Claims Denied by Insurers

Getting a hospital or practice's claims paid on first submission, instead of reworking, appealing, or writing off claims the insurer refused, most of which could have been prevented before they were sent.

AI is suitable with conditions

Hospitals and practices hold years of submitted claims and the insurer's response to each, which is the labelled history a model can learn from, and in a retrospective study a language model trained on one health system's clinical notes predicted at admission which claims would be denied with an AUC of 87%. It is conditional rather than strong because much of the problem sits in process rather than prediction. One vendor's analyses of hospital claims put roughly 40 to 50% of denials at the front end, where eligibility, authorisation, and claim data checks apply, and judge most denials potentially avoidable, although many of those need judgement rather than a field check. Practices that reduced denials credit front-desk training, denial task forces, and coding training rather than models. A model is worth adding only once those checks and processes are in place, as a way to rank the remaining risky claims for review, and it is limited by insurer reason codes that are often too vague to learn a cause from.

Problem Statement

A hospital or practice treats a patient, sends the insurer a claim, and waits to learn whether it will be paid.

A growing share come back denied, and each denial has to be corrected and resubmitted, appealed, or written off.

The market for denial software presents this as a prediction problem.

The sources behind this playbook point first at process: a large share of denials start at a step before the claim was sent, such as an eligibility check or a prior authorisation, and most are judged potentially avoidable.

Business Context

Hospital claims data analysed by Change Healthcare, a revenue cycle vendor, put the initial denial rate at around 12% in the first half of 2022, up from 9% in 2016.

The same analysis found registration and eligibility to be the largest single cause, and judged around 82% of denials potentially avoidable.

It split that figure into 31% it called avoidable and 51% it called situationally avoidable, so a potentially avoidable denial is not always one a field check would have caught.

Within registration and eligibility, most denials concerned coordination of benefits between insurers and benefit maximums rather than a missing or inactive plan.

Its 2020 analysis traced roughly half of all denials, and every denial it classed as unrecoverable, to the front end of the revenue cycle.

Many practice leaders report a rise too: in a 2024 poll by the Medical Group Management Association, 60% said their denial rate had risen against the same period a year earlier.

Denial rates differ sharply by insurer type.

A study of physician remittance data found fee-for-service Medicaid denied claims at a rate 17.8 percentage points higher than fee-for-service Medicare, and estimated the revenue challenged across the sector at $11 billion a year on conservative assumptions, possibly as high as $54 billion.

A later study by the same research group estimated that physicians lose 18% of Medicaid revenue to billing problems, against 4.7% for Medicare and 2.4% for commercial insurers, and found that physicians accept fewer Medicaid patients where billing is harder.

Denials are therefore a question of access to care as well as of revenue.

Why This Is Hard

The first difficulty is that the cause and the symptom sit in different teams.

Patient access staff at the front desk capture the insurance details and track authorisations, coders and clinical documentation staff shape the claim, and a separate denials team sees the insurer’s response after the claim has been adjudicated.

When the denials team corrects and resubmits a claim, the error that caused it goes on happening at the front desk unless somebody routes it back.

The second difficulty is that insurer rules are numerous and keep changing.

Each insurer, and often each plan, has its own list of services that need prior authorisation, its own coverage policies, and its own filing deadlines, and practitioners describe keeping pre-submission edits current as continuous work.

The third difficulty is that the label is poor.

Insurers return claim adjustment reason codes with each denial, but the codes are often too general to identify the real cause.

In federal data on marketplace plans for 2024, analysed by KFF, the largest category of in-network denial reason was other, at 36%, while lack of prior authorisation or referral was 9% and medical necessity 5%.

Any model or report that treats reason codes as causes inherits that vagueness.

The fourth difficulty is that not every denial is the provider’s mistake.

The HHS Office of Inspector General found that 13% of prior authorisation denials and 18% of payment denials it sampled from Medicare Advantage insurers met Medicare’s coverage rules, and attributed most of the wrongful payment denials to human error in manual review or to insurer systems that were not programmed correctly.

A denial programme that only looks inward will miss denials that should simply be appealed.

