A health system with 3,000 employed physicians runs a leakage reduction initiative. Dashboards, incentives, a physician liaison team, messaging about keeping care in the family.
Meanwhile a primary care physician in that system sends a patient with a complex hand problem to a surgeon at a competing hospital, because she trained with him and knows he is excellent.
Two floors away from her clinic, in her own system, is a hand surgeon with a subspecialty interest in exactly that problem.
She has never met him. She does not know he exists. His name appears in the EHR provider directory as "Orthopedic Surgery," which is also true of eleven other people.
The system will record this as leakage and treat it as a loyalty problem.
It is not a loyalty problem. She routed along the only trust graph she has, because her employer never gave her an expertise graph.
The numbers on internal blindness
Survey data from referral technology research puts the scale of this clearly. From a study of 200 physicians:
- 79 percent refer patients out of network.
- Roughly 34 percent of those out-of-network referrals are avoidable with better information about colleagues' specialties and clinical focus.
- 72 percent refer to the same provider by habit rather than searching for the best match.
- 45 percent find it hard to determine who is even in network.
That third figure is the one worth pausing on. Nearly half of physicians cannot reliably determine who is in their own network. Not who is good. Who is in it.
And the financial framing that health systems use makes the scale of the concern obvious. Industry benchmarks put referral leakage at a substantial share of employed primary care referral revenue, with vendor estimates running into tens of millions of dollars per hundred affiliated physicians. Those figures come from companies selling referral management software and should be treated as directional, and the direction is not in dispute: systems consider this among their largest recoverable revenue problems.
Why consolidation made this worse
Health system consolidation has produced organizations of a size at which the informal mechanism cannot possibly work.
In a 40-physician group, everybody knows everybody. Expertise is discovered by lunch. If you need to know who is good at something, you ask three people and get a consistent answer.
In a 3,000-physician system spanning multiple hospitals, dozens of clinics, and several acquired practices across a region, the informal mechanism fails completely. Most physicians have never met most of their colleagues and never will.
And the acquisition pattern compounds it. A system that grows by acquiring practices inherits physicians whose entire professional network was formed elsewhere. Their trust graph points outward, at the people they trained with and the specialists they used for a decade before the acquisition.
Consolidation grew the employment graph far faster than it grew the trust graph, and nothing was built to close the gap.
The directory cannot fix this, for the reasons this series keeps finding
The natural institutional response is a better internal directory, and most systems have built one. They do not work, and the reasons are now familiar.
Directories index employment, not expertise. They record specialty, location, and network status. The hand surgeon with the specific subspecialty interest appears as "Orthopedic Surgery," because that is the taxonomy the directory has.
Profiles are self-maintained and go stale. This is the documented failure mode of every expertise-locator system ever deployed, in healthcare and outside it. Older enterprise research found large gaps between organizations wanting colleague-expertise search and being able to do it, with self-maintained profiles as the culprit. Academic research networking systems were built expressly to solve this and largely stagnated for the same reason.
Physicians have no incentive to maintain data for their employer. Filling in a detailed expertise profile is unpaid administrative work whose benefit accrues to the organization. Predictably, it does not happen, and what does get entered is written once and never revisited.
And there is no corroboration. A self-described expertise claim is unverified, which means a referring physician has no more reason to trust the directory entry than a stranger's assertion.
What physicians actually use instead
The referral literature is consistent and points at exactly one thing.
Physicians refer to people they trained with. Research examining more than 40,000 referrals found primary care physicians referring to residency and fellowship co-trainees at 26.2 percent against a 21.4 percent baseline, with the effect driven by post-graduate co-training rather than medical school.
So the actual algorithm inside a large health system is:
- Do I know someone who does this? If yes, refer there.
- If not, refer to whoever I referred to last time (the 72 percent).
- If neither, search the directory and hope.
Note that steps one and two are unaffected by network status, which is precisely why leakage initiatives underperform. A physician following step one is not defying the system. They are using the only reliable information they have.
The uncomfortable implication for health systems is that the trust graph their physicians use was formed years earlier, somewhere else, and points at competitors. Employing someone does not transfer it.
The idle capability problem
There is a second cost, and it receives almost no attention because it does not appear on a dashboard.
When expertise inside a system is unknown, it is unused.
The hand surgeon two floors away with the specific subspecialty interest is doing general orthopedic work he could do less of, while cases suited to his particular expertise travel elsewhere. The internist who has quietly become the regional expert in a complex condition sees a general panel because nobody knows to send those patients to her. The intensivist with unusual experience in an uncommon presentation is never called for it.
This is a pure deadweight loss. The system employs the capability, pays for it, and does not route work to it.
And it damages the physicians as well. Building genuine subspecialty expertise and finding it unused is a documented contributor to professional dissatisfaction. The physician who wants to do the thing they are best at, and cannot because nobody knows they do it, is a retention risk in a market where retention costs are enormous.
The onboarding hour nobody uses
There is a moment in every physician's employment when this problem could be solved almost for free, and it is universally wasted.
New physician onboarding at a large health system covers billing, the electronic record, compliance modules, badge access, parking, and human resources paperwork. It is thorough, it is mandatory, and it is entirely about how to operate the organization's systems.
It contains almost nothing about who the other physicians are.
Consider what an hour spent differently would produce. A new internist told: here are the six specialists in this system whose subspecialty focus most commonly matters for your panel, here is what each of them actually wants to see, here are three colleagues who trained where you trained, and here is who to call at 2 a.m. for each of the four situations that will come up in your first year.
That hour would shape a decade of referral behavior. It costs one hour. It is not on any onboarding checklist I have seen.
Instead the new physician spends her first two years discovering all of this by accident, forming referral habits from whichever colleague she happened to meet first, and, per the survey evidence, keeping those habits for the rest of her tenure because 72 percent of physicians refer by habit.
The organization has exactly one moment when a physician's referral defaults are still unset, and it spends that moment on the expense reimbursement system.
What would actually work
Corroborated expertise, not self-reported profiles. The distinction between "I say I do this" and "four colleagues confirm they have sent me these cases and the outcomes were good" is the entire difference between a directory that fails and one that works. Corroboration is also cheap: it takes seconds and it is exactly the judgment colleagues make informally anyway.
Sub-specialty granularity. "Orthopedic surgery" is not a routing key. "Complex hand and wrist, including revision" is.
Make the co-training graph visible inside the system. This is the highest-value and lowest-cost intervention available. Systems know where their physicians trained; the data is in their credentialing files. Surfacing "three of your colleagues trained at the same program as this specialist" activates the trust mechanism physicians already use, but pointed inward.
Portable, member-owned expertise records. Here is why an employer-owned directory can never fully solve this: physicians have no reason to maintain data for an employer they may leave, and everything they enter is lost when they do. A record the physician owns, maintains for their own benefit, and shares with employers by consent, is the only version that stays current across a career.
And an honest reframe of the leakage problem. A system that treats out-of-network referral as disloyalty will keep losing. A system that treats it as evidence that its physicians lack an expertise map, and then builds one, addresses the actual cause. The Kyruus finding that roughly a third of out-of-network referrals are avoidable with better information is the whole argument.
What you can do now
If you are a physician in a large system
Find out who your colleagues actually are. Ask your department, once, for a list of who does what at sub-specialty granularity. Most departments have never produced one and the exercise is genuinely revealing.
Tell people what you do. Not your specialty. What you actually focus on and want to see. A single email to referring colleagues is more effective than any directory entry you will ever complete.
Ask where your colleagues trained. It is a normal collegial question and it maps the trust graph inside your own building, which nobody else has done.
Audit your own habit. With 72 percent of physicians referring habitually, your defaults were probably set years ago on old information. One review a year is enough.
If you lead a system
Stop calling it leakage and start calling it a search failure. Roughly a third of out-of-network referrals are described as avoidable with better information. That is a discoverability problem your organization can solve, unlike physician loyalty, which it cannot mandate.
Surface where your physicians trained. You already hold this data in credentialing. Making it visible activates the strongest documented referral trust signal in the literature, at essentially zero cost, and no system I am aware of does it.
Fund the expertise map properly, with corroboration. Self-maintained profiles have failed everywhere they have been tried. Peer attestation is the only mechanism with any track record.
Ask your specialists what they are not seeing. The specialist with unused subspecialty capability knows exactly what they wish they were getting. Nobody asks, and the answer is a free map of your own idle capacity.
If you build referral technology
Taxonomy is the product. Specialty-level matching cannot solve a sub-specialty problem, and every additional feature built on top of a coarse taxonomy inherits its failure.
Design for the physician, not the administrator. A tool that exists to reduce leakage will be experienced as surveillance. A tool that genuinely helps a physician find the right colleague will reduce leakage as a side effect and will actually be used.
Frequently asked questions
How often do physicians refer outside their own network? Frequently. Survey research covering 200 physicians found 79 percent refer patients out of network, with roughly 34 percent of those out-of-network referrals described as avoidable with better information about colleagues' specialties and clinical focus.
Why do physicians refer out of network when an in-network option exists? Primarily because they know the outside specialist and do not know the inside one. Referral research finds 72 percent refer to the same provider by habit and 45 percent find it hard to determine who is in network, while physicians route preferentially to people they trained with regardless of employment.
Do internal provider directories solve this? Generally not. Directories index employment and specialty rather than sub-specialty expertise, rely on self-maintained profiles that go stale, offer physicians no incentive to keep them current, and provide no peer corroboration of stated expertise. Expertise-locator systems have failed for these reasons across industries.
What is the strongest predictor of where a physician refers? Prior relationship, particularly co-training. Research examining more than 40,000 referrals found physicians referring to residency and fellowship co-trainees at 26.2 percent versus a 21.4 percent baseline, with the effect driven by post-graduate training rather than shared medical school.
Why does consolidation make internal expertise harder to find? Because informal discovery works in small groups and fails at scale. A physician in a 3,000-clinician system will never meet most colleagues, and acquired physicians arrive with professional networks formed entirely elsewhere, pointing at specialists outside the new organization.
What is the cost of unknown internal expertise? Two costs. Referral revenue leaving the system, which is what organizations measure, and idle clinical capability, which nobody measures: specialists with genuine sub-specialty expertise doing general work because no one knows to route the relevant cases to them, which is also a documented contributor to professional dissatisfaction.
The bottom line
A health system employs three thousand physicians and has no idea what most of them know.
Its directory records their specialty and their location, which is what the compliance function requires and what the referral decision cannot use. Its physicians, lacking any better instrument, route to the people they trained with, three quarters of them by pure habit, and nearly half cannot even establish who is in their own network.
The organization records the result as leakage and responds with dashboards and messaging about loyalty.
Meanwhile the hand surgeon two floors away with exactly the right subspecialty interest continues doing general cases, unaware that a colleague in the same building has been sending his ideal patients to a competitor for six years, for the entirely rational reason that she knows the competitor and has never heard of him.
The expertise is employed. It is paid for. It is two floors away.
Nobody built the map.
Part of a series on the missing professional infrastructure of healthcare. Previously: The Doctor in the Family
Evidence note: referral behavior figures come from Kyruus referral trends research surveying 200 physicians, a vendor source, and are identified as such. Referral leakage financial benchmarks come from industry and vendor compilations and are directional rather than measured. Co-training referral effects come from Pany and McWilliams in Health Services Research (2021). Directory inaccuracy figures come from CMS provider directory review. Evidence on expertise-locator system failure includes dated enterprise research and the documented stagnation of academic research networking platforms.
