An 81-year-old patient’s journey through cancer care raises a provocative question: Why is healthcare investing in AI to manage administrative failures instead of preventing them upstream?
“The next challenge for clinical AI is not only better task performance, but better operational follow-through across the care journey.” -Saif Khairat and Charles Safran, in a peer-reviewed perspective published in npj Health Systems on August 5, 2026.
The referral that never arrived
Healthcare can now generate clinical notes with artificial
intelligence, analyze medical images, automate billing tasks, and process
mountains of data in seconds.
Yet one of healthcare’s most important messages can still
fail to reach its destination: a referral.
Consider the story of Judy Brauning, whose cancer journey
was published by VCU Massey Comprehensive Cancer Center in Richmond, Virginia,
on October 15, 2026.
Brauning experienced persistent abdominal pain and made
repeated visits to an emergency department. Initially, her symptoms were
attributed to constipation. Further evaluation, including a colonoscopy and
other testing, did not reveal the cause. Eventually, she was referred to a
gynecologist, but she faced a three-month wait for an appointment.
A biopsy subsequently revealed endometrial cancer.
Brauning chose VCU Massey Comprehensive Cancer Center for
specialized treatment. Then, according to the center’s published account, she
discovered that her primary care doctor had not submitted the referral.
She did what many patients have learned to do when
healthcare’s administrative machinery stalls: she called the cancer center
herself.
She secured an appointment with a gynecologic oncologist two
days later.
Her treatment journey included extensive surgery,
chemotherapy, an infection that led to sepsis, and subsequent treatment. Later
scans showed no evidence of cancer, and she transitioned to maintenance
immunotherapy.
The published account was written by Lynne Blasi. It is a
story of resilience, specialized cancer care, and a patient determined to move
forward.
But it also raises an uncomfortable question.
Why should a patient have to become her own referral
coordinator to keep the healthcare system moving?
To be precise, the published story does not establish that
the missing referral caused the cancer to progress or changed Brauning’s
clinical outcome. We should not confuse a documented administrative failure
with proof of clinical causation.
The failure is still worth examining.
A referral is not simply a document. It represents a
clinical decision, a transfer of information, an expectation of follow-up, and
a handoff of responsibility.
When that handoff fails, the patient may be the only person
who realizes that the next step has not happened.
And that is where a human story becomes a systems question.
Healthcare’s technology paradox: more tools, same old
problems
Here is the contrarian argument.
Healthcare does not always need another tool to process the
mess. Sometimes it needs a system that prevents the mess from being created.
That sounds obvious. Yet much of healthcare technology is
organized around what happens after something goes wrong.
A claim is rejected, so a system identifies the denial.
Documentation is incomplete, so software flags the missing
information.
An authorization is delayed, so staff receive an alert.
A referral remains unresolved, so a dashboard displays an
overdue task.
These capabilities can be useful. Visibility matters.
Automation can reduce repetitive work. AI can identify patterns that humans
might otherwise miss.
But there is a fundamental distinction between identifying a
failure and preventing one.
Imagine a clinic invests in an expensive referral-management
platform. The interface is attractive, the dashboard is impressive, and the
vendor demonstrates how easily staff can track outstanding referrals.
Six months later, a referral still goes missing.
The software may make the problem easier to find. It may
even provide a timestamp showing exactly when the process stalled.
But if the underlying workflow still permits a referral to
remain unsubmitted without detection, the fundamental problem persists.
Congratulations: the organization has digitized the
possibility of noticing a failure.
This is not an argument against technology. It is an
argument against confusing technological sophistication with operational
reliability.
Healthcare has a remarkable ability to turn a simple
administrative problem into a complex digital workflow.
A missing piece of information becomes a task. The task
becomes a work queue. The work queue generates an alert. The alert requires a
staff member to investigate. The investigation creates another record. The
record produces a report.
Eventually, someone asks whether the organization needs
another employee to manage the alerts.
The original question remains unanswered: Why was the
information missing in the first place?
The next breakthrough in healthcare AI may not be a
machine that fixes mistakes faster. It may be a system that helps prevent
predictable mistakes from occurring at all.
A timely warning from healthcare technology researchers
This argument is not simply a contrarian opinion about the
future of AI.
On October 14, 2026, the Peterson Health Technology
Institute (PHTI) released a report examining the barriers to realizing AI’s
potential in healthcare. The institute warned that existing payment incentives,
administrative complexity, and flawed workflows could limit the value of new
technology.
Caroline Pearson, PHTI’s executive director, put the concern
bluntly in an interview with Axios:
“If we don’t start doing something differently, we’re going
to layer technology on top of existing broken workflows, everything gets more
expensive and we don’t get the benefits we hope for.”
— Caroline Pearson, executive director, Peterson Health
Technology Institute, quoted by Axios, October 14, 2026.
The warning deserves attention because it challenges one of
the healthcare technology industry’s most comfortable assumptions: that adding
AI to an existing process necessarily makes that process better.
It does not.
AI can accelerate a flawed workflow. It can increase the
volume of activity. It can make administrative work cheaper for one
organization while shifting costs to another. It can even make a dysfunctional
process appear more productive because more tasks are completed per hour.
The more important question is whether the technology
improves outcomes, expands access, or lowers the total cost of care.
PHTI’s analysis of administrative AI makes a similar point:
deploying AI on top of flawed workflows, fragmented data, and misaligned
incentives can exacerbate underlying problems. Realizing the technology’s
potential requires redesigning the processes in which it operates.
For independent practices, the implications are practical.
Before buying another AI tool, ask whether it will reduce
the amount of work required to resolve a problem or merely accelerate the
existing resolution process.
Will it prevent missing information, or simply flag it?
Will it reduce recurring exceptions, or generate a more
sophisticated exception report?
Will it make the revenue cycle more predictable, or make the
organization more efficient at chasing unpredictable outcomes?
These are not minor implementation questions. They determine
whether technology produces real operational value.
The referral and the claim: different consequences,
similar information challenges
A missing referral and a denied insurance claim are not
equivalent failures.
One can interrupt a patient’s path to clinical care. The
other may delay reimbursement, require additional administrative work, or
affect a practice’s cash flow.
Clinical safety must remain the priority when the two are
discussed together.
Nevertheless, both processes depend on information moving
reliably from one step to another.
A referral requires a clinical decision, appropriate
documentation, correct routing, receipt, and follow-up.
A claim requires accurate patient and insurance information,
appropriate documentation, coding, applicable authorization, submission,
adjudication, and payment or resolution.
Both can become unnecessarily complicated when information
is missing, inconsistent, delayed, or disconnected from the action it should
trigger.
Consider several examples.
Eligibility: A patient’s insurance information is
incomplete or outdated. The problem may not become visible until a claim is
rejected.
Authorization: A required authorization is not
identified early enough. Staff must investigate, contact the payer, and
determine how the issue affects the scheduled service.
Documentation: A clinical record does not adequately
support the billed service. The billing team must request clarification or
correct the claim when appropriate.
Coding: Documentation and coding are misaligned. A
claim may require review, correction, or appeal.
Referral management: A referral is created but not
submitted, received, scheduled, or completed as intended.
In each example, the downstream problem may be the first
visible sign of an earlier information or coordination failure.
This leads to an important distinction:
The department that discovers a problem is not
necessarily the department that created it.
A billing team may own the denial resolution without having
caused the original error.
A practice administrator may be responsible for following up
on a referral without having been responsible for its submission.
A clinician may be asked to clarify documentation because an
upstream process failed to capture the necessary information consistently.
Blaming the person who inherits the problem is an easy way
to avoid investigating its origin.
The more productive approach is to trace the problem
backward.
Where did the required information originate? Who had access
to it? When should it have been available? What prevented it from reaching the
next step?
That is how a practice moves from repeatedly correcting
errors to reducing the conditions that create them.
The hidden price of administrative work
Administrative work is not inherently wasteful.
Healthcare needs documentation, clinical review,
authorization processes, compliance controls, accurate coding, financial
reconciliation, and communication between organizations.
The problem is the unnecessary work layered on top of those
legitimate responsibilities.
A staff member may spend time verifying insurance
information that should already be available.
A biller may repeatedly contact a payer about an issue that
could have been identified earlier.
A physician may respond to documentation queries that arise
from an inconsistent process.
A practice manager may maintain a spreadsheet because the
organization’s existing systems do not provide reliable visibility into
unresolved work.
These tasks consume time even when no additional patient
care is delivered.
They also create opportunity costs.
Every hour spent investigating a preventable administrative
failure is an hour unavailable for other work. Depending on the situation, that
could mean less time for patient communication, clinical documentation, staff
development, operational improvement, or revenue cycle analysis.
The cost is not always visible on a single invoice.
It may appear as overtime, delayed collections, increased
staffing needs, employee frustration, or a growing backlog of unresolved tasks.