How The Solutions Fit Together

Eligibility, authorisation, and claim edit checks before submission are the floor.

They target the front-end and claim data errors that make up a large share of denials, they can be explained to any member of staff, and they are what practitioners describe as their first remedies.

Denial work queues sorted by root cause run beside them, routing each denial to staff who know that insurer and sending a recurring cause back to the team that owns it.

This is what turns individual fixes into fewer denials next month, and it is how the pre-submission checks learn which rules are missing.

A denial risk score supports both once they are in place, ranking the claims that passed every check by how likely they are to be denied, so billers review the riskiest ones before submission.

Prediction from admission notes is a discovery tool for large health systems, flagging encounters early so utilisation review and clinical documentation staff can check the authorisation and documentation the insurer will look for.

That use is this playbook’s judgement, because the published denial model was tested only on past encounters and nobody has shown that acting on its flags prevents denials.

None of these decides on its own.

Each tells a person which claim or encounter to look at, and that person decides what is corrected, resubmitted, documented, or appealed.

Implementation Considerations

Measure the starting point by root cause, not by reason code alone, by sampling denials and tracing each to the step that caused it.

Fix eligibility, authorisation, and missing or invalid claim data first, because in the vendor’s analyses they are the largest front-end and back-end categories and none of them needs a model.

Keep insurer rules in one maintained place that the scheduling, registration, and claim edit steps all read, so a rule change is made once.

Store each claim exactly as submitted alongside the insurer’s response, because a model or a root-cause report built on claims as later corrected will learn from data that never reached the insurer.

Treat a change in an insurer’s authorisation list or coverage policy as a break in the history, and retrain or re-baseline any model across it.

Keep any model’s output to a ranked review list, sized to the number of claims billers can actually review each day.

ROI Considerations

The return comes from three places, and they should be counted separately.

Revenue that is no longer lost, from denials that are prevented or that are now worked before their appeal deadline.

Staff time recovered from rework, which only becomes a saving when prevention reduces the volume of denials rather than speeding up the work on them.

Faster cash, from claims paid on first submission rather than after a resubmission cycle.

A process improvement study at a pediatric hospital revised the emergency department registration form, cut missing or incomplete insurance fields by 67%, and reports that this helped reduce denials, so a change to a form belongs on the list of options before any software is bought.

Security, Privacy, and Regulation

Every data set in this problem is protected health information.

The HIPAA Privacy Rule permits a provider to use it for payment, which the rule’s definitions say includes billing, claims management, and determinations of eligibility or coverage.

Passing it to an outside vendor brings further HIPAA obligations, and this playbook’s advice is to send any vendor only the fields it needs.

Clinical notes are the most sensitive input, so a model that reads them should run inside the provider’s own controlled environment.

Since January 2026, CMS’s interoperability and prior authorisation rule has required Medicare Advantage, Medicaid, and CHIP payers to decide prior authorisation requests within 72 hours for urgent requests and seven calendar days for standard ones, and the payers it covers to give a specific reason for every prior authorisation denial.

The same rule requires those payers to publish prior authorisation metrics each year and to offer a prior authorisation API from January 2027, although the decision timeframes and API do not apply to drugs.

A provider should capture those specific reasons and decision dates as data, since they are better labels than the general reason codes most claim denials carry.

The insurer side is changing too.

CMS’s WISeR model, running from January 2026 to the end of 2031 in New Jersey, Ohio, Oklahoma, Texas, Arizona, and Washington, uses technology-assisted prior authorisation review for selected services in traditional Medicare, with every recommendation for non-payment determined by licensed clinicians and participants paid a share of the spending their reviews avert.

Where The Evidence Disagrees

Change Healthcare’s 2020 and 2022 analyses judge most denials potentially avoidable, but no source here measures how much a model adds once the deterministic checks are in place.

The vendor indexes list too little investment in analytics and AI among the causes of denials, while the practices in the 2024 MGMA poll that reduced denials credited training and denial task forces.

The strongest published model evidence comes from one health system, and its structured comparison lacked several claim fields, including the prior authorisation number, so it is not a test against well-maintained rules.

