversation. That gap is intentional: an accurate record with a missing field is safer than a confident record with a wrong field.

Always verify any medication information in your summary against your prescription and your pharmacy records. Kin gives you a starting point and a written record. Your pharmacist and your prescribing doctor are the authoritative sources on what you should actually be taking.

Action items and follow-up instructions

The language of action items tends to follow recognizable patterns. "Schedule a follow-up," "come back in," "I would like you to get," "make sure to call if." These constructions signal that an instruction is being given, and Kin is trained to identify them and extract the specific actions they describe.

Conditional instructions require a slightly different approach because they involve a relationship between a trigger condition and a response action. "If the pain returns and is rated above a 7, go to the emergency room" is two pieces of information that have to be understood together to be useful. Kin captures these as linked pairs when they appear clearly in the recording. When the connection is ambiguous, the individual pieces are included but the relationship between them may be noted as uncertain.

What Kin does not do

Kin transcribes and summarizes the conversation that occurred. It does not have access to your medical record, your lab results, your imaging, or any clinical information that was not discussed in the room during the recorded visit. If your doctor reviewed a lab report and summarized the results verbally, Kin captures what the doctor said. The underlying lab data itself is not part of the record.

Kin also does not interpret or diagnose. If your summary includes a clinical term that is unfamiliar to you, Kin presents it as your doctor stated it and does not add clinical meaning that was not present in the conversation. Interpreting the medical significance of findings is your care team's role.

We also want to be clear that no automated system achieves perfect accuracy on spontaneous medical speech. Recording quality, ambient noise, accents, and speech patterns all affect the reliability of the transcription. Kin performs well across a wide range of conditions, but errors occur. That is why reviewing your summary and flagging anything that does not match your recollection of the visit is a useful habit. Your memory of the appointment and the summary should be read together, not as alternatives to each other.

The patient-side design intent

Most medical transcription tools that have come to market were built for the physician side of the conversation: to reduce documentation burden for clinicians, to populate electronic health records automatically, to support billing workflows. Those are real problems worth solving.

Kin's design focus is different. We built around the patient side of the conversation: what the patient heard, what they were told to do, and what information they need to carry forward into the rest of their care. The categories we extract, the language we use in the summary, and the way we structure the output are all oriented toward what is useful to a person managing their own health, not to a clinical documentation system.

That orientation shapes every decision we make about how the processing works, from which phrases trigger extraction to how ambiguity is handled to what format the summary takes when it arrives in your account.