AI Claims Review: Catch E/M Downcoding Inside Paid Claims

The claim says paid. The deposit arrives. The denial queue stays quiet. But the payer has reimbursed a level-four office visit as a level three, and nobody has compared the submitted service with the adjudicated result. Repeat that across a busy practice and the problem is not simply a lower collection rate. It is a payment exception that the usual denial dashboard may never show.

As practices close September 2026 and prepare their next operating budgets, AI-assisted claims review deserves attention beyond the familiar debate over prior authorization. The practical question is what happens after an evaluation and management claim reaches the payer: Was its code accepted, reduced, bundled, or priced differently—and can the practice explain why? Automated edits, statistical models, and clinical reviewers can produce similar-looking results. An adjustment alone does not reveal which one made the decision.

The response should be neither a blanket accusation against artificial intelligence nor an automatic appeal of every reduction. Practices need a narrow, repeatable audit of paid E/M claims that connects coding support, payer policy, contract terms, and actual reimbursement. That work starts by separating a legitimate correction from a payment decision that deserves a challenge.

1. A paid claim can still contain an adverse payment decision

Downcoding occurs when a payer adjudicates a service at a lower code level than the practice submitted. For office and other outpatient E/M services, that may mean payment corresponding to a lower-level visit. However, a smaller-than-expected payment is not, by itself, proof of downcoding. Deductibles, coinsurance, contractual adjustments, multiple-service rules, and an incorrect fee schedule can all produce a shortfall.

A payer may also leave the submitted code visible while reducing the allowed amount under a payment policy. That is an underpayment question, but it is not necessarily a code substitution. The distinction matters: a coding appeal, a pricing dispute, and a contract escalation may go to different departments and follow different deadlines.

Start with three values: the submitted code, the expected allowed amount under the applicable agreement, and the adjudicated allowed amount. Then examine the explanation of payment, adjustment reasons, and any separate review notice. Do not compare charges with the check and call the difference a loss.

Electronic remittance data will not always explain the decision completely. An 835 transaction may need to be paired with the payer portal, correspondence, or a claim inquiry. When the explanation is missing, record the case as unresolved—not as confirmed AI downcoding. Accurate classification makes the eventual dispute stronger.

2. Ask what changed, not just whether AI was involved

The most useful question is often not “Did you use AI?” It is “What rule produced this result, and what evidence was considered?” A conventional rules engine can apply an inappropriate edit consistently. A sophisticated model can flag a claim for a qualified reviewer who reaches a defensible conclusion. The technology label does not settle the coding issue.

Payer communications sometimes describe broad payment-integrity initiatives without explaining their effect on individual claims. Practices should preserve the actual policy version, its effective date, the affected product, and the notice received. A policy for a commercial product should not automatically be treated as applicable to Medicare Advantage, Medicaid managed care, or another employer plan administered by the same company.

When a reduction is unexplained, send a focused inquiry that can support the next step. Avoid unsupported allegations about an algorithm's design or training data; those rarely help the payer identify and reverse a specific claim decision.

• Was the submitted procedure code changed, or was its allowance reduced without a code change?
• Was the determination based on claim data, submitted records, utilization patterns, or another stated criterion?
• Which payment policy, coding standard, or contract provision supports the determination?
• What records were reviewed, and what finding was made about this encounter?
• Which reconsideration or appeal route applies, and what is its filing deadline?

3. Test office-visit coding against the actual CPT framework

For office and other outpatient E/M visits, the central coding question is generally whether the selected level is supported by medical decision making or qualifying total physician or other qualified health care professional time on the date of the encounter. Code-specific requirements still matter; not every service in the broader E/M family follows an identical selection method.

When medical decision making determines the level, evaluate the problems addressed, the amount and complexity of data reviewed and analyzed, and the risk of patient management. Two of those three elements generally must meet or exceed the requirements for the selected level. A lengthy diagnosis list does not establish that each condition was addressed, and prescription drug management alone does not automatically establish the entire MDM level.

When time determines the level, verify the applicable threshold and count only qualifying activities. Exclude clinical staff time and time associated with separately reported services. The record should support the reported total; a generic statement that the visit was complex does not substitute for time documentation.

An AI-generated note can be beautifully organized and still fail this test. Repeated histories, imported laboratory results, and a long assessment do not necessarily show analysis or management. Conversely, a concise note may support the billed level well. The audit should test the encounter against coding requirements, not reward length or penalize brevity.

4. Separate coding validity from payer payment authority

A code can be clinically and technically supported while its payment remains disputed under a payer policy. Equally, an expected contract allowance does not rescue a code that the documentation fails to support. Treat these as two separate questions: Was the service coded correctly, and did the payer adjudicate that correctly coded service according to the governing terms?

The AMA's CPT instructions provide the coding framework. Medicare coverage and payment requirements add rules for Medicare claims. Commercial contracts, incorporated reimbursement policies, and applicable law can introduce other considerations. No single label—“industry standard,” “CMS compliant,” or “proprietary edit”—answers every question across every line of business.

Do not assume that CMS requirements governing prior authorization create a universal appeal deadline or human-review right for every post-service commercial payment dispute. Prior authorization, medical necessity, code validation, and contracted reimbursement are related, but they are not interchangeable regulatory categories.

For recurring reductions, have contracting staff identify the relevant reimbursement and dispute provisions. Determine whether the payer incorporated the policy into the agreement and whether any required notice was provided. If a claims representative cites a policy that took effect after the disputed service date, preserve that discrepancy. Material questions about contract enforceability or state requirements belong with qualified counsel, not an improvised argument in a claim note.

