It is 2:15 in the morning at a 180-bed hospital in a mid-sized American city. A hospitalist is standing at a workstation looking at a 34-year-old woman with a fever that will not break, ferritin above 15,000, falling platelets, and a liver that is starting to misbehave. She has read the differential. She knows what the pattern might be. She knows that if it is what she suspects, the next 48 hours matter enormously and the treatment decision is not something you get from a summary.
What she wants is not an article. What she wants is a specific sentence from a specific human being: "I have seen eleven of these. Here is what the third day usually looks like, here is the thing that fooled me the first time, and here is my cell number if it goes the other way."
She is a physician in a country with more than a million active physicians. Somewhere among them, several dozen have managed exactly this presentation. She has no way to find a single one of them tonight.
So she does what everyone does. She texts two people from residency. One is asleep. One replies at 6:40 a.m. with a thoughtful paragraph and an apology that this is not really their area. She posts a de-identified version in a group chat. She reads more. She makes a good decision, probably, on her own, at 3 a.m., and she never finds out whether the person with eleven cases in their head would have said something different.
That is not a knowledge problem. The knowledge exists. That is not an access problem in the usual sense. She had a phone, a license, an internet connection, and colleagues. That is a routing problem, and it is the most expensive unsolved problem in the professional infrastructure of medicine.
The sentence that gives the whole thing away
Walk into any hospital, any group practice, any department meeting, any nursing station, and listen for a while. You will hear a version of the same sentence, over and over, from people at every level of seniority:
> "Does anyone know someone who..."
Does anyone know someone who has managed pregnancy in a Fontan patient. Does anyone know someone who has done this revision. Does anyone know someone who has actually run an ambient scribe evaluation. Does anyone know a psychiatrist taking new patients. Does anyone know someone at another center who has been through this.
That sentence is a diagnostic finding. It tells you, unambiguously, that the profession has industrial-grade infrastructure for proving that a person is qualified and almost nothing for finding the specific person who knows the specific thing.
Sit with how strange that is.
Medicine built the most verified workforce on earth
Give the profession its due, because the achievement is real and it is enormous.
Before a physician touches a patient in the United States, the following has happened. They were selected from a competitive applicant pool. They completed a four-year doctoral degree. They passed a multi-part national licensing examination. They completed three to seven years of supervised residency training under a national accreditation body. Many completed additional fellowship training. They obtained a state license, verified through primary-source verification of every credential. They obtained board certification through a specialty board and then maintained it through recurring assessment. They were credentialed and privileged by a hospital, which independently verified all of the above, queried the National Practitioner Data Bank, checked references, and reviewed their claims history. Then a payer credentialed them again. Then another payer credentialed them again.
The Federation of State Medical Boards counted 1,082,187 actively licensed physicians in its 2024 census. The Association of American Medical Colleges counted 1,032,365 active physicians in 2024, of whom 866,460 were in direct patient care. The World Health Organization puts the global health workforce above 70 million people. Every one of them sits inside some version of that verification apparatus.
No other profession on earth is checked this thoroughly. Not law. Not aviation. Not finance.
And here is the punchline. After all of that verification, the system can tell you that Dr. Reyes is board certified in gastroenterology and works at St. Anne's. It cannot tell you whether Dr. Reyes has performed a third-space endoscopic procedure in the last three years, whether she has ever managed the complication you are staring at, whether she would take a call from a stranger at 2 a.m., or whether she is even still doing gastroenterology in any recognizable sense.
The system knows where she works. It does not know what she knows.
Naming the thing: the credential-expertise gap
Let us give this a name, because naming it is how you start measuring it.
The credential-expertise gap is the distance between what a directory records about a clinician and what that clinician actually knows and does right now.
Directories record credentials because credentials are what institutions are legally and financially obligated to verify. A hospital must verify a license to grant privileges. A payer must verify certification to pay a claim. Every entry in every healthcare directory exists to satisfy a compliance requirement, and compliance requirements are about eligibility, not capability.