This playbook’s position, which is a judgement rather than a sourced finding, is that a model should earn its place against a complete set of pre-submission checks on the same claims.

The size of the avoidable share also varies between analyses, and the figures most often quoted come from revenue cycle vendors, so this playbook states it as most denials rather than as a single number.

The trend is not uniform either: hospital claims in the vendor’s data show the initial denial rate rising from 9% to around 12%, while the same 2024 MGMA article reports a single-specialty practice benchmark of 8% denied on first submission, unchanged since 2019.

The sources also point in different directions on insurers’ use of automation, with CMS expanding technology-assisted prior authorisation review in WISeR while the Inspector General has documented insurer denials that met coverage rules, and none of the sources measures what automated review has done to denial accuracy.

Symptoms

The share of claims denied on first submission has risen at many hospitals, and the denials team falls further behind its work queue.
Denials cite eligibility, missing prior authorisation, or missing claim data for patients whose details the front desk believes it checked.
Insurers ask for more medical records and question medical necessity more often, so clinically valid care is held up in documentation requests.
Cash arrives later and days in accounts receivable grow, while a backlog of small denials is quietly written off because nobody has time to work them.
The same denial reason recurs from the same insurer every month, because the denials team fixes each claim but nobody fixes the step that caused it.

Root causes

Registration and eligibility denials are the largest single category in one vendor's hospital claims data, and most of them involve coordination of benefits between insurers or benefit maximums rather than simply an inactive plan.
Prior authorisation rules differ by insurer, plan, and service and change often, and authorisation denials far more often involve an authorisation that is invalid, does not match what was delivered, or was refused than a service delivered with no authorisation at all.
Claims leave with missing or invalid data, or with coding and modifier errors, because pre-submission checks do not reflect each insurer's current rules.
Insurers have increased clinical validation of claims, asking for medical records and challenging medical necessity after the care was given.
The work is split between teams: a large share of denials start at the front end, but the back-end denials team is the one that sees them, so the cause is never fed back.
Denial reason codes returned by insurers are often vague, and in public insurer data the largest category of denial reason is simply other, so root-cause analysis starts from poor information.

Business impact

Revenue lost outright on denials that are never reworked or that miss the insurer's appeal deadline.
Staff cost of reworking and appealing claims, part of it spent on denials that a check before submission would have prevented.
Delayed cash and higher days in accounts receivable.
Reduced access to care, since research finds physicians accept fewer Medicaid patients in states where billing Medicaid is harder.
Patients billed for care their insurer refused, or delayed while an authorisation dispute is resolved.

Traditional approaches

Real-time eligibility checks against the insurer before or at registration.
Authorisation tracking by schedulers and patient access staff, using each insurer's published list of services that need approval.
Claim scrubbers and practice management system edits that check codes, modifiers, and required fields before submission.
Denials work queues sorted by insurer and reason, worked by staff who know each insurer's rules.
Process improvement methods such as Lean Six Sigma on the front-end workflow, and regular meetings with insurer representatives about recurring denials.

Solution approaches

Each of these is written for this problem specifically, not borrowed from a generic catalogue.

Recommendation

Start with deterministic checks before submission, covering eligibility, prior authorisation against each insurer's current list, and claim edits, and feed every denial back to the team that owns its root cause through work queues sorted by insurer and reason. Add a denial risk score on the claims that pass those checks only when the remaining denials are too varied for rules, and use it to decide which claims a biller reviews before submission, not to change the claim. Prediction from admission notes is for large health systems with their own clinical language model, as an early warning to utilisation review staff. Billers decide what is corrected and resubmitted, clinicians and clinical documentation staff decide every medical necessity and documentation response, and no model writes or sends an appeal on its own.

Data requirements

Submitted claims high sensitivity: Every claim as submitted, with insurer, plan, procedure and diagnosis codes, modifiers, place of service, billing and rendering provider, charges, and the authorisation number if any, stored as it was at submission.
Remittance and denial history high sensitivity: The insurer's response to each claim line, with payment amount, claim adjustment reason codes, remark codes, and the outcome of any resubmission or appeal, used as the label for every model and every root-cause report.
Registration and eligibility records high sensitivity: