“For too long, fee-for-service payment has rewarded volume over value, fueling a reactive system of sick care.” — Mehmet Oz, MD, CMS Administrator

The claim was clean. The afternoon was ruined.

Imagine this scene in a small physician practice. It is an illustrative story, not a documented patient case or an OnnX customer testimonial.

The physician has finished the visit. The patient has gone home. The note is signed. Everyone has done something useful.

Then the claim comes back.

The authorization covers a different service. The referral attachment contains an old insurance card. A payer name looks almost identical to another payer name. Somewhere between the front desk and submission, a small mismatch became an expensive afternoon.

The billing specialist opens another tab.

The practice manager sends another message.

The physician receives another request to clarify something that felt perfectly clear yesterday.

Nobody delivered more care. Everybody did more work.

Now enter the software demonstration. A cheerful dashboard promises to generate appeal letters in seconds.

Wonderful. We have taught the computer to write faster apologies for mistakes the clinic should never have made.

That is the uncomfortable question behind AI medical billing: Are we improving the practice, or simply making its broken habits more productive?

My view is straightforward. The most valuable billing technology should help clinics prevent avoidable failures, protect their knowledge, and preserve human judgment. Faster processing matters. Fewer unnecessary loops matter more.

As a physician and founder of OnnX, I am interested in technology that gives independent clinics more control over their operations. That requires honesty about what AI can do, what it cannot do, and what a vendor presentation tends to skip.

The future of billing should contain fewer administrative middlemen. It should also contain fewer software subscriptions acting like administrative middlemen with better typography.

Seven tips before you buy another promise

For busy physicians and clinic owners, start here:

1. Follow one failed claim backward. Find the earliest point where someone could reasonably have prevented the problem.
2. Choose one workflow. A narrow pilot produces clearer evidence than a grand transformation announcement.
3. Separate claim acceptance from payment. Passing an electronic check does not establish coverage or medical necessity.
4. Ask what happens to your information. Training, storage, retrieval, human access, and deletion are different questions.
5. Measure total work. Include review, corrections, meetings, patient calls, and exceptions.
6. Protect the exit. Know how to export records, revoke access, and keep operating during an outage.
7. Reward prevention. Celebrate the missing authorization caught before the visit as well as the denial overturned afterward.

These steps sound modest. That is their advantage.

Practices rarely need another complicated framework before they can identify the insurance field that keeps breaking their week.

The real product is a reliable handoff

Medical billing errors often begin before anyone selects a billing code.

A patient changes insurance. A referral arrives as a scanned document. A scheduled procedure changes. A clinical note omits the detail needed to support the service. Staff copy information because the systems do not exchange it reliably.

Then the billing department inherits the consequences.

Calling this only a billing problem is like blaming the kitchen sink for a roof leak. The water arrives there. That does not mean it started there.

For independent physician practices, the useful unit of improvement is the handoff: patient to scheduler, scheduler to front desk, clinician to coder, practice to clearinghouse, payer response to follow-up team.

At each handoff, ask three questions. What information must be correct? Who owns the check? What happens when the information is missing?

AI may help extract data, compare documents, flag contradictions, and draft explanations. Fixed rules may handle simpler checks more reliably and cheaply. Experienced staff must still resolve ambiguity.

The point is not to worship one technology. It is to make the next person less dependent on detective work.

A reliable workflow carries context forward. A fragile workflow carries questions forward.

That distinction belongs in every buying decision.

Recent news: Medicare just supplied a practical stress test

The news window here is November 3–9, 2026. The three dated references near the end were published on November 4. The broader operational lessons are intended to remain useful after this news cycle.

CMS has released its 2027 Medicare Physician Fee Schedule final rule, with policies generally taking effect on or after January 1, 2027. The agency also published related Shared Savings Program changes and an explanation of its policy direction.

For clinic owners, this is a reminder to distinguish a proposed rule, a finalized policy, and an implemented software update. Those are three separate events, sometimes separated by a surprising amount of confusion.

The immediate task is to identify affected services, confirm effective dates, assign an owner, and test representative workflows against authoritative requirements. Medicare policy does not automatically determine every commercial payer’s implementation.

The practical connection to AI is simple: automation depends on maintained rules. A system can execute an obsolete instruction with remarkable consistency.

My interpretation of this week’s developments is that practices should evaluate technology on its ability to manage change, not merely process today’s queue. Ask how policy updates are reviewed, approved, tested, and rolled back.

Do not mistake this interpretation for a CMS endorsement of a particular billing product. It is an operational lesson drawn from a changing payment environment.

Statistics: adoption is evidence of interest, not proof of value

The AMA’s 2026 physician AI survey reported that 81% of respondents used AI professionally, compared with 38% in 2023. It also found that 70% saw opportunities to automate tasks contributing to burnout, while 86% considered privacy assurances important and 88% emphasized safety and efficacy validation.

These are survey findings about physician use and attitudes. They do not establish that every AI product improves collections, reduces denials, or saves a specific amount of staff time.

For your own practice, four measures deserve more attention than a vendor’s largest percentage:

• Initial denial rate: Initially denied claims divided by adjudicated claims in a defined cohort.
• First-pass acceptance: Initial submissions accepted by the specified receiving system, using a consistent denominator.
• A/R over 90 days: Outstanding receivables older than 90 days divided by total outstanding receivables, using the same aging convention.
• Rework time: Staff minutes spent correcting, resubmitting, appealing, and resolving preventable problems.

Document whether you count claims, lines, encounters, or dollars. Those measures answer different questions.

A declining denial percentage can conceal rising denied dollars. Faster acceptance can coexist with slower payment. A beautiful dashboard can display all of this in calming blue.

A number becomes useful when its definition survives a meeting.

Expert opinion round-up: three perspectives, three practical lessons

These perspectives come from published statements and commentary. They are not interviews, endorsements of OnnX, or evidence that the experts agree with every recommendation here.

John Whyte, MD, MPH: keep physicians involved

In the AMA’s 2026 announcement, John Whyte emphasized that AI should support physicians and be implemented responsibly. His comments connect enthusiasm about reduced administrative burden with the need for safety and trust.

My practical takeaway: physician participation belongs in implementation decisions. A billing workflow can affect documentation, clinical communication, and patient trust. Give a physician responsibility for reviewing those connections.

Do not invite clinical input only after the software has become mandatory.

Michael Turken, MD, MPH: rethink where work happens

In a June 20, 2026 KevinMD essay, Michael Turken argued for tools that help patients and caregivers manage more care safely outside traditional encounters. He highlighted substantial work that families perform without formal recognition.

My practical takeaway: make home information useful to the care team without treating every patient action as a billable event. His essay offers a provocative design argument. It does not establish payer coverage.

Visibility is valuable. Reimbursement still requires its own evidence.

Mehmet Oz, MD: examine what payment rewards

In the November 4 HHS announcement, Mehmet Oz criticized payment incentives centered on service volume and described the administration’s rationale for the new rule. That is a policy position, not proof of its eventual outcomes.

My practical takeaway: evaluate whether your internal incentives reward accurate care delivery, useful coordination, and fewer errors. Counting submissions alone can reward activity while hiding avoidable work.

Together, these perspectives suggest a useful discipline: keep clinicians accountable, support patients realistically, and measure the work that matters.

Ghost work deserves recognition, not a billing costume

The attached story describes ghost work: the symptom logs, medication checks, wound observations, and care coordination that patients and caregivers perform at home.

The phrase captures something real. Much of healthcare happens between appointments, without an audience and without a neat place in the record.

AI may help organize that information. It might summarize a patient’s reported symptoms, identify missing entries, or prepare a draft for clinical review. But a structured summary does not automatically become a covered service.

Documentation is evidence. It is not permission.

Before billing any remote service, verify the applicable code, payer policy, patient eligibility, medical necessity, consent, staff qualifications, time or data requirements, supervision, and restrictions on overlapping services. Requirements differ by service and payer.

Do not infer that a patient answering questions creates reimbursable physician work. Do not turn automated processing time into human clinical time. Do not count the same work twice because two applications observed it.

The FDA’s ACNU framework concerns an additional condition for nonprescription drug use. It is not a general Medicare reimbursement pathway. Its existence does not establish coverage for AI-generated summaries or home questionnaires.

The ethical opportunity is to reduce the distance between useful information and appropriate care. The financial opportunity must be demonstrated separately.

Otherwise, ghost work becomes ghost revenue: impressive in a presentation, difficult to locate in the bank account.

The privacy trap is real. The explanation needs work.

The original story warns that using a shared AI model automatically teaches competitors your billing secrets. That description confuses several different processes.

Inference is using a model to produce an output. Routine inference does not inherently update the model’s underlying weights. A system can use information during a session without incorporating it into a shared model.

Separate risks still deserve serious attention: retained prompts, shared retrieval systems, training permissions, fine-tuning, human review, insecure logs, misconfigured access, and subcontractor practices.

Those risks depend on the product, architecture, configuration, contract, and actual operating controls.

Ask the vendor to answer these questions in writing:

• Can prompts, outputs, attachments, or feedback be used for model training?
• Are records retained, and for how long?
• Can another customer retrieve our information or derived materials?
• Who can inspect content, including support staff and subcontractors?
• What happens to backups, logs, and customized models after termination?
• Can we restrict secondary uses and verify those restrictions?

For practice intelligence, include more than patient identifiers. Your payer escalation notes, internal checklists, staffing assumptions, and negotiated terms may be commercially sensitive.

However, privately hosted infrastructure is not automatically safe. Someone must patch it, control access, test recovery, and respond to incidents.

Privacy is a property of the whole arrangement. It is not a location on a server diagram.

Your AI bill may be cheaper per task and higher per month

Another useful idea in the story is the token tax: the accumulated cost of repeated AI processing.

The specific enterprise spending figures in the transcript are not used here because they were not independently established for this article. Small practices should be especially cautious about borrowing hospital-scale numbers.

Still, the underlying arithmetic deserves attention.

A workflow may extract a document, search a record, draft an answer, check the answer, retry after an error, and ask a reviewer to resolve an exception. Each step can consume resources.

Lower unit prices do not guarantee a lower total bill when the number of steps grows.

Ask for cost per completed, validated task. Include platform charges, integration, storage, support, human review, and implementation labor. Also ask whether failed attempts and retries are billable.

Consider a hypothetical pilot. Monthly technology costs are $600. Setup costs of $1,200 are spread across six months, adding $200 monthly. Staff save 30 hours but spend eight hours reviewing outputs. At an assumed $30 hourly labor value, the 22 net hours represent $660 of capacity.

Against $800 in monthly allocated costs, that leaves a $140 shortfall before other benefits. These invented figures illustrate a method, not an OnnX result.

Time saved may improve service without reducing payroll. Treat capacity, cash savings, and recovered collections as separate outcomes.

Myth busters: six claims worth challenging

Myth one: a clean claim guarantees payment

Reality: Acceptance, adjudication, and payment are different stages. A claim can pass a format check and still fail a coverage requirement. Always name the stage when reporting performance.

Myth two: more automation means less work

Reality: Automation can move work into exception handling, validation, or patient complaints. Measure the complete workflow. A shorter queue in one department can be a longer queue wearing someone else’s badge.

Myth three: shared models always learn from your prompts

Reality: That is not an inherent feature of inference. Review actual training permissions, retention settings, access boundaries, and retrieval design. Ask precise questions instead of accepting precise-sounding fear.

Myth four: a BAA makes the product safe

Reality: A business associate agreement is an important legal arrangement when applicable. It does not replace risk analysis, secure configuration, monitoring, or workforce training.

Myth five: AI can make missing documentation disappear

Reality: AI can highlight a gap. It must not invent a clinical finding or rewrite history to support a charge. An elegant sentence can still be false.

Myth six: the best system needs no human review

Reality: Review should match risk and demonstrated performance. Some narrow administrative checks may need little intervention. Ambiguous coding and clinical assertions require stronger oversight. Confidence is a software output, not professional licensure.

Pitfalls: where sensible pilots become expensive hobbies

The first pitfall is starting too broadly. When a practice changes intake, coding, eligibility, and appeals simultaneously, improvement becomes difficult to attribute. Failure becomes equally difficult to diagnose.

The second is buying a demonstration rather than a workflow. Demonstrations use cooperative documents. Clinics receive crooked scans, conflicting dates, and insurance cards photographed beside a sandwich.

Test the awkward cases. They are part of the business.

The third is double-counting benefits. If an avoided denial also reduces rework, do not casually add overlapping estimates as though they represent independent cash gains. Show assumptions clearly.

The fourth is forgetting patient communication. A billing change that increases confusing messages or unexpected balances can damage trust even when internal processing improves.

The fifth is silent policy drift. A once-correct rule becomes outdated. Someone assumes the vendor updated it. The vendor assumes the clinic configured it. The payer declines the opportunity to resolve the misunderstanding philanthropically.

The sixth is treating staff concerns as resistance. The person who says, “That will fail on our referrals,” may be providing the cheapest risk assessment you will ever receive.

The seventh is confusing momentum with evidence. Weekly meetings, new dashboards, and enthusiastic emails are activities. Better outcomes require a baseline, comparable follow-up, and enough observation time.

A pilot without a decision rule is a subscription waiting to happen.

Legal implications: accountability survives the software contract

Start with a practical premise: a practice should not assume that outsourcing a task transfers every responsibility attached to it.

For systems handling protected health information, assess the parties’ HIPAA roles, permitted uses, safeguards, and required agreements. HHS guidance explains that cloud providers handling electronic PHI may be business associates even when they cannot decrypt the information. Appropriate agreements and risk management remain necessary.

For claims, maintain support for the services actually delivered. AI-generated text cannot excuse false facts, unsupported codes, duplicate billing, or inaccurate representations. The legal consequences depend on the conduct and applicable law; a software error is not automatically fraud.

For sensitive legal discussions, involve counsel before placing privileged materials in an AI system. Privilege questions depend on jurisdiction, facts, confidentiality, purpose, and arrangements with the provider. This article does not rely on the transcript’s asserted court ruling or its blanket claim that vendor-hosted AI eliminates privilege.

For contracts, review data rights, permitted secondary uses, incident notification, subcontractors, export obligations, termination assistance, liability allocation, and limitations of liability with qualified counsel.

Keep approval records and change histories. If a claim is challenged later, the practice needs to explain what happened, which information supported it, and who approved the action.

An audit trail should explain a decision, not merely prove that a button was clicked.

Ethical considerations: the patient is not your integration layer

There is a tempting way to reduce administrative burden: ask the patient to do it.

Upload the card. Reenter the medication list. Complete the same questions in another portal. Confirm details the clinic already holds. Call the insurer. Repeat until morale improves.

That may improve an internal productivity report while making care harder to access.

For patient-centered automation, measure the patient’s work too. Count repeat requests, abandoned forms, confusion, and requests for help. Offer alternatives for patients with limited connectivity, language barriers, disability, or low digital confidence.

Caregiver information also requires careful handling. Useful family participation should not erase patient preferences or appropriate access boundaries.

Review whether performance varies across patient groups and document types. A system that handles typed English well may perform differently with translated material, dictated notes, or difficult scans. Investigate the failure rather than blaming the patient.

Be clear when automated messages are used. Provide a practical route to a person, especially for disputes, financial hardship, or clinically concerning information.

And avoid incentives that encourage more billable activity without clear clinical purpose.

The most efficient billing system can still be a poor healthcare system if it makes patients carry the friction.

Ethical design asks whether the work disappeared, moved, or became invisible. Those are different outcomes.

Practical considerations: the small practice advantage

Small clinics have fewer resources. They may also have shorter communication lines and a clearer view of recurring problems.

Use that advantage.

Bring a physician, a front desk representative, and a billing specialist together for 30 minutes. Review a handful of recent failures. Ask where information was first available, where it changed, and where someone first noticed the problem.

Do not begin with blame. Begin with sequence.