Nobody, anywhere, is paid to record what a clinician actually knows.
Think about what that means in practice. "Board certified in cardiology" is a fact about an examination, often taken decades ago. It says nothing about who does cardiac sarcoidosis, who does adult congenital, who has managed forty cases of a rare cardiomyopathy, who has abandoned catheterization entirely for a device practice, or who is the person other cardiologists quietly call when they are stuck.
That last category is the most valuable professional information in medicine. It exists only in individual human memory. It is transmitted only by conversation. And it dies when people retire.
The passport analogy
Here is the cleanest way to see the shape of the problem.
Imagine a country with the finest passport control system ever built. Every traveler is biometrically verified. Every document is authenticated against its issuing authority. Fraud is essentially impossible. The system can tell you with total confidence that the person in front of you is who they say they are and is legally entitled to be here.
Now imagine that this country has no maps, no street signs, no addresses, and no directory. You are verified beyond all doubt, and you cannot find anyone.
That is medicine. Extraordinary verification. No wayfinding.
The reason this persists is not incompetence. It is that verification has a paying customer (the institution that would otherwise bear legal and financial risk) and routing does not. Nobody gets sued for failing to help a hospitalist find the right person at 2 a.m. There is no line item for it. So it does not exist.
Three graphs medicine does not have
To move from complaint to something buildable, you need to be specific about what is missing. Three distinct structures are absent, and conflating them is why previous attempts have failed.
1. The expertise graph: who knows what
This is the map of what people have actually done, at what volume, how recently, in what setting.
Not "trained in interventional cardiology" but "has performed 40 to 60 of these per year for the last five years, including 12 in patients with this specific anatomy." Not "pediatric neurology" but "has personally managed nine children with this ultra-rare epilepsy syndrome and knows what the second-line agent does at month four."
Nothing in healthcare stores this. Board certification does not. Employment records do not. Publication records emphatically do not, because the clinicians with the deepest hands-on exposure often publish least. A prolific author and a high-volume operator are frequently different people, and only one of them is legible to any existing system.
Exposure has a peculiar mathematical property that matters enormously for design. For any given rare condition, exposure is sparse per clinician but dense across a large population. Almost nobody has seen it; somebody definitely has. That is precisely the shape of problem that networks solve and directories cannot, and it is the mathematical reason scale matters here in a way it does not matter for, say, a content feed.
2. The trust graph: who is actually turned to
This is different from the expertise graph and often confused with it.
The expertise graph says who knows. The trust graph says who gets asked. The two overlap far less than you would expect, and the second one is the one clinicians actually use.
The trust graph in medicine is real, measurable, and almost entirely invisible to software. Its strongest known edge is co-training. Research published in Health Services Research examining more than 40,000 referrals found that primary care physicians referred to physicians they trained with at a rate of 26.2 percent versus a 21.4 percent baseline, a statistically significant preference, and that the effect was driven by residency and fellowship co-training rather than by medical school. People trust the people they were in the trenches with, and they keep trusting them for decades.
Notice what this implies. Medicine already has a large, durable, high-quality trust graph. It is encoded in who trained where and when. And essentially no clinical system uses it, because the person you trained with usually works for a competitor, and no institution has any incentive to help you route work outside its own walls.
This is a recurring pattern worth internalizing: the most valuable professional relationships in medicine cross institutional boundaries, and every existing system stops at those boundaries.
3. The memory graph: who learned what, and where it went
The third missing structure is institutional and professional memory.
Healthcare preserves documents obsessively and destroys context routinely. The protocol survives. The person who knows why step three exists, what they tried before that failed, and which patient population it quietly does not work for, leaves.
The rate of that loss is not subtle. Registered nurse turnover in US hospitals ran 17.6 percent in the most recent NSI National Health Care Retention Report, at an average replacement cost of $60,090 per departing bedside nurse and $5.19 million a year for the average hospital. Research published in The Permanente Journal found physicians now leaving clinical practice at a mean age of 48.1, roughly nine years younger than an earlier cohort. Hospital executive turnover runs near a fifth per year.