For small and medium-sized practices, the burden can be
particularly difficult because the organization has fewer people available to
absorb exceptions.
A large health system may have dedicated teams for
eligibility, authorization, coding, denials, and analytics. An independent
clinic may depend on a handful of staff members to cover several functions.
When the same avoidable problem occurs repeatedly, it does
not merely create inconvenience. It competes with the practice’s limited
operating capacity.
This is why a useful assessment should ask more than how
many claims were denied or how many tasks remain open.
It should ask how much effort is being spent correcting
problems that could have been prevented earlier.
What the evidence tells us about prior authorization
The administrative burden of healthcare is well documented,
even though the causes and consequences differ across organizations and
processes.
The American Medical Association’s 2025 physician survey
reported that physicians and their staff handled an average of 40 prior
authorization requests per week, consuming an average of 13 hours of physician
and staff time weekly.
The survey also reported that:
- 95%
of surveyed physicians said prior authorization delays access to necessary
care.
- 79%
reported that patients abandon treatment because of prior authorization
challenges.
- 92%
reported that prior authorization negatively affects clinical outcomes.
- 94%
said prior authorization contributes to physician burnout.
- 26%
reported that prior authorization had led to a serious adverse event for a
patient in their care.
These figures are survey findings, not universal estimates
of the experience of every physician or patient. They reflect respondents’
reported experiences and should be interpreted in that context.
Nevertheless, the volume of reported work is difficult to
ignore.
Forty requests each week is not an occasional interruption.
Thirteen hours represents substantial staff capacity devoted to a single
administrative process.
And prior authorization is only one part of the broader
administrative workload.
Practices must also manage registration, eligibility,
referrals, documentation, coding, claims, remittance advice, denial resolution,
payer correspondence, and patient billing.
Some of this work is necessary. Some can be simplified. Some
may be avoidable if information is captured, structured, and transferred more
reliably.
The objective should not be to eliminate every
administrative requirement. It should be to distinguish necessary controls from
unnecessary rework.
A process that protects patients, supports appropriate care,
and ensures accurate reimbursement is not waste simply because it requires
effort.
A process that repeatedly corrects the same preventable
information gap deserves scrutiny.
Why closed-loop follow-up matters
The Agency for Healthcare Research and Quality has supported
research into closed-loop diagnostic processes designed to improve follow-up of
tests, referrals, and evolving symptoms.
The principle is straightforward: creating an order is not
the same as completing the intended action.
A test may be ordered but never completed.
A result may become available without the appropriate
clinician reviewing it.
A referral may be created without the receiving specialist
obtaining the information needed to proceed.
A message may be sent without confirmation that the
recipient can act on it.
AHRQ’s work on closed-loop diagnostics highlights the need
to design systems that support reliable follow-up rather than relying solely on
individual memory and vigilance.
The specific completion rates reported in research vary by
setting and process; they should not be generalized into a claim that a fixed
percentage of all referrals fail nationally.
The broader lesson is more important than any single number.
A digital record proves that something was recorded. It
does not necessarily prove that the intended outcome occurred.
A status marked “sent” is not necessarily proof of receipt.
A claim marked “submitted” is not proof of payment.
An authorization request marked “in progress” is not proof
that the planned service can proceed.
A referral marked “created” is not proof that the patient
has been connected with the intended specialist.
Reliable workflows need clear definitions of completion,
ownership of the next action, and a way to surface unresolved exceptions.
That is what makes a process closed-loop rather than merely
digital.
Three expert perspectives that should change how
practices think
Russell S. Phillips, MD: follow-up must be designed
Dr. Russell S. Phillips of Harvard Medical School has been
associated with AHRQ-supported work on closed-loop diagnostic processes.
The practical lesson is that follow-up should not depend
entirely on memory or on a patient noticing that something has gone wrong.
A practice can define the information required at each
handoff, assign responsibility for the next action, establish completion
criteria, and specify how unresolved cases are escalated.
The same discipline can improve administrative and revenue
cycle processes.
A claim should not be treated as resolved merely because it
left the practice’s system. A referral should not be considered complete merely
because someone entered an order.
The next step needs to be visible and verifiable.
Gordon D. Schiff, MD: systems should support diagnostic
safety
Dr. Gordon D. Schiff of Harvard Medical School and Brigham
and Women’s Hospital has contributed to work on diagnostic safety and
closed-loop follow-up.
The lesson for healthcare leaders is not that individual
accountability is unimportant. It is that a system built around perfect human
memory is fragile.
If a process requires a staff member to check several
systems, remember a deadline, infer what is missing, contact another
organization, and reconstruct the context independently, the process is
vulnerable.
A stronger system makes the required action and its status
easier to see. It identifies unresolved steps and routes them to the right
person.
Caroline Pearson: AI needs the right incentives and
workflows
As executive director of PHTI, Caroline Pearson has
highlighted the risk that healthcare could layer AI onto existing broken
workflows without achieving the hoped-for improvements in cost and outcomes.
Her warning is particularly relevant to medical billing.
AI can assist with coding, documentation, claims, and denial
analysis. But if the underlying process rewards more administrative activity
rather than better outcomes and lower avoidable cost, the technology may
accelerate the activity without reducing the burden.
The lesson is not to avoid administrative AI. It is to
assess whether the technology is improving the entire process or simply making
one step faster.
The contrarian thesis: stop automating problems we should
have prevented
Healthcare AI is often marketed around the task it can
perform.
It can write the note faster.
It can identify the likely code.
It can summarize the medical record.
It can prepare a denial appeal.
It can extract information from a payer message.
It can classify an error and recommend a correction.
These capabilities can be valuable. But they invite a
question that often receives less attention:
Why did the work become necessary in the first place?
Imagine a hypothetical workflow.
First, information is captured incompletely.
Then a claim is prepared.
A system detects a problem.
An AI model classifies the problem.
Another tool recommends a correction.
A staff member reviews the recommendation.
A workflow routes the task.
Someone submits the corrected claim.
A dashboard records the outcome.
The organization may have built a highly sophisticated
system for resolving a problem that could sometimes have been prevented before
the claim was created.
That is not necessarily a failure of AI. It may be a failure
of the process in which AI was deployed.
The distinction is between using AI to process downstream
consequences and using intelligence upstream to reduce the conditions that
create those consequences.
The latter is harder.
It requires understanding where information originates, how
it changes, which context must be preserved, and which next action depends on
it.
It requires connecting clinical, operational, and financial
information without pretending that every system can be replaced or every
exception can be automated.
It requires knowing when information is incomplete and when
a human must intervene.
But it offers a more compelling objective than simply
processing more work per hour.
The goal should be less avoidable work, not merely faster
avoidable work.
What this means for independent clinics
Independent practices do not have unlimited budgets,
technical staff, or implementation capacity.
They need technology that works with the systems they
already use and solves clearly defined problems without creating an entirely
new operational burden.
That makes upstream problem-solving particularly important.
A practice should begin by identifying one recurring issue
that consumes substantial staff time, delays reimbursement, or creates
operational risk.
It might be eligibility-related rejections, missing
authorization information, recurring documentation gaps, delayed referral
completion, or claims that repeatedly require manual correction.
Then the practice should ask:
- Where
does the problem first originate?
- What
information is required to prevent it?
- When
should that information be available?
- Who is
responsible for ensuring it reaches the next step?
- What
evidence confirms that the step was completed?
- What
happens when the process fails?
- How
will the practice measure whether the intervention works?
These questions can reveal whether the organization needs
better training, clearer ownership, a revised workflow, improved system
configuration, an integration, or a new technology.
Not every problem requires AI.
Sometimes the best intervention is a more reliable
registration process, a clearly assigned task, a standardized documentation
requirement, or an explicit escalation rule.
Technology should follow the diagnosis of the operational
problem, not substitute for it.
A practical framework: find the failure before it becomes
a denial
Step 1: Choose one recurring problem
Do not begin with the ambition to transform the entire
revenue cycle.
Select one issue with a meaningful operational or financial
impact. Examples include repeated eligibility rejections, authorization-related
delays, missing documentation, or a particular category of denial.
A narrow starting point makes it easier to identify causes
and evaluate whether a change works.
Step 2: Trace the problem backward
Follow the issue through the workflow.
If a claim is rejected, examine the relevant steps from
scheduling and registration through documentation, coding, claim preparation,
and submission.
If a referral remains incomplete, examine the process from
the clinical decision through order creation, supporting documentation,
transmission, receipt, and scheduling.
Identify the earliest point at which the problem could
reasonably have been prevented.
Step 3: Separate root causes from symptoms
A denial code is a clue, not always a complete explanation.
A missing authorization may reflect an unclear policy, a
missed requirement, an ownership gap, or a communication failure.
A registration error may reflect outdated information or
inconsistent data entry.
Group recurring problems by root cause rather than treating
each occurrence as a separate incident.
Step 4: Define the handoff
Specify what information must be present, who owns the next
action, where the information must go, what confirms completion, and how
exceptions are escalated.
If these responsibilities are unclear, automation may only
make the ambiguity move faster.
Step 5: Use technology where it adds value
Existing EHRs, practice-management systems, referral tools,
eligibility services, revenue cycle platforms, and workflow automation may
already provide useful capabilities.
AI may help identify patterns, classify issues, summarize
information, or prioritize exceptions for review.
Choose the intervention that addresses the identified cause,
not the one with the longest feature list.
Step 6: Measure before and after
Establish a baseline before implementation.
Measure how often the problem occurs, how much time it
consumes, how long resolution takes, and what downstream consequences it
creates.
Repeat the measurement after the change using the same
definitions.
For many operational improvements, a 30- to 90-day
reassessment can provide an initial indication of whether performance is
changing. The appropriate period depends on the workflow and the volume of
cases.
Step 7: Look for unintended consequences
A new alert may reduce missed tasks while overwhelming staff
with notifications.
An AI tool may speed up coding while increasing the need for
review.
An automation may move work out of one department only to
create additional reconciliation work elsewhere.
Measure the entire process, not just the step that became
faster.
Measure outcomes, not just activity
A practice trying to reduce administrative noise should
track a balanced set of operational, financial, and quality measures.
Key Metrics for Identifying Upstream Healthcare Failures
Claims: Track the first-pass acceptance rate to determine
whether claims pass initial submission edits.
Denials: Monitor the initial denial rate to identify how
frequently claims are rejected or denied.
Rework: Analyze rework by root cause to uncover recurring
failures that generate unnecessary administrative work.
Revenue: Measure days in accounts receivable to understand
how long payments remain outstanding.
Collections: Track the net collection rate to assess how
effectively the practice collects expected allowed amounts.
Authorization: Measure the time from authorization request
to decision to identify delays in the approval process.
Referrals: Monitor referral completion rates to determine
whether patients reach the intended next step in care.
Follow-up: Track unresolved items beyond their target
deadlines to identify tasks that remain open too long.
Staffing: Measure staff hours spent on rework to quantify
labor consumed by correcting preventable problems.
Patient experience: Track reported administrative barriers
to identify avoidable coordination problems affecting patients.
The strategic question: Which recurring failures can be
prevented upstream rather than repeatedly corrected downstream?
Metric definitions matter. First-pass acceptance,
clean-claim rate, and first-pass resolution may be calculated differently by
different organizations. Practices should define the numerator, denominator,
exclusions, and observation period before comparing results.
A lower denial rate is useful, but it does not explain why
denials changed.
Faster processing is useful, but it does not prove that
total workload declined.
More closed tasks are useful, but only if the tasks
represent meaningful progress.
The most important question is whether the practice is
reducing the frequency and cost of avoidable failures.
The tools are changing. The need for process discipline
is not.
Several categories of technology can support more reliable
information handling.
EHR and practice-management systems can support
structured documentation, task ownership, and work queues.
Referral-management systems can help track outgoing
referrals, receiving organizations, scheduling status, and unresolved handoffs.
Eligibility and authorization tools can help staff
identify coverage information and payer requirements earlier.
Revenue cycle platforms can help classify denials,
manage follow-up, and monitor financial performance.
Workflow automation can route tasks and move approved
information between systems.
AI-assisted analysis can help identify patterns,
summarize available information, and prioritize cases that require human
review.
But no category of software replaces clear ownership.
Before purchasing a tool, ask whether it integrates with the
systems already in use, preserves an audit trail, handles incomplete
information safely, and demonstrates value against a defined baseline.
Ask what happens when the system cannot determine the
correct next step.
Ask whether the tool reduces the number of exceptions or
simply generates another list of them.
And ask how the practice will know whether it is working.
If a vendor can show how many tasks the software processes
but cannot define a meaningful outcome, the practice should be cautious.
Regulation and interoperability: necessary, but not
sufficient
The Centers for Medicare & Medicaid Services’
Interoperability and Prior Authorization Final Rule, CMS-0057-F, includes
requirements intended to improve data exchange and prior authorization
processes.
Some operational requirements generally began in 2026, while
several API requirements have compliance dates generally beginning in 2027,
depending on the payer and provision.
These changes matter because better information exchange and
greater transparency can reduce certain administrative barriers.
But interoperability does not automatically create
operational reliability.
Two systems can exchange data successfully while the
receiving organization still lacks a clear owner for the next action.
