A family medicine physician is looking at a 61-year-old patient with a new pancreatic mass on a CT she ordered three days ago.
She is employed by a 400-bed health system that acquired her practice two years ago. The system has a gastroenterology group, four physicians, down the hall from her clinic. It also has a GI oncologist across town, in independent practice, whom this physician has referred to for eleven years, since they worked together at her first job out of residency. That physician calls her back the same day, personally, and has walked three of her patients through a Whipple procedure with outcomes she has watched closely enough to trust completely.
She sends the referral across town.
Two weeks later, her practice manager forwards an email from the system's network performance team. Her out-of-network referral rate is in the 85th percentile for her division. Would she be willing to review the in-network specialist roster during her next huddle.
She has never met the employed GI oncologist down the hall. She does not know his complication rate, whether he returns calls, or what he does when a case goes sideways. The dashboard does not ask her that. It asks her to change her behavior, and the behavior it is measuring is trust, priced as a percentage.
The health system is trying to fix a network problem with a network it does not own: the one inside her head.
The leak is enormous, and it is not new
Start with the scale, because it is larger than most administrators are prepared to say out loud.
66.7 percent of Medicare beneficiaries' office visits with specialists occurred outside their assigned ACO, according to a JAMA Internal Medicine analysis of Medicare claims. Primary care leakage was far lower, at 8.7 percent, which tells you the mismatch is concentrated exactly where referral choice matters most: the specialist visit, the one the PCP actively selects rather than the one the patient books directly.
The leakage rate was not flat. It ranged from 54.3 to 68.1 percent depending on spending quartile, rising with patient complexity. The sicker the patient, the more likely the referral left the network.
And it barely moved. A follow-up analysis in the American Journal of Managed Care found that specialist leakage "decreased minimally" from 2010 to 2014, even as ACO contract penetration climbed to 85 percent in specialties where it mattered most. Health systems have spent a decade building the contractual apparatus of narrow networks. The referral pattern underneath it has hardly noticed.
This is not a story about physicians who do not understand their contracts. It is a story about two graphs occupying the same organizational chart and answering to different owners.
Two graphs, two owners, opposed incentives
Every referral a physician makes sits at the intersection of two entirely separate maps.
The contract graph is built by the health system, the payer, or the ACO. It says which specialists are in network, which have signed risk-sharing agreements, which the finance department wants to keep the downstream imaging, surgery, and infusion revenue flowing to. It is maintained by contracting teams, physician-liaison staff, and increasingly by referral-management software. It changes when a contract changes.
The trust graph is built by the referring physician over years. It says who answers the phone, who reports back with a legible plan rather than a scanned discharge summary, who treats her patients the way she would want a family member treated, who she trained with, who has never let a complicated case go sideways without calling her first. It is maintained entirely in one person's memory and it changes when a relationship changes, not when a contract does.
A systematic review in the Journal of General Internal Medicine, pooling seven studies and 1,575 referring providers, found that referrers "rely on subjective factors and assessments to evaluate quality" when choosing where to send a patient. Not the metrics on a scorecard. Not network status. Personal, accumulated, largely undocumented judgment.
These two graphs are drawn by different people, for different purposes, updated on different timelines, and they are currently reconciled by exactly one mechanism: pressure on the physician to conform the trust graph to the contract graph, applied through dashboards, compensation formulas, and huddles about "in-network referral requirements."
Nobody is building the reverse mechanism: fixing the contract graph to reflect where trust already lives.
Leakage is priced as loss, and the loss is real
The financial argument for stopping leakage is genuine, even where the specific numbers are shaky.
Health-system consultancies routinely cite figures in the range of 10 to 30 percent of system revenue lost to leakage, and a widely circulated estimate puts the per-physician cost at $821,000 to $971,000 a year. Those numbers come from vendors selling leakage-reduction software and consulting engagements, and they have not been independently verified in the sources reviewed for this article. Treat them as directional, not audited.
What is independently documented is the scale of the underlying behavior. If even a modest fraction of the leakage measured in the JAMA Internal Medicine and AJMC studies represents genuinely recoverable revenue, the dollar figure is large for any individual system: a 400-bed hospital with a specialist referral base generating eight figures in downstream revenue does not need to recapture much of a 66.7 percent leakage rate before the number matters to a CFO.