Every one of those departures is a deletion. And because thousands of organizations are deleting and rewriting simultaneously, the aggregate effect is that healthcare pays over and over to rediscover things it already knew.
Exhibit A: the largest consult channel in medicine has no router
If you want to see the credential-expertise gap operating at scale, look at the curbside consult.
The curbside is the hallway question, the text to a friend, the "quick one for you" at the coffee machine. It is not a fringe behavior. It is arguably the highest-volume clinical decision channel in the profession.
A study published in JAMA surveying subspecialists found that 87.5 percent had fielded at least one curbside consultation in the prior week, at a mean of 3.6 per week, while primary care physicians requested about 3.2 per week. More recent work suggests the channel has simply migrated to phones. A 2026 study in PLoS One at a tertiary center reported that 98.9 percent of clinical consultations between colleagues ran through WhatsApp, with 63.7 percent of respondents expressing concern about the legal implications.
Do the arithmetic and the scale is startling. If even 300,000 subspecialists field 3.6 curbsides a week at roughly ten minutes each, that is on the order of nine million hours a year of expert clinical judgment, given away free, recorded nowhere, credited to no one, and routed entirely by who happens to be in your phone.
Now the uncomfortable part. A study in the Journal of Hospital Medicine compared curbside advice against a formal consultation on the same patients. The information conveyed in the curbside was inaccurate or incomplete in 51 percent of cases. Management advice differed after formal consultation in 60 percent of cases, rising to 92 percent when the initial information had been wrong.
That is not an argument against curbsides. Curbsides are fast, collegial, and often exactly right. It is an argument about infrastructure. The profession runs an enormous, high-stakes decision channel with no routing, no record, no quality layer, and no compensation for the person answering. Research presented at IDWeek in 2025 found that more than 70 percent of internal medicine and infectious disease clinicians had received no formal education whatsoever in how to give or receive a curbside.
And here is the structural point that explains why this has never been fixed. Formal alternatives exist and work well. Electronic consultation programs report turnaround times of 1.2 to 2.6 days and avoid somewhere between 32 and 70 percent of face-to-face specialist referrals, with payers paying real money per consult. So why has the curbside not simply been replaced?
Because e-consults are bought by an employer or a payer and route only within a contracted panel. They work beautifully inside the walls. The hard question, the one where you genuinely do not know anyone, is by definition the one that crosses the walls. The channel that would help most is precisely the channel nobody sells.
Exhibit B: the diagnostic odyssey is a search failure wearing a lab coat
Now take the same gap and look at it from the patient's side.
The rare disease diagnostic odyssey is usually described as a knowledge problem: these conditions are obscure, most clinicians will never see one, and diagnosis is genuinely hard. All true. But look closely at where the years actually go.
The odyssey averages roughly 4.7 years in Europe and 7.6 years in the United States. Patients see up to eight physicians and collect two to three misdiagnoses along the way. The EveryLife Foundation puts it at more than six years and nearly 17 doctor visits, hospitalizations, and health-related trips. German claims data quantified the cost precisely: direct costs of the pre-diagnosis period ran EUR 26,999 per suspected rare-disease patient against EUR 3,561 for matched controls, a 7.6-fold excess, with 7.3 specialists consulted and 50 distinct working diagnoses recorded per patient.
Fifty working diagnoses. Sit with that number for a second. That is not a knowledge vacuum. That is a system generating hypotheses energetically and failing to route the patient to the one person whose pattern recognition would collapse the whole thing in a single visit.
And that person usually exists. Often in the same country. Sometimes in the same health system.
What does a clinician do today when they suspect something rare? A study in Belgium found that 73 percent of general practitioners could not name a single rare-disease information source, and that across physician groups, 51 to 63 percent said what they actually wanted was direct contact with an expert. Not a database. A person.
So they ask a colleague. They post in a group chat. They email a former attending. They search PubMed for authors and cold-email strangers. They try Orphanet or a centers-of-excellence list, which lists institutions, not the specific human with the pattern in their head. Or they refer to the nearest plausible subspecialty and hope.
The formal routing mechanisms that do exist show exactly how hard this is to solve institutionally. The European reference network system for rare endocrine conditions logged 144 expert panels in four years across 111 reference centers, and had to pay EUR 200 per completed panel to stimulate participation, with clinicians citing login difficulty, lack of time, and lack of awareness as barriers. The NIH Undiagnosed Diseases Network accepts roughly 30 to 42 percent of applicants and takes six to eight weeks to reach an acceptance decision, after months of record assembly.
Ninety-one percent of pediatricians report low confidence diagnosing rare disease. They are not wrong to lack confidence. They are being asked to do pattern recognition on a pattern they have never seen, when a colleague three states away has seen it eleven times.
The rare disease bottleneck is not primarily knowledge scarcity. It is expertise routing latency.
Exhibit C: your directory is describing a person who no longer exists
Here is the part of the problem almost nobody talks about, and it is arguably the most damaging, because it corrupts even the information we do have.
Expertise is not static. It grows, drifts, narrows, and decays. Directories update on discrete events: a new license, a new board certification, a new employer. The reality they are supposed to describe changes continuously.
The evidence on decay is uncomfortable and worth stating plainly. A systematic review by Choudhry and colleagues in Annals of Internal Medicine found that 32 of 62 evaluations, or 52 percent, showed decreasing performance with increasing years in practice across measured outcomes. Research by Tsugawa and colleagues found 30-day patient mortality of 10.8 percent for hospitalists under 40 versus 12.1 percent for those over 60, with an important exception: the difference disappeared among high-volume physicians. Volume, in other words, is protective. Recency and repetition are what actually matter, and neither appears in any directory.
Can clinicians self-report this accurately? A systematic review by Davis and colleagues in JAMA found that in 13 of 20 comparisons, physician self-assessment bore little, no, or an inverse relationship to external measures of competence, and that the mismatch was worst among the least skilled and the most confident. Self-report fails precisely where you need it most.
Meanwhile the directories are not merely conceptually stale, they are factually broken. A CMS review found that 48.74 percent of Medicare Advantage directory locations had at least one inaccuracy, and that 40 percent of flagged errors were still wrong 500 days later. In behavioral health it is worse: research in Health Affairs analyzing Oregon Medicaid claims found 58.2 percent of directory listings were "phantom" providers who saw zero enrollees, rising to 67.4 percent among prescribers. A US Senate Finance Committee secret-shopper investigation obtained an appointment only 18 percent of the time.
Now combine the three findings and look at what a referring physician is actually working with. A list that is roughly half wrong on the basics, describing practice focus that may be decades out of date, self-reported by people whose self-assessment is demonstrably unreliable.
Then ask why 72 percent of referrers just refer to the same person they always refer to. They are not being lazy. They are routing around a broken instrument using the only reliable one they have, which is memory.
The metric that would change everything, and nobody publishes it
Medicine is superb at measuring intervals. Door-to-needle. Door-to-balloon. Time to antibiotics. Referral-to-appointment. When the profession decides an interval matters, it measures it obsessively, benchmarks it publicly, and drives it down relentlessly.
There is one interval that governs the quality of every non-protocolized decision in medicine, and it has never been measured.
Expertise routing latency: the time from a clinician recognizing they need someone who knows X to receiving a substantive answer from a qualified human.
Look at the range that interval currently spans:
- Minutes, via a curbside to a friend, with quality unmeasured and roughly a coin flip on completeness.
- One to two hours, via an informal app-based consult service in Japan, where 52 percent of questions were resolved in chat.
- 1.2 to 2.6 days, via a funded e-consult program, if your employer bought one and your question fits its panel.
- 26 days on average for a new-patient specialist appointment in large US metros, and 34.5 days in dermatology.
- Six to eight weeks for an undiagnosed-disease program decision, after months of record assembly.
- Four years, if you are the patient in a diagnostic odyssey.
The same clinical question can take two hours or four years depending entirely on which channel it happens to fall into, and no clinician has any tool that shows them the fastest qualified path for the question in front of them.
Because it is not measured, it is not managed. Because it is not managed, it is not budgeted. Because it is not budgeted, the specialist hours that would collapse it are spent instead on 26-day appointments that deliver 18-minute answers.
Access in medicine is treated as a workforce problem. A large part of it is a routing problem. The hours often exist. They are allocated to the wrong channel.
Why the incumbents have not fixed this
The reasonable objection at this point is: there are enormous, well-funded companies serving physicians. Why has none of them closed the gap?
The answer is not that they are incompetent. It is that each is structurally prevented, and the structure is more interesting than the failure.
Doximity is the verification and distribution rail of American medicine: more than three million members, roughly 85 percent of US physicians, $644.9 million in fiscal 2026 revenue, clinical AI tools deployed in more than 140 health systems. It could add a "find a colleague" feature tomorrow. What it cannot add is an obligation to answer, because obligation is a membership covenant rather than a product feature. And its revenue comes from pharmaceutical marketing, hiring, and enterprise subscriptions, which means it must maximize reach and measurable impressions. A network whose economics require selling access to physicians cannot credibly promise physicians that nobody is buying access to them.
Sermo claims more than 1.5 million healthcare professionals and solves the disclosure problem through anonymity. But anonymity destroys the thing that makes an answer worth having. You cannot route a hard question to the right person if you do not know who anyone is, and you cannot rely on an answer whose author has no standing to lose.
LinkedIn has the world's largest professional graph and no clinical verification, no specialty semantics, no confidentiality, and an engagement model that rewards visibility over accuracy. It can tell you someone's job title. It cannot tell you who they call at 2 a.m.
ResearchGate, ORCID, and PubMed capture what people wrote, not what people did. Clinical, procedural, and operational expertise is overwhelmingly unpublished.
OpenEvidence deserves particular attention, because its trajectory is the single most important fact in this landscape. Founded in 2022, it reports more than 757,000 registered verified US physicians and daily use by over 40 percent of American physicians across more than 10,000 hospitals, reaching roughly a million queries in a single day in March 2026. It is the fastest-adopted clinical tool in modern American medicine.
And it answers questions from literature. It cannot tell you who to call. It cannot be accountable for the residual. It cannot help with any problem whose answer is a person rather than a citation.
Specialty societies and the AMA historically performed the trust-routing function. That capacity is thinning. AMA membership has fallen from roughly 75 percent of US physicians in the 1950s to about 15 percent today, with roughly 47 percent of current members being students and residents rather than practicing physicians.
WhatsApp and group chats are what clinicians actually use, and they work by being small. That is exactly why they cannot solve rare-expertise discovery, which requires being large.
Recruiters, locum agencies, and expert networks monetize these coordination failures directly. The US locum tenens market runs about $9.6 billion. The expert network industry reached roughly $3 billion in 2025, with firms charging clients $1,000 to $1,400 an hour for expert calls while paying the expert $200 to $500. Their margin is the opacity. An intermediary paid for access has no incentive to make the underlying graph transparent.
Put it all together and the residual is stark. A clinician using every one of these products simultaneously still cannot find the person who has actually done the rare thing, still cannot get a hard question to a qualified peer with any expectation of a reply, and still cannot discover who has already solved the problem their institution is about to start on.
What the gap actually costs
Skeptics are right to demand numbers before accepting that a coordination failure is an economic one. Here are the ones that hold up.