Then choose a correction that fits ordinary staffing. A solution that requires the practice manager to inspect every record forever is a temporary workaround with a managerial title.

Prefer small, observable changes: a required authorization check for one procedure group, a payer routing rule for one recurring mismatch, or an extraction tool that highlights uncertain fields before entry.

Budget for training and protected review time. Staff cannot simultaneously maintain full production and conduct a careful pilot without a tradeoff somewhere.

Document the fallback before go-live. Who handles incomplete records? Who can pause the tool? Which tasks continue manually? Where are pending items tracked?

A small practice does not need to imitate a hospital’s governance department. It needs clear responsibility, useful records, and a workable recovery plan.

Sophistication is knowing which safeguards matter. Complexity is buying all of them without knowing why.

Step by step: a bounded pilot physicians can actually manage

Step one: choose one recurring failure

Select a problem with enough volume to observe and a clear owner. Examples include authorization mismatches, wrong payer routing, or missing information in referral documents. Define the scope narrowly enough that staff can explain it without a slide deck.

Step two: establish a baseline

Review a recent, representative period. Record volume, error frequency, staff time, and downstream consequences. Use enough cases to understand normal variation. A small sample can guide exploration, but it cannot support sweeping performance claims.

Step three: map the earliest useful intervention

Identify where the relevant facts first exist. If the mismatch can be caught during scheduling, do not wait until claim submission. Earlier is useful only when the information is available and the check will not unnecessarily delay care.

Step four: separate rules from interpretation

Use a deterministic check for clear requirements, such as missing fields or a mismatched identifier. Consider AI when the task involves interpreting variable documents. Route uncertainty to someone with the authority and context to resolve it.

Step five: test in observation mode

Compare suggestions with actual staff decisions before allowing consequential actions. Include incorrect dates, duplicate documents, ambiguous payer names, and incomplete scans. Record false alarms and missed problems. Reviewers should see the supporting source, not just a confident answer.

Step six: agree on targets and stop conditions

Set local targets for error reduction, rework, review burden, cost, and safety. For example, a practice might require lower correction time without increased missed authorizations. These are practice-specific pilot criteria, not universal industry standards.

Step seven: run a controlled live pilot

Train the team, limit permissions, and monitor exceptions. Use a defined patient or service cohort where appropriate. Document concurrent changes in payer mix, staffing, volume, or policy so they do not masquerade as software benefits.

Step eight: decide and preserve the learning

Choose to expand, revise, pause, or stop. Allow sufficient follow-up for adjudication and collections before judging financial outcomes. Keep successful rules and failure examples even if the vendor relationship ends. The practice should retain its learning.

Tools, metrics, and resources: a useful minimum

You can begin with existing practice management reports, remittance information, clearinghouse responses, a secure issue log, and a basic time study.

The tools should answer practical questions:

• What failed, and at which stage?
• What source information supports the correction?
• How often does the same cause recur?
• How much review does the intervention create?
• What changed for staff, patients, and cash flow?

Use payer manuals and authoritative coding resources for requirements. Use CMS materials for relevant Medicare policies. Use qualified coding, compliance, legal, and security support when the question exceeds the team’s expertise.

For an AI tool, request a concise system description covering intended use, limitations, information access, evaluation methods, update process, and human oversight. A polished model card can help, but it does not replace local testing.

For staff, provide a short operating guide with screenshots or examples, escalation contacts, and pause instructions. Make it easy to report a wrong answer without starting an argument about the future of technology.

For metrics, review both false positives and false negatives. A system that flags every claim may find every problem while saving nobody any time.

Also track overrides. Frequent overrides can indicate poor rules, unclear instructions, or useful staff knowledge that has not been incorporated.

The best operational tool makes the next decision clearer.

Insights: three reframes that change the buying conversation

First, billing is an information quality problem with financial consequences. This reframing makes scheduling, intake, clinical documentation, and payer rules part of the solution. It also explains why a downstream tool may struggle with an upstream omission.

Second, independence requires operational control. A practice can outsource useful work while retaining access, oversight, and an exit path. Conversely, owning a server does not guarantee control if nobody understands the workflow running on it.

Third, staff knowledge is infrastructure. The billing specialist who recognizes a recurring payer mismatch holds practical knowledge that deserves documentation, review, and preservation.

Ask that person to explain the signal. Convert the explanation into a maintained rule where possible. Keep the source and effective date. Invite correction when circumstances change.

This is especially relevant when modernizing legacy software. Old screens can contain outdated habits. They can also encode decades of useful exceptions. Replacing the interface before understanding the rules risks discarding the part that actually works.

The right question is not whether the old system looks modern. It is which knowledge deserves to survive.

A clinic’s advantage may live in the workaround everyone forgot to document.

Modernization should make that knowledge visible before making it portable.

What failure should look like in an honest article

I cannot offer a dramatic OnnX case study here without verified pilot evidence. Inventing one would make the article more exciting and the advice less trustworthy.

Instead, consider a clearly hypothetical failure.

A clinic tests document extraction. The tool reads most fields correctly. Everyone is encouraged. Then the team discovers that reviewers spend longer checking the outputs than they previously spent entering the fields.

The technical demonstration succeeded. The operational pilot failed.

That is useful information.

The next move might be to narrow the document types, highlight only uncertain fields, improve source quality, or stop the project. The wrong move is to redefine success as adoption because the contract has already been signed.

Another hypothetical: an appeals assistant produces persuasive letters, but the practice never fixes the scheduling error generating those denials. Appeal throughput improves. The underlying defect continues.

Here, the failure is choosing the wrong intervention point.

For OnnX, this reinforces the standard I want to apply: prove a bounded improvement before claiming a broad transformation. That is a development principle, not a statement that every capability discussed here is already available.

Trust grows when a founder can distinguish a product goal from a measured result.

A failed pilot can save money. An unmeasured success can spend it indefinitely.

Questions worth asking in the next vendor meeting

Ask the vendor to show a failure before showing another success.

Which documents cause trouble? Which payer changes require manual updates? When does the system refuse to act? What does a reviewer see when the recommendation is uncertain?

Then ask for a complete journey through one ordinary claim. Watch the handoffs. Count the systems. Notice where someone has to copy information or infer what happened elsewhere.

Ask who maintains the connection when an interface changes. Ask what support includes. Ask how long implementation takes for a clinic with your staffing and your software, rather than an ideal customer with an unusually patient IT team.

For interoperability, ask for tested mappings and exception handling. HL7 is a family of standards, and FHIR is itself an HL7 standard. Calling them inherently incompatible languages oversimplifies the problem. Reliable exchange also depends on interfaces, terminology, identity matching, permissions, and local configuration.

AI extraction may help with unstructured material. It cannot make those dependencies disappear by calling itself a universal translator.

Finally, ask: If this does not work, what exactly do we keep?

The answer should include your records and an understandable account of what the system did. Depending on the agreement, it may also include configurations, rules, and transition support.

The exit conversation is part of the purchase conversation.

The Monday morning test: can your team use this?

Before approving a pilot, imagine an ordinary Monday with two absent employees, a full schedule, and an insurer changing its portal password requirements at the least poetic moment.

Can the team still explain the process?

If the answer depends on one unusually knowledgeable employee being available, you have identified a dependency. Document it before adding automation around it.

Try a simple claim review huddle. Select five recent problem claims. Remove unnecessary patient identifiers from the discussion materials. For each claim, identify the first incorrect or missing fact, the person who could have seen it, and the action that would have helped.

Keep payer errors separate from practice errors. Otherwise, staff may be held responsible for decisions outside their control, and the improvement plan will chase the wrong cause.

Then ask what evidence would change your mind about the proposed solution.

Perhaps the missing information was unavailable until after the service. Perhaps the authorization was correct but the payer processed it incorrectly. Perhaps the field was accurate in one system and altered during transfer.

Each possibility points toward a different intervention. One calls for a later check. Another calls for escalation. Another calls for interface repair.

Diagnosis before automation is a useful rule for administrative medicine too.

Assign one person to maintain the issue log and another to review the clinical or coding implications when needed. Agree on a short weekly review. Keep the meeting focused on recurring causes, unresolved exceptions, and actions with owners.

Avoid making the log a museum of old frustrations. Every active issue needs a next step or a reason to close it.

For patient-facing changes, test the wording with someone outside the billing team. Ask them to explain the message back. If they cannot tell whether they owe money, need to provide information, or should simply wait, revise it.

Confusion is work. It usually returns as a phone call.

Finally, run a short downtime exercise. Revoke a test account, simulate an unavailable connection, and confirm that pending tasks remain visible. Do this safely in a test environment or a carefully controlled exercise, without interrupting patient care.

Record who can authorize a pause, contact support, and restore service. Confirm that the fallback preserves privacy and does not lose charges or clinical messages.

These exercises will not look impressive in a launch announcement. They will look very impressive to the employee trying to keep Tuesday from becoming a backlog.

Readiness is the ability to recover on an ordinary day, with ordinary people, using instructions they can actually find.

Before expanding, ask staff whether they would choose to keep the tool if the decision were theirs. Their answer is not the only measure, but it can reveal hidden review burdens that collection reports and vendor satisfaction surveys routinely miss.

Future outlook: the scarce resource will be trustworthy judgment

AI capabilities will continue to change. Prices, products, and regulations will change with them. Predictions about a particular model’s dominance are less useful to a clinic than habits that survive those changes.

Three habits seem durable.

Maintain clear source information. Know which document, policy, or clinical observation supports a claim.

Maintain accountable decisions. Know who can approve a recommendation, resolve a conflict, and stop an unsafe process.

Maintain reversible operations. Know how to change a rule, recover a record, and continue when a vendor becomes unavailable.

More capable agents may eventually perform longer administrative sequences. That makes permission design more consequential. Reading a document, editing a record, submitting a claim, and sending a patient message should not automatically share the same authority.

The winning practice will not necessarily own the most AI. It will understand which decisions can be delegated, which require review, and which should remain firmly human.

For independent clinics, this is an opportunity to build stronger operations without surrendering professional identity. But technology should earn responsibility through evidence.

Trust should expand with demonstrated reliability, not with the length of a product roadmap.

Frequently asked questions

Will AI replace our billing staff?

Some tasks may become easier to automate. Others will require more interpretation, oversight, and patient communication. Plan around actual workload and performance. Do not promise staffing reductions before measuring review time, exceptions, and service quality.

Where should a small clinic start?

Choose a frequent, bounded problem with a clear owner and measurable consequences. Missing authorization information or recurring payer routing errors may be useful candidates. The right starting point depends on your practice’s actual failures, not a generic vendor checklist.

Does using a shared model train it on our information?

Not inherently. Inference, retention, retrieval, and training are separate processes. Review the vendor’s terms, technical arrangements, configuration, and subcontractors. A privacy assurance should identify the specific uses that are prohibited and the controls supporting that promise.

Can patient-generated information create billable services?

It can support care and may contribute to qualifying services when applicable requirements are satisfied. Patient activity alone does not establish reimbursement. Verify the service, payer, medical necessity, documentation, consent, staff involvement, and other relevant conditions before billing.

Is private hosting always the safest choice?

No. Private hosting changes responsibilities and may offer useful control, but safety depends on implementation and maintenance. Compare access controls, monitoring, patching, recovery, contracts, and staffing. Choose an arrangement the practice can govern competently.

How soon should a pilot show financial results?

Operational signals may appear before collections mature. Match the evaluation period to the outcome. Review time and detected errors can be assessed early; adjudication, payment, and aging require longer follow-up. Avoid drawing a cash-flow conclusion from an unfinished claim cohort.

What evidence should we request from a vendor?

Request a clear intended use, representative evaluation results, failure examples, implementation requirements, security information, and contract terms. If customer references are offered, ask whether their setting resembles yours. Evidence from a different specialty or workflow may not transfer reliably.

Should we eliminate every billing intermediary?

Remove unnecessary handoffs, duplication, and unclear accountability. Retain useful expertise and infrastructure when they improve outcomes. The goal is a simpler, more accountable operation. A specialist who prevents errors may add more value than a supposedly direct workflow that creates them.

Final thoughts: make the second attempt unnecessary

Independent physicians do not need another reason to work harder after the patient leaves.

They need accurate information, sensible rules, capable staff, and technology that reduces avoidable friction.

Denial prevention does not mean every denial is preventable. Payer errors, coverage disputes, changing policies, and clinical complexity remain. It means refusing to normalize the errors a practice can reasonably catch earlier.

That is the direction I want OnnX to help advance: fewer unnecessary handoffs, clearer accountability, and practical evidence of improvement for small and medium-sized clinics.

The best result may be surprisingly undramatic. A correct payer. A confirmed authorization. A documented service. A claim that does not return for the same avoidable reason.

No victory music. No heroic appeal. Just an afternoon returned to the practice.

The most valuable administrative work may be the work you no longer need to do.

Get involved: change the conversation

If your clinic became twice as fast at billing tomorrow, would it solve the problem—or merely process the same mistakes twice as quickly?

Comment with the one recurring billing failure you would eliminate first, without sharing patient information, and help other physicians learn from your experience.

Repost this article so another physician or clinic owner can question an expensive habit before buying a faster version of it.

References

1. Calendar Year 2027 Medicare Physician Fee Schedule Final Rule
   CMS’s November 4, 2026 fact sheet explains finalized physician payment policies. A useful starting point for reviewing affected services, effective dates, and billing workflow updates.
2. CMS Finalizes CY 2027 Physician Fee Schedule to Shift from Sick Care to Healthcare and Curb Waste
   This November 4, 2026 HHS announcement presents the administration’s policy rationale and includes Dr. Mehmet Oz’s statement quoted beneath this article’s title.
3. Calendar Year 2027 Medicare Physician Fee Schedule Final Rule—Medicare Shared Savings Program
   CMS’s November 4, 2026 summary outlines changes relevant to accountable care participation, financial incentives, and practice planning.

4. Your Medical Billing Problem May Not Be a Billing Problem: The Hidden Cost of Creating Work You Should Never Have to Do
   Explores how recurring billing failures begin upstream—and how tracing their causes can help practices reduce corrections, delays, and unnecessary administrative work.
5. We Trained Doctors to Save Lives. Then We Made Them Professional Clickers.
   Examines the connection between physician administrative burden, fragmented workflows, and technology that creates activity without necessarily improving care or operations.
6. Healthcare Has a Strange Business Model: We Pay People to Fix Mistakes We Pay Software to Create
   Challenges the assumption that faster automation always means better performance, emphasizing information quality, reliable handoffs, and prevention of avoidable billing failures.

About the author

Dr. Daniel Cham is a physician, medical consultant, and entrepreneur working at the intersection of medical technology, healthcare management, and medical billing. As founder of OnnX, he focuses on practical ways to reduce administrative burden and strengthen independent clinical practices. He writes about the operational decisions connecting patient care, information quality, and innovation. Connect with Dr. Cham on LinkedIn.

A note on scope

This article offers general education and professional commentary. It does not provide individual medical, legal, billing, or compliance advice. Requirements vary by payer, service, jurisdiction, and effective date. Consult appropriately qualified professionals before applying these ideas to specific decisions. Illustrative scenarios and calculations are identified as hypothetical and do not represent verified customer results.

Keep the conversation moving

Explore practical ideas, operational lessons, and perspectives on the decisions shaping healthcare, independent practice, and responsible innovation.

Visit the personal website and read the blog. Listen on Spotify, watch on YouTube, or follow on Substack.

Join the conversation on X, Facebook, TikTok, Instagram, and Threads.

You can also find discussions on Reddit and Quora.

Knowledge becomes progress when someone puts it to work. Start with one better question.

P.S. Visit the Featured section of my LinkedIn profile for any currently available free downloads; check each resource’s access details there.

#MedicalBilling #HealthcareAI #IndependentPhysicians #PracticeManagement #RevenueCycleManagement #DenialPrevention #PhysicianEntrepreneur #HealthcareInnovation #OnnX