So the incentive to act is real. The question is what "acting" should mean, and the industry has answered it almost entirely with software that maps the contract graph more precisely and pushes it harder at the referring physician, rather than software that asks why the trust graph diverged from it in the first place.
What the evidence says trust actually buys
Here is the finding that should reframe the entire leakage conversation, and it rarely appears in a leakage-reduction sales deck.
A 2021 study in Management Science, drawing on 284,604 Medicare Advantage patients and 3.4 million Medicare beneficiaries, found that a one standard deviation increase in referral concentration (meaning a PCP sending a larger share of referrals to a smaller, more consistent set of specialists) produced 3.9 percent lower utilization and 7.4 percent lower spending, with no decline in quality.
Read that carefully. Concentrated, relationship-based referral, the kind built by exactly the informal trust mechanism a leakage dashboard treats as noise, is associated with lower spending and unchanged quality. It is not the enemy of value-based care. It may be one of its most under-used levers.
A 2021 study in Cancer Medicine adds a specialty-level view of why physicians deviate from pure clinical judgment when they do stray from trust. Among community oncologists, 48 percent said they weigh insurance considerations in referral decisions, versus 15 percent of tertiary-center oncologists (a statistically significant difference, p equals 0.003), and 59 percent cited financial considerations as their biggest referral barrier. This is the mirror image of the leakage story: it shows referrers who are not defecting from network incentives but constrained by them, sometimes away from the specialist they would otherwise choose.
Put the two findings together and a coherent picture emerges. Referral concentration built on trust looks efficient. Referral fragmented by insurance friction looks costly. The health system's leakage dashboard cannot tell the two apart, because it only sees where the referral landed, never why.
Why nobody owns the reconciliation
This is the structural diagnosis, and it explains why the mismatch has persisted for over a decade of ACO contracting without anyone seriously trying to fix it.
The health system wants to capture referrals, not to understand or honor the trust that generated them. Its referral-management tools (Definitive Healthcare, Trilliant Network Explorer, Doctivity, Kyruus, Epic's own "preferred provider" surfacing) all answer one question: where should this referral go to keep revenue and risk inside our contracts. None of them ask why the physician trusts who she trusts, because that answer does not change what the system is selling.
The payer wants a narrow, controllable network. Its interest in trust ends at the point where trust threatens the network's boundaries.
Doximity, the platform that already holds most of the raw material for a trust graph (residency history, specialty, activity), monetizes physician attention through advertising and pharma relationships. Asking members to rank and vouch for peers, the way a trust graph requires, would create exactly the kind of adversarial content that threatens those relationships. It has not built this, and there is no commercial reason for it to start.
Physician-liaison teams, the closest thing to a human reconciliation layer, spend their time "pounding the pavement" trying to build new referral relationships toward the contract graph. Nobody is tasked with the opposite job: documenting the existing trust graph and asking what it would take to make the in-network option earn a place on it.
The result is a referrer with no fiduciary-aligned party. Every institution touching her referral decision wants something from the outcome. Nobody is simply trying to help her make and record the best-informed choice and be honored for it.
The compliance boundary that keeps everyone honest, and stuck
Some of the paralysis here is not laziness. It is genuine legal caution, and it deserves to be named rather than waved past.
Employed physicians whose compensation is structured around productivity face real Stark Law and Anti-Kickback Statute exposure if referral requirements or incentives are drawn too tightly around financial relationships rather than clinical judgment. A system cannot simply pay physicians a bonus for sending referrals to a specific specialist without walking into exactly the kind of self-referral and kickback scrutiny those statutes exist to prevent. Patient-choice obligations under Medicare add another layer: a beneficiary generally retains the right to see any specialist, regardless of what a system's contracting team prefers.
This is precisely why leakage-reduction efforts have converged on the softest available lever, dashboards, huddles, and preferred-provider surfacing in the EHR, rather than harder financial or administrative levers. It is also why a trust-based alternative has to be built with the same legal discipline: any mechanism that documents and routes referral trust needs firm guardrails against becoming a disguised steering or pay-for-placement arrangement. The absence of those guardrails, not the absence of interest, is a large part of why nobody has built this cleanly.
What would actually work
Document the trust graph before trying to change it. Before a system spends another dollar on leakage-reduction software, it should ask its referring physicians, condition by condition, who they actually trust and why. Most systems have never collected this in a structured way, which means every leakage intervention is aimed at a target nobody has actually mapped.
Separate reachability, report-back, and outcomes from network status entirely. A referrer's trust is built on concrete, measurable behaviors: does the specialist answer, does a legible plan come back within a defined window, does the patient report being treated well. These can be tracked and improved without ever touching contract status, and doing so is the fastest way to make an in-network specialist trust-competitive.
Treat leakage as a signal to investigate, not a defect to suppress. A high leakage rate for a specific condition or specialty is diagnostic information about where the in-network option is failing trust, whether through slow reporting, poor reachability, or a genuine quality gap. Suppressing the signal with a dashboard mandate destroys the information before anyone reads it.
Give the trust graph a portable, physician-owned home. The value in a referring physician's trust relationships should not evaporate when she changes employers, and today it entirely does, because it lives only in her memory and her personal contacts. A record she owns and carries is also the only version of this graph that any party can trust was not built to serve a contracting agenda.
Build reciprocity into the in-network relationship, not just access to it. The Management Science finding on concentrated referral suggests the goal is not maximizing in-network volume but maximizing the number of specialists worth concentrating referrals toward. That requires specialists who actively earn trust: answering promptly, reporting back, treating patients well, not merely appearing on a roster.
Keep any trust-verification mechanism free of pay-for-placement. The moment a specialist can buy a better trust signal, the entire mechanism collapses back into the contract graph it was meant to correct, and it invites exactly the Anti-Kickback and Stark scrutiny that has kept systems cautious for good reason.
Measure the concentration effect, not just the leakage rate. A leakage percentage tells a CFO how much revenue left. It does not tell anyone whether the referrals that stayed in network were concentrated among trusted specialists in the way the Management Science data suggests drives savings, or scattered by default among whoever appeared first on a preferred-provider list.
What you can do now
If you are a referring physician
Write down why you refer to whom you refer, condition by condition. Most physicians carry this reasoning entirely in memory. Externalizing it, even in a private note, is the first step toward being able to demonstrate it is clinical judgment rather than habit, and it is the raw material for any future system that tries to honor rather than override your trust graph.
Ask your system what "in-network" actually offers on reachability and report-back. If a preferred specialist cannot commit to answering within a defined window or sending a legible plan back, that is a legitimate, documentable reason for the gap your dashboard is flagging, and it is worth saying so explicitly rather than absorbing the pressure silently.
Push back on volume-only leakage metrics with your own data. If you can show that your out-of-network referrals cluster around complex cases, faster report-back, or better outcomes you have personally observed, you are giving your leadership the diagnostic information the dashboard cannot generate on its own.
If you lead a health system or ACO
Before building another leakage dashboard, survey your own referring physicians on trust. Ask them, by specialty and condition, who they would refer to if network status were not a factor, and why. This single survey will tell you more about where your in-network specialists are failing than a year of leakage percentages.
Fund reachability and report-back as a quality investment, not a compliance one. The JGIM finding that referrers rely on subjective quality signals means the fastest way to reduce genuine leakage is to make your in-network specialists demonstrably better on the dimensions referrers actually weigh.
Stop treating leakage as uniformly bad. The Management Science data suggests the goal is referral concentration around trusted specialists, not maximum in-network capture. A system that recaptures leaked referrals by winning trust will keep them. A system that recaptures them by mandate will see the underlying trust deficit resurface the moment enforcement relaxes.
If you build referral or network software
Build for the reconciliation, not just the contract graph. Every major referral-management vendor answers "where did referrals go." None answer "why did the referrer trust who she trusted, and what would it take to shift that trust in network." That gap is the product opportunity, and it requires physician-owned data, not just claims data.
Design compliance in from the start. Any trust-signal mechanism needs to be built with Stark, Anti-Kickback, and anti-steering constraints as first-class requirements, not an afterthought bolted on before a legal review. A trust graph that cannot survive that review is not a viable product, however useful the underlying data would be.
Frequently asked questions
What is referral leakage in healthcare? Referral leakage describes patients or referrals moving outside a health system's or ACO's contracted network of specialists. A JAMA Internal Medicine analysis found 66.7 percent of Medicare beneficiaries' specialist office visits occurred outside their assigned ACO, compared with just 8.7 percent of primary care visits.
Why do doctors refer patients out of network? Primarily because of trust built over years of working with a specific specialist: reachability, timely and legible report-back, and observed outcomes. A 2022 systematic review in the Journal of General Internal Medicine, covering seven studies and 1,575 providers, found referrers "rely on subjective factors and assessments to evaluate quality" rather than network status alone.
Does referral concentration actually lower healthcare costs? Yes, according to the best available evidence. A 2021 study in Management Science, using data on 284,604 Medicare Advantage patients and 3.4 million Medicare beneficiaries, found a one standard deviation increase in referral concentration produced 3.9 percent lower utilization and 7.4 percent lower spending with no decline in quality.
Is it legal for a hospital to require employed physicians to refer in network? It is legally constrained rather than simply prohibited or permitted. Stark Law and the Anti-Kickback Statute limit how tightly compensation or referral requirements can be tied to financial relationships, and Medicare patient-choice protections generally preserve a beneficiary's right to see the specialist of their choice regardless of a system's contracting preferences. This is a large part of why systems rely on softer levers like dashboards and preferred-provider lists rather than hard mandates.
How much money do hospitals actually lose to leakage? The widely cited figures, 10 to 30 percent of system revenue and $821,000 to $971,000 per physician per year, come from vendor and consulting estimates that have not been independently audited. What is independently documented is the leakage rate itself (66.7 percent of specialist visits, JAMA Internal Medicine 2014) and that it has moved only minimally since, per a 2018 AJMC follow-up analysis.
Do community oncologists refer differently than academic oncologists because of insurance? Yes. A 2021 study in Cancer Medicine found 48 percent of community oncologists weigh insurance considerations in referral decisions, versus 15 percent of tertiary-center oncologists (p equals 0.003), and 59 percent of community oncologists cited financial considerations as their single biggest referral barrier.
The bottom line
Two-thirds of Medicare specialist referrals leave the network a health system built to hold them, and that number has barely moved in a decade of dashboards, huddles, and preferred-provider lists. The industry's answer has been to build the contract graph more precisely and push it harder at the referring physician.
The evidence says this is aimed at the wrong target. Concentrated, trust-based referral is associated with lower spending and no loss of quality. The physicians generating "leakage" are, in a meaningful share of cases, doing exactly what the value-based care literature says should be rewarded: sending patients to specialists who answer, report back, and deliver outcomes they have personally verified.
Nobody owns the reconciliation between the network a payer contracts and the network a physician trusts, because every party that could build it has an incentive pointed somewhere else. The system wants capture. The payer wants narrow networks. Doximity wants advertising relationships undisturbed. The referring physician, who holds the only data that actually matters, has no fiduciary-aligned party and no way to make her trust graph legible to the people trying to override it.
That family medicine physician is still going to send her pancreatic mass patient across town, because eleven years of reachability and honest report-back beats a roster entry every time. Her system will keep calling this leakage, and keep buying software to describe it more precisely, without ever measuring the one thing that would actually change her mind: whether the specialist down the hall has earned it.
Part of a series on the missing professional infrastructure of healthcare. Previously: Imported Disease Expertise Deserts
Evidence note: leakage rates are from McWilliams et al., JAMA Internal Medicine (2014), analyzing Medicare claims data, and from a 2018 AJMC follow-up analysis. The referral-concentration and spending finding is from a 2021 Management Science study covering 284,604 Medicare Advantage patients and 3.4 million Medicare beneficiaries. The community-versus-tertiary oncologist referral behavior finding is from Cancer Medicine (2021). The subjective-quality-factors finding is from a systematic review in the Journal of General Internal Medicine (2022), pooling seven studies and 1,575 providers. The 10 to 30 percent revenue loss and $821,000 to $971,000 per-physician figures are vendor and industry estimates reported by Healthcare Finance News and have not been independently audited; they are presented here as directional claims, not verified findings. The opening scenario is a composite illustration built from the patterns documented in this evidence, not a specific reported case.
