A doctor opens the inbox at the end of a clinic day. There are lab results, refill requests, and messages from patients. Most can wait until morning. A few cannot. The hard part is knowing which is which before you have read all of them.

This is where AI belongs in medicine. It should do the sorting, so the doctor can do the judging.

What we mean by help

At Thyra, we build an electronic health record for independent practices. Our team includes a practicing physician, and that shapes one rule we do not bend: the clinician decides. Software can read faster than a person. It cannot be responsible for the patient.

So we ask a simple question of every AI feature. Does it give the doctor time back, or does it take the decision away? If it only takes the decision away, we do not build it.

The inbox is where this gets real

Inbox work is a good test because the stakes are uneven. A message about a parking question and a message about chest pain look the same in a long list.

Thyra's Smart Inbox classifies and routes incoming results and messages, so the ones that need action come up first. The doctor still reads the message and still decides what to do. The tool changes the order, not the responsibility.

What the evidence says

We are not the first to look at this, and we would rather point to real data than make claims of our own. Kaiser Permanente in Southern California studied an AI tool that labels patient portal messages by topic and urgency. The study covered more than 3 million messages from over 1 million patients, over two years.

  • The tool classified messages correctly 81.0% of the time, compared with 44.0% for the older system that relied on patients to pick a category.
  • For messages the tool flagged as high acuity, the time before a clinician first read them dropped by about 17 hours.

That is a result about triage. Nobody in that study handed the decision to a machine. The tool put the right message in front of the right person sooner. The study was published in JAMA Network Open in January 2026, and we link it below.

Where we draw the line

We hold ourselves to a few plain rules when AI touches a patient record.

  • AI output is labeled as AI, so nobody mistakes it for the clinician's own words.
  • A clinician reviews AI work before anything is signed or sent.
  • Review takes real time. A good design counts that time instead of pretending it is free.
  • When the tool is unsure, it says so and hands the case to a person.

The third rule is the one people skip. If a product saves ten minutes of typing but adds ten minutes of checking, it has saved nothing. We would rather be honest about that than sell speed we cannot show.

Why this matters for small practices

Independent practices do not have a floor of staff to catch what falls through. The doctor is often the last line of defense and the first person to feel the cost of a missed message. Good AI gives that doctor back the hour at the end of the day, without asking them to trust it blindly.

The goal is not a doctor who does less. It is a doctor who spends the day on the patient in front of them.

If you run a practice and want to see how Smart Inbox works on real clinic traffic, you can start a 30-day pilot.

Source: Performance of an Intelligent Messaging Tool for Clinical Triage, JAMA Network Open, January 2026. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2843966