An API can transmit an authorization status. It cannot
guarantee that a practice acts on the information correctly.
A standard can improve the format of a message. It cannot
guarantee that every handoff is acknowledged or that every unresolved exception
reaches the right person.
The technology standards and the workflow must work
together.
Practices should not wait for every regulatory requirement
to take effect before reviewing their own processes.
Five myths keeping healthcare stuck
Myth 1: If we have an EHR, we have a reliable digital
workflow.
An EHR can store information without ensuring that the next
action occurs. Digital documentation is not proof of completed follow-up.
Myth 2: More automation always means less work.
Automation can reduce repetitive tasks, but it can also
create alerts, exceptions, reconciliation, and new monitoring responsibilities.
Measure the total burden.
Myth 3: Denials are solely a billing department problem.
Billing teams often resolve denials that originate in
registration, eligibility, authorization, documentation, coding, or payer
policy. Resolution responsibility does not necessarily identify the root cause.
Myth 4: AI can fill in missing context.
AI can identify gaps and help interpret available
information. It cannot reliably recover facts that were never captured or
verify assumptions without appropriate evidence.
Myth 5: Patients should simply be more proactive.
Patients should be encouraged to ask questions and
participate in their care. But patient engagement should complement reliable
systems, not replace them.
A patient who calls to check on a referral may uncover a
failure. That does not mean the patient should have been responsible for
detecting it.
Legal, ethical, and safety considerations
Upstream intelligence must be built around privacy,
security, clinical accountability, and appropriate oversight.
Protect patient information. Organizations should
determine whether a technology vendor requires access to protected health
information, whether a business associate agreement is required, and whether
the proposed use is permitted under applicable law and contracts.
Limit access appropriately. Role-based permissions,
audit logs, security controls, and retention policies should be considered.
Keep clinical decisions under appropriate professional
oversight. AI may identify a missing record or recommend that a referral
status be reviewed. It should not silently make clinical decisions that require
qualified professional judgment.
Preserve the distinction between fact and inference.
Generated summaries, extracted information, and inferred possibilities should
not be confused with original documentation.
Avoid unsupported promises. No technology should
claim to eliminate all denials, guarantee reimbursement, or prevent every error
without adequate evidence.
Evaluate fairness and accessibility. Digital
workflows should not create additional barriers for patients with disabilities,
limited digital access, language needs, or difficulty navigating complex
systems.
Measure unintended consequences. Faster processing
should not come at the expense of documentation quality, privacy, safety, or
appropriate review.
The objective is not automation at any cost. It is more
reliable information and responsible action.
OnnX: moving intelligence upstream of the claim
This is the problem space behind OnnX, an early-stage
AI-powered healthcare intelligence initiative focused on independent physician
practices and small to medium-sized clinics.
The core thesis is that the claim is often where a problem
becomes visible, not where the problem began.
A claim may be rejected because of an eligibility mismatch,
but the mismatch may have originated during registration.
A denial may involve missing documentation, but the
documentation gap may have started during the clinical encounter.
A payment may be delayed by an authorization issue that was
not identified early enough.
A billing team may spend hours chasing information that was
available earlier but did not reliably reach the right system or person.
The originating cause can be distributed across clinical,
operational, insurance, and administrative workflows.
OnnX’s product vision is to provide an upstream intelligence
layer that helps practices understand and reduce recurring information failures
without requiring them to abandon their existing EHRs, billing systems,
vendors, or established workflows.
The conceptual model is:
Capture → Structure → Preserve → Propagate → Act → Learn
- Capture:
Collect relevant information at the appropriate point in the process.
- Structure:
Organize information so people and software can interpret it consistently.
- Preserve:
Retain important context across handoffs and systems.
- Propagate:
Make relevant information available to the next authorized person, system,
or workflow.
- Act:
Surface gaps and appropriate next steps while maintaining human oversight.
- Learn:
Use practice-specific patterns and measured outcomes to improve future
performance.
The ambition is not to add another dashboard announcing that
a practice has 200 unresolved tasks.
It is to help explain why those tasks keep appearing and
identify opportunities to prevent their recurrence.
This matters for smaller practices, which often have limited
staffing, multiple technology vendors, and little capacity for complex
implementation projects.
They need technology that augments existing systems and
reduces unnecessary administrative work.
The capabilities described here are a product vision, not a
claim that every feature has already been validated or deployed. They must be
tested in real practices, with real workflows, and against measurable outcomes.
The evidence that matters is straightforward: fewer
avoidable errors, less repetitive rework, improved visibility into unresolved
problems, and more predictable administrative performance.
The future of healthcare AI: fewer exceptions, not just
faster processing
Healthcare technology has traditionally focused on systems
of record and systems of action.
Systems of record store information. Systems of action help
people perform tasks.
The next opportunity is to improve the intelligence
connecting those functions: what information is required, where it originated,
what is missing, where it needs to go, and whether the intended action
occurred.
This is not simply another chatbot or a larger language
model.
It is an approach to using context across fragmented
information environments so that predictable gaps can be identified earlier and
appropriate action can follow.
In a medical practice, it might help identify recurring
eligibility problems before they generate another batch of rejected claims.
It might surface a referral that has not reached the next
stage.
It might help a billing team discover that apparently
different denials share a common upstream cause.
It might distinguish a payer-specific requirement from a
practice-specific documentation gap.
Such capabilities require reliable information, thoughtful
integration, governance, human oversight, and evidence that the intervention
actually works.
Healthcare is too complex for any credible technology to
promise that all uncertainty can be eliminated.
But complexity is not an excuse for accepting every
recurring failure as inevitable.
We can build systems that are explicit about what they know,
what they do not know, and what remains unresolved.
We can make responsibility for the next action visible.
We can use AI to reduce unnecessary work rather than merely
making unnecessary work faster.
And we can judge innovation by whether it improves outcomes
and reduces avoidable burden, not simply by the number of tasks it processes.
That is the opportunity behind upstream intelligence.
What Judy Brauning’s story should make us ask
Judy Brauning’s story is not evidence that a particular AI
product would have prevented her experience. We do not have enough information
to make that claim.
It is a real patient story that raises a broader question
about how healthcare coordinates important next steps.
When a referral fails to reach its destination, the failure
may become visible only after someone investigates.
When documentation is incomplete, the problem may become
visible only when a claim is reviewed.
When authorization requirements are missed, the consequences
may emerge only after a service is delayed or a payment is disputed.
The organization may eventually resolve the problem.
But resolution is not prevention.
And the fact that a capable employee can repair a failure
does not mean the failure was necessary.
Healthcare has extraordinary clinicians, dedicated
administrative teams, sophisticated medical technology, and increasingly
powerful AI.
The challenge is to connect those capabilities to more
reliable processes.
We should celebrate technology that helps physicians make
better decisions and patients receive better care.
We should also ask harder questions about the administrative
work that should never have existed in the first place.
Perhaps the next major healthcare innovation will not be
another tool that helps us manage the consequences of fragmented information.
Perhaps it will be a system that makes fewer fragments in
the first place.
Perhaps the most impressive AI demonstration will not be a
machine that explains a denial in seconds.
It will be a practice that no longer spends hours resolving
that same preventable denial.
Healthcare does not need to automate every problem it has
learned to tolerate. It needs to identify which problems should never have been
created.
Frequently asked questions
What does Judy Brauning’s story reveal about healthcare
administration?
Her published story describes a referral that had not been
submitted to the cancer center she selected. She contacted the center herself
and secured an appointment. The account illustrates how a failed handoff can
leave patients responsible for detecting and navigating an administrative gap.
Did the missing referral cause Judy Brauning’s cancer to
progress?
The published account does not establish that causal
relationship. It documents the referral issue and her subsequent treatment, but
readers should not infer that the administrative failure caused her cancer to
progress or changed her outcome.
How are referral failures connected to medical billing?
The processes are different, but both depend on accurate
information, clear responsibility, and reliable handoffs. Billing problems can
arise when eligibility, authorization, documentation, coding, or other
information is missing or inconsistent earlier in the process.
What is upstream intelligence in healthcare?
Upstream intelligence means organizing and using relevant
information earlier in a process so that predictable gaps can be identified
before they create downstream rework, delays, or errors. It should support
human judgment rather than replace clinical accountability.
Can AI eliminate medical billing denials?
No credible solution should promise to eliminate all
denials. Some require clinical judgment or arise from coverage limitations,
payer rules, or other factors. AI may help identify patterns and prevent
certain avoidable errors, but results must be measured in real-world workflows.
What should an independent medical practice measure
first?
Start with one recurring problem that creates substantial
work or financial impact. Measure its frequency, root causes, resolution time,
staff burden, and downstream consequences. Establish a baseline before
implementing a change and repeat the measurement afterward.
How can a practice improve referral follow-up?
Define ownership, required information, expected timelines,
confirmation of receipt, completion criteria, and escalation procedures. Track
unresolved referrals and review exceptions rather than assuming that an order
or transmission guarantees completion.
What is OnnX trying to achieve?
OnnX is an early-stage healthcare intelligence initiative
exploring how independent practices can identify and reduce upstream
information failures that contribute to administrative and billing rework. Its
product vision is to augment existing systems and workflows, with effectiveness
to be validated through practice assessments and measurable outcomes.
Final thought: what if the real problem is the work we
keep creating?
Judy Brauning had to intervene in her own care to move a
referral forward. Her determination is admirable. The administrative gap is the
part worth examining.
In medical billing, the same question deserves attention
whenever a team repeatedly chases information, corrects preventable errors, or
reconstructs context that should have been available earlier.
We can continue investing in tools that make these
activities faster.
Or we can ask whether some of these activities should exist
at all.
That is a harder question than choosing a software vendor.
It requires examining systems, incentives, handoffs, and accountability. It
requires measuring outcomes rather than celebrating implementation. And it
requires recognizing that the people managing administrative failures are often
compensating for weaknesses they did not create.
For OnnX, that is the central opportunity: move intelligence
upstream, preserve context, and help practices prevent avoidable administrative
noise before it reaches the claim.
Not every problem can be prevented. Not every delay is
avoidable. Not every denial is an error.
But recurring failures deserve more than another workaround.
If healthcare can use AI to write a denial appeal in
seconds, why are we not equally focused on preventing the avoidable denial in
the first place?
I welcome perspectives from physicians, practice owners,
medical billers, administrators, and healthcare technology leaders.
What is one recurring administrative task in your
practice that should never have existed?
Share your experience in the comments. Repost this article
if you believe healthcare innovation should focus not only on doing work
faster, but on preventing unnecessary work altogether.
P.S. I have a free resource in the Featured section
of my LinkedIn profile. No signup required.
About the author
Dr. Daniel Cham is a physician, medical consultant,
and entrepreneur focused on healthcare technology, medical practice operations,
and the administrative challenges affecting independent clinics. He is the
founder of OnnX, an AI-powered healthcare intelligence initiative exploring how
practices can reduce downstream billing noise by addressing information gaps
upstream.
His work focuses on the intersection of clinical care,
operational intelligence, medical billing, and practical AI adoption.
Disclaimer
This article is for educational and informational
purposes only. It is not medical, legal, billing, coding, reimbursement, or
compliance advice. Judy Brauning’s story is based on a publicly available
patient profile from VCU Massey Comprehensive Cancer Center. The article does
not establish that the documented referral delay caused cancer progression or
altered her clinical outcome. Statistics are attributed to their cited sources
and should be interpreted in the context of the underlying research and survey
methods. OnnX’s capabilities described here represent a product vision and
should not be interpreted as proof of validated clinical or financial outcomes.
Follow Dr. Daniel Cham
·
Listen to the
podcast on Spotify
·
Subscribe
and watch on YouTube
·
Follow
Dr. Cham on Instagram
References and further reading
- VCU
Massey Comprehensive Cancer Center. “Judy Brauning: ‘Massey was the
right place for me.’” October 15, 2026.
https://www.masseycancercenter.org/news/2026/judy-brauning-massey-was-the-right-place-for-me/ - Axios.
“The barriers to realizing AI’s promise in health care.” October 14, 2026.
https://www.axios.com/2026/10/14/ai-health-care-spending - Peterson
Health Technology Institute. “Administrative AI: Current Use and
Potential Impact.”
https://www.phti.org/research-insight/administrative-ai-current-use-and-potential-impact/ - American
Medical Association. “AMA survey: Prior authorization reform pledge
falls short with physicians.”
https://www.ama-assn.org/press-center/ama-press-releases/ama-survey-prior-authorization-reform-pledge-falls-short-physicians - Agency
for Healthcare Research and Quality. “Closed-Loop Diagnostics.”
https://www.ahrq.gov/patient-safety/resources/learning-lab/closed-loop-long-desc.html - Centers
for Medicare & Medicaid Services. “CMS Interoperability and Prior
Authorization Final Rule CMS-0057-F.”
https://www.cms.gov/newsroom/fact-sheets/cms-interoperability-prior-authorization-final-rule-cms-0057-f
Hashtags
#HealthcareAI #MedicalBilling #RevenueCycleManagement
#HealthcareInnovation #PatientSafety #PhysicianEntrepreneur #HealthTech
#MedicalPracticeManagement #ArtificialIntelligence #OnnX