Referral misrouting. Referral leakage is estimated to drain on the order of $150 billion a year from US healthcare, with leakage rates between 20 and 65 percent depending on service line and mid-sized systems losing $200 to $500 million each. MGMA reported in 2025 that 38 percent of referrals stall entirely, usually because nobody follows up.
The diagnostic odyssey. A 7.6-fold excess in pre-diagnosis direct costs, at roughly EUR 27,000 per patient, across populations of roughly 30 million people in the US and 30 million in Europe.
Uncompensated expert time. Roughly nine million hours a year of subspecialist judgment given away through an unrouted channel, worth something on the order of $1.5 to $2 billion at conservative hourly valuations.
Wasted specialist appointments. With e-consults resolving 32 to 70 percent of questions that would otherwise become face-to-face referrals, and each avoided referral worth roughly $150 to $500 in payer spend, the misallocation is measured in billions.
Professional isolation. Research in Annals of Internal Medicine attributed roughly $4.6 billion a year in physician turnover and reduced clinical hours to burnout, at about $7,600 per employed physician.
Duplicated institutional work. Only 18 percent of health systems report a mature AI governance structure. Essentially all of the rest are independently drafting the same policies and independently evaluating the same three ambient scribe vendors, at roughly 6,100 hospitals.
None of these are knowledge problems. Every one is a matching, trust, or routing problem.
What you can actually do about this
Diagnosis without treatment is just commentary. Here is what is actionable now, by role, without waiting for anyone to build anything.
If you are a practicing clinician
Write down your own exposure, with dates. Not your CV. A working list: what you have personally managed, roughly how many, how recently, in what setting. "Fourteen cases of X, most recent 2026, community hospital." You will be astonished how much of your genuine expertise appears nowhere in any document about you. This is the single highest-leverage professional inventory most clinicians have never done.
Audit your own routing latency for one month. Every time you think "I need someone who knows about this," note the timestamp, what you needed, what you did, and when you got a useful human answer. Most clinicians who try this are shocked by two things: how often the honest answer is "never," and how much of their routing runs through three or four people they happened to train with.
Treat your co-training cohort as infrastructure, not nostalgia. The evidence says this is the strongest durable trust edge you have. Most people let it decay because there is no maintenance mechanism. Build one deliberately: a real roster with current focus, not just a dormant group chat.
When you curbside, structure it. State explicitly whether you are asking a general educational question or seeking patient-specific advice. Give the answerer the information they need to be right, since half of curbside information is incomplete. Note what you were told. This protects both of you, and it is the difference between a channel and a liability.
If you run a department, group, or service line
Measure time-to-qualified-answer as an operational metric. Pick your ten hardest recurring question types. Measure honestly how long it takes someone on your team to reach a qualified human. Publish it internally. You cannot improve an interval nobody has ever timed.
Map your own expertise inventory before you buy anything. Most departments cannot answer "who here has managed X" without asking around. That is an internal version of the same failure, and it is fixable in an afternoon with a spreadsheet.
Stop treating referral quality as a directory problem. Your referrers are already routing on trust and ignoring your list. Ask them who they actually send to and why. The gap between their answer and your directory is your real referral strategy.
If you build products or set policy
Separate "trained in" from "currently does," permanently. Any expertise field without a date is a liability. Any expertise claim without peer corroboration is marketing.
Assume decay. Design attestations that expire. An expertise claim from 2011 that has never been renewed is data about 2011.
Stop optimizing directory accuracy for addresses. Phone numbers and locations are the least consequential inaccuracy. Practice-focus recency is the one that misroutes patients, and no attestation campaign fixes it. Only peers can.
Frequently asked questions
What is the credential-expertise gap? It is the distance between what a healthcare directory records about a clinician (license, board certification, employer, training) and what that clinician actually knows and does now (procedures performed recently, conditions personally managed, current practice focus). Directories record credentials because institutions are legally required to verify them. Nobody is paid to record expertise, so nobody does.
What is expertise routing? Expertise routing is moving a specific question, case, or need to the specific verified person best able to act on it, fast enough to matter. Healthcare has world-class verification infrastructure and almost no routing infrastructure. The measurable version is expertise routing latency: the time from recognizing you need someone who knows X to getting a substantive answer from a qualified human.
Why can't I just look up which doctor has treated a rare disease? Because exposure is not recorded anywhere. Board certification describes an examination. Employment records describe a location. Publication records describe writing, and the clinicians with the deepest hands-on exposure often publish least. Centers-of-excellence lists name institutions rather than the specific person with the pattern recognition. The information exists only in individual memory.
Are provider directories really that inaccurate? Yes, and it is well documented. A CMS review found 48.74 percent of Medicare Advantage directory locations had at least one inaccuracy, with 40 percent of flagged errors still wrong 500 days later. In behavioral health, Health Affairs research found 58.2 percent of Oregon Medicaid directory listings were providers who saw zero enrollees, and a Senate Finance Committee secret-shopper study obtained an appointment only 18 percent of the time.
Does clinical AI solve this? No, and it sharpens the need. Tools like OpenEvidence, used daily by more than 40 percent of US physicians, answer questions from literature extremely well. They cannot tell you who has personally managed your case, cannot take accountability for the decision, and cannot help when the answer you need is a person rather than a citation. As generically correct answers approach zero cost, the scarce asset becomes the accountable human who will take the call.
Isn't the curbside consult already solving this? Partly, and unreliably. It is enormous: 87.5 percent of subspecialists fielded one in the prior week. But it routes by acquaintance rather than expertise, and comparison against formal consultation found curbside information inaccurate or incomplete 51 percent of the time with management advice changing in 60 percent of cases. It is a channel, not infrastructure.
Why haven't Doximity or LinkedIn built this? Structural constraints, not missing features. Doximity's revenue requires selling access to physicians, which is incompatible with promising physicians that nobody is buying access to them, and it cannot create an obligation to answer because that is a membership covenant. LinkedIn has no clinical verification or specialty semantics. Sermo's anonymity destroys the accountability that makes an answer valuable. E-consult vendors are contractually scoped to an employer's or payer's panel, which is exactly the boundary hard questions cross.
The bottom line
There are more than a million active physicians in the United States and more than 70 million health workers on earth. Every one of them has been verified, examined, licensed, credentialed, and re-credentialed. It is the most thoroughly checked workforce in human history.
And at 2:15 in the morning, a good hospitalist with a hard case cannot find the one person who has seen it eleven times.
That is not a failure of knowledge, or of training, or of individual effort. It is a missing layer of professional infrastructure. Medicine built verification because institutions needed it, and never built routing because no institution owned the problem.
The gap is nameable: the credential-expertise gap. It is measurable: expertise routing latency. It is expensive: billions in misrouted referrals, wasted appointments, prolonged diagnostic odysseys, and uncompensated expert time. And it is, unlike most problems in healthcare, structurally tractable, because the expertise already exists. Nobody needs to be trained. Nobody needs to be hired. The people with the answers are already out there, licensed, verified, and reachable in principle.
They just cannot be found.
That is the work.
This article is the first in a series examining the missing professional infrastructure of healthcare. Subsequent pieces go deep on the curbside consult as an unrouted market, the measurement of expertise routing latency, the co-training trust graph medicine already has and never uses, and what accountable human expertise is worth in an era of free machine answers.
Evidence note: figures cited here come from AAMC and FSMB workforce data, JAMA, Annals of Internal Medicine, Journal of Hospital Medicine, Health Affairs, Health Services Research, PLoS One, CMS reviews, US Senate Finance Committee investigation materials, company financial disclosures, and industry reports. Where a figure originates from a vendor or trade source rather than peer-reviewed or government data, it is identified as such in the text. Several widely circulated healthcare statistics were deliberately excluded because they could not be verified to a primary source.
