A paper in the American Economic Review this month compared nurse practitioners and physicians across 1.1 million emergency department visits. Physicians came out ahead on average. NPs came out ahead in 38 percent of random head-to-head matchups, and thirty-day mortality showed no significant difference (1)(2). I wrote separately about what that study does and does not say.

This piece is about the setting I actually work in. Every visit I do is a video visit, and virtual urgent care has quietly become one of the largest places in American medicine where the question of who is on the other end of the call gets settled by economics rather than by clinical reasoning. The emergency department findings do not carry over to it cleanly, and the reasons they do not are worth their own article.

What changed about who answers

The staffing model in this corner of medicine changed fast, and most patients never noticed. Direct-to-consumer telemedicine launched staffed largely by physicians and runs now largely on nurse practitioners and physician assistants. The reason was never clinical. An NP costs less per hour and is easier to hire at scale, and when the product is a cheap video visit available at two in the morning, that arithmetic decides the roster. About 13 percent of emergency visits nationwide are now handled by NPs (2). In virtual urgent care the share is far higher, and those platforms are built for volume and speed.

The training gap is worth stating precisely, and I would rather use nursing education’s own accrediting standards than anybody’s talking points. The National Task Force standards that govern NP programs set a minimum of 750 direct patient care clinical hours for a population-focused track (4). No residency is required after that. A physician arrives at the same video visit having done clerkships and then residency, which the AMA puts somewhere between 10,000 and 16,000 hours (3). And about 87 percent of the 461,000 licensed NPs in this country trained in programs focused on primary care (5), which is preparation for the undifferentiated but mostly well patient rather than for the sick one.

Now let me argue against my own instinct, because the data do not go where a physician would expect. If thinner training meant missing sick patients, NPs in the Chan and Chen cohort should have under-admitted the dangerous presentations. They did the reverse. For sepsis, stroke, and heart failure they were substantially more likely to admit (2). The signature of that whole paper is a lower threshold to spend a resource when the picture is murky, which is caution rather than blindness, and it predicts over-referral out of telemedicine rather than under-referral.

Here is why I do not think that settles the question. The emergency department is a rigged environment for this comparison, rigged in a way that flatters everyone working inside it. The patient is physically present. A triage nurse has already taken vitals and assigned an acuity level. Somebody has laid eyes on them before the clinician decides anything, which means most of the signal that says this one is sick arrives free. Take that away, which is precisely what a video visit does, and the task becomes generating suspicion from a history and a small rectangle of video. Chan and Chen did not measure that task, and as far as I can tell nobody has.

Drop that mechanism into a video visit and the two available levers are prescribe empirically or send the patient somewhere with hands, which are exactly the two the study found NPs pulling more often. In telemedicine that shows up as more empiric courses and more requests to be seen in person today.

There is already independent evidence pointing the same direction. Ray and colleagues looked at children treated for acute respiratory infections and found antibiotics prescribed at 52 percent of direct-to-consumer telemedicine visits, against 42 percent at urgent care and 31 percent at a primary care office. Guideline-concordant management ran 59 percent in telemedicine versus 78 percent with the child’s own doctor (6). That was a finding about the modality, not about who was staffing it. Stack a modality that pushes toward empiric treatment on top of a staffing model that also pushes that way, and you have compounded the same bias twice.

The part nobody audits

Every telemedicine urgent care company writes its own clinical practice guidelines. Sinusitis follows this pathway, dysuria follows that one, and here is the list of complaints that must be routed to an in-person evaluation no matter how well the patient looks on camera. The documents are usually sensible. The problem lives downstream of the document.

They are not always followed, and at these volumes nobody is checking most of the time. Chart review is a sampling exercise and the sample is thin. A clinician who has drifted off the pathway, who has learned that the fastest route through a shift is a prescription and a click to the next patient, can drift a long while before anything notices. That is not a nurse practitioner problem. I have watched physicians do the same thing under the same incentives. It does interact badly with a workforce carrying fewer residency-trained reps for the case the guideline did not anticipate, and those are the visits that hurt people.

If I ran one of these companies, the number I would want on the dashboard is one I have never seen a platform publish. What fraction of visits departed from the pathway, broken out by clinician, with the bad outcomes tracked forward past the end of the encounter. Visit volume gets measured continuously. Guideline adherence gets measured occasionally. What happened to the patient afterward gets measured close to never.

Spotting the sick one is the whole job

Strip a virtual urgent care shift down to its clinical core and it is one decision repeated all day. Not what is the diagnosis. Does this person need to be somewhere with hands, equipment and a laboratory, and do they need to be there now.

Nearly everything else is forgiving. Choose the wrong antibiotic for a sinusitis and the patient is annoyed and books a second visit. Miss the atypical presentation of an acute coronary syndrome, or a subarachnoid bleed, or an ectopic pregnancy, or the child whose respiratory rate nobody counted because nobody asked, and that visit becomes the first line of a malpractice complaint or worse.

The instruments available for that decision are thinner than patients realize. No vitals unless the patient owns a cuff and a pulse oximeter and knows how to use them. Nothing to palpate. No chance to watch someone walk across a room, which is one of the most informative things I ever did in person. What remains is the history, and a history taken properly is the entire safety margin. The questions that carry weight cost time a productivity dashboard works against.

Now add the payment model, which pulls the wrong way all shift. Most virtual urgent care pays by the visit, or by the hour with a visit target attached. Nobody gets paid extra for the encounter where you spent eleven minutes on a history, decided something was off, and sent the patient to an emergency department without writing a prescription. On the dashboard that was your least productive visit of the day. It may have been the only one that mattered.

That visit scores badly a second time after you log off, because the patient rates you and at many platforms the rating feeds into your pay. Satisfaction scores capture real things, whether you listened and whether somebody felt taken seriously at eleven at night, and they are worth measuring. The trouble is timing. Satisfaction can be measured the second a visit ends. Whether the decision was correct often cannot be measured for weeks. The fast signal gets attached to the paycheck and the slow one does not, which is a design problem rather than anybody being a villain.

That pull is documented, and the study documenting it looked at physicians rather than nurse practitioners, which is rather the point. Across 8,437 direct-to-consumer telemedicine visits for respiratory infections handled by 85 physicians, 66 percent ended with an antibiotic. Patients who received one gave five stars 90.9 percent of the time, against 72.5 percent for those who left with no prescription (7). Individual physicians prescribed antibiotics anywhere from 19 to 90 percent of the time, and the heavier prescribers scored better. None of this is a telemedicine invention. Fenton and colleagues reported more than a decade ago that the most satisfied patients in a national cohort carried higher expenditures and higher mortality (8), an association rather than a mechanism and argued over ever since.

Nobody should stop asking patients what they thought. But a rating collected sixty seconds after the video ends cannot see whether the person who needed to be sent somewhere actually got sent. Until that number sits beside the satisfaction number, the clinician who declines the antibiotic and recommends the emergency department absorbs a cost nothing credits back.

A specific failure mode grows out of that, and it has nothing to do with anybody’s credential. You see forty upper respiratory infections across a shift. Thirty-nine are exactly what they appear to be. Your threshold quietly resets, because it has been rewarded for resetting all day, and the fortieth looks like the other thirty-nine right up until you ask the question that separates them. Volume does this to everyone. I have felt it happen to me.

This is where inpatient training earns its keep, and the claim is easy to overstate so let me be precise. I am not saying patients should see only physicians. I am saying the particular skill at issue, sorting the person who needs closer evaluation from the person who can be managed at home, gets built by having been responsible for patients who were admitted and then deteriorated overnight. Residency is years of exactly that, with somebody senior checking your reasoning. It is not the only route to that skill. It is the most reliable one we have.

Referral also assumes a mechanism exists, which is something you learn fast working inside one of these services. A good deal of virtual urgent care cannot order anything at all. Not a basic metabolic panel, not a chest film, not a cardiology consult. Payers frequently will not honor an order originating from a telemedicine encounter, reserving that authority for a primary care physician or an in-person urgent care, and platforms understandably do not build plumbing for something insurers will not pay for. Some services do have that capability. Several large national ones do not, and a patient cannot tell which sort they have reached before the visit begins.

Which narrows the toolkit further than most patients imagine. If I cannot send someone for a troponin, the disposition decision stops being one tool among several and becomes the only one. Every bit of uncertainty has to collapse into a single sentence. This can wait. Go be seen today. Go now.

That may loosen. Curbside specialist consults built into the encounter, where a cardiologist or a dermatologist can be pulled onto the call while the patient is still sitting there, are beginning to appear and would change this arithmetic considerably. They are not part of everyday practice yet. Plan around what exists today.

One more piece gets ignored. Referring is not the same as protecting. When a telehealth visit generates a recommended test or referral, patients complete it less often than when the same recommendation comes from an in-person visit. Across 4,133 orders, completion inside the expected window ran 58 percent for in-person against 43 percent for telehealth (9). That study looked at elective primary care work rather than emergency escalation, so do not stretch it. The direction still matters. Telling somebody to go to the emergency department is a recommendation, not a handoff, and a real fraction will not go.

Which brings me to the one staffing opinion I will state without hedging. Virtual urgent care is a poor first job. A nurse practitioner or physician assistant fresh out of school should not be making disposition decisions alone, on camera, without vitals, at production speed. Neither pathway requires a residency, so nobody has yet watched them be wrong somewhere it got caught and corrected. Spend a few years somewhere with a hallway and a colleague in it.

So when patients ask whether it matters that they drew a nurse practitioner instead of a physician, this is my honest answer. On the ordinary complaint, which is most of them, it does not matter much, and the 38 percent figure is why I say that with a straight face. On the visit where something is quietly wrong and nobody has said so out loud yet, training and repetitions matter more, and the average NP has fewer of both. That is not about anybody being careless. It is what 750 hours buys against several thousand, and any NP a decade into acute care has closed most of that distance.

For patients

You can ask who you are seeing and what their background is, and no reasonable clinician will take offense. Do not treat the letters after the name as the whole answer, because a randomly chosen NP outperforms a randomly chosen physician 38 times out of 100 (1). How long they have been doing acute care is the more useful question.

Two practical things. If you call with chest pain, a severe headache, abdominal pain, or a child who is breathing fast, expect to be sent somewhere with hands and equipment, and understand that this is the visit working rather than failing. And if you are told to be seen in person today, go. On many platforms that instruction is the entire product of the encounter, and it only helps if you act on it.

For clinicians

If you build or run these services, the actionable finding from the emergency department data is the complexity gradient rather than the average effect. Route sore throats and rash checks broadly. Route diagnostic ambiguity and high-acuity complaints to whoever has the most reps with them, and make that assignment on individual performance data rather than on license class. That is a solvable engineering problem and almost nobody is solving it.

If you work in them, two suggestions. Keep a private count of how many visits in a shift you closed without writing a prescription, because that number tells you something a satisfaction score cannot. And treat your last hour differently from your first, since threshold drift is worst when you are tired and the pattern has looked the same all day.

The Bottom Line

Virtual urgent care asks one question of a clinician, over and over, all day. Does this person need to be somewhere with hands, equipment and a laboratory, and do they need to be there now. Nearly everything else about the visit is forgiving.

The pressures working against that question are structural rather than personal. Payment by volume. Ratings collected sixty seconds after the video ends. Guidelines nobody has the capacity to audit at scale. And on many platforms no ability to order a test or generate a referral at all. None of that is a nurse practitioner problem and none of it is anybody being a villain. It is what happens when the fast measurements get attached to the paycheck and the slow ones never get taken.

If I were staffing one of these services, I would not put anyone into this work as a first job. And if I were the patient, I would care far less about the credential on the screen than about how many years the person behind it has spent deciding who was sick.