5. Build an exception report that finds silent reductions

A useful report starts with a bounded population: one payer product, office E/M services, and a defined service-date range. Include adjudicated claims rather than only denied claims. Compare expected and actual allowed amounts, while keeping patient responsibility separate from payer liability. Otherwise, a deductible-heavy population can look like a downcoding campaign when the contracted allowance was applied correctly.

Use a fee schedule that matches the provider, product, location where relevant, and date of service. If the contract uses a percentage of a reference schedule, confirm the specified schedule and version rather than assuming the current Medicare amount applies. An outdated expected-payment table can manufacture hundreds of false exceptions.

A practice using medical billing services should require this comparison explicitly. Posting remittances and working denials are not substitutes for validating paid claims. Agree on the reporting fields, ownership of policy research, and the evidence needed before an exception is closed.

Keep the report small enough to investigate. The first objective is to identify a reproducible pattern, not to generate a spreadsheet nobody can finish.

• Claim and encounter identifiers, dates of service, payer product, and rendering clinician.
• Submitted E/M code and modifiers, plus any identifiable adjudicated code change.
• Expected allowance, actual allowance, patient responsibility, and calculated variance.
• Adjustment explanations, policy references, and records-request history.
• Dispute deadline, assigned owner, current status, and final resolution.

6. Validate the pattern before launching bulk appeals

Select a manageable sample of flagged encounters across clinicians, code levels, and dates. Have a qualified reviewer examine both the documentation and the payment explanation. Where practical, the reviewer should make an independent coding assessment before seeing the desired financial outcome. That reduces pressure to justify the originally billed level simply because money is at stake.

Classify each sampled case as supported coding with an unexplained reduction, supported coding with an identified payment-policy dispute, unsupported coding, or a noncoding payment issue. Add an insufficient-information category rather than forcing an uncertain case into the appeal pile. If records are missing, retrieve them before drawing conclusions.

Consider a hypothetical level-four established-patient visit. The practice selected the level using time, and the signed record supports qualifying total time under the applicable CPT requirements. The payer's explanation addresses only MDM. The appeal should identify the permitted time-based selection method and the supporting documentation—not merely argue that the patient was complicated.

The reverse finding matters too. If a practice selected the level using MDM but the record supports only one element at the required level, the right response may be coding education and claim correction rather than an appeal. Preserve the original record. Any legitimate amendment should follow documentation policy and retain its date and authorship; never retrofit the chart to manufacture support.

7. Make the appeal easy to evaluate and hard to misroute

An effective appeal names the disputed determination immediately. Identify the submitted service, what the payer changed, the requested correction, and the basis for that correction. “Please reprocess” is not enough. Neither is a lengthy clinical narrative that leaves the reviewer to infer the coding argument.

Submit the relevant encounter documentation through the payer's accepted secure channel. Point to the specific portions supporting MDM or time, and identify the coding guidance, payer policy, or contract term at issue. If the payer reviewed incomplete records, explain what was absent and supply it. Keep the argument factual and encounter-specific.

Know whether the payer expects a corrected claim, a reconsideration, a provider payment dispute, or another appeal type. Those routes are not interchangeable. Sending an unchanged duplicate claim can create extra work without preserving dispute rights. A representative's promise to “look into it” should not be assumed to suspend a filing deadline.

When the decision remains unexplained, request the applicable rationale and clinical or coding review through the available process. Track proof of submission and the response. Escalate a documented pattern through the payer's provider relations or contracting channels, but continue protecting individual claim deadlines unless the payer confirms a different arrangement in writing.

8. Put guardrails around your own automation

Practices should apply the same skepticism to their own tools that they apply to payer tools. Software can cluster reductions, retrieve policy references, and draft an appeal outline. It should not independently invent a clinical rationale, insert undocumented management decisions, or select the highest-paying code that seems plausible.

AI-assisted documentation and appeal products require an appropriate privacy and security review. Determine what protected health information they receive, where it goes, whether a business associate agreement is required and in place, and how the vendor handles retention and secondary use. A publicly available chatbot is not automatically an acceptable destination for patient records.

If a medical billing services partner uses automation, ask who validates coding arguments, who approves submissions, and how erroneous drafts are intercepted. Contracts and workflows should make those responsibilities visible. “Human in the loop” is meaningful only if the human has the information, competence, and authority to stop the work.

Avoid solving the dispute by transferring the reduction to the patient. Whether an amount is collectible depends on the contract, benefit terms, applicable law, and valid patient responsibility—not simply on the payer's refusal to pay the billed level. An internal adjustment should not become a patient statement by default.

9. Measure recovered revenue—and what it costs to recover

The first operating cycle should answer three questions: How often is the practice being reduced, how often are those reductions defensible, and what does resolution cost? Track confirmed reductions against an appropriate denominator, such as adjudicated office E/M claims for the same product. A raw case count cannot distinguish a worsening policy problem from ordinary growth in visit volume.

Measure dollars restored, appeal outcomes, time to resolution, and staff effort. Separate coding wins from contract-pricing corrections. A high overturn rate may justify a payer-level escalation, while a high rate of unsupported original coding calls for targeted education. Neither result should disappear into a single denial-management percentage.

Set financial priorities without abandoning accuracy. Repeated modest reductions can justify a systemic challenge even when each individual appeal is expensive. Conversely, pursuing every small variance without first validating the fee schedule can consume more labor than it returns. Preserve filing rights while deciding which cases need clinical review and which can follow an established workflow.

The goal is not to prove that every automated payer decision is wrong. It is to make every material payment difference explainable. When a practice can show the submitted service, the coding support, the governing payment terms, and the exact shortfall, it turns an argument about AI into a concrete reimbursement dispute. That is the evidence a payer can act on—and the control a paid-claims dashboard should have been providing all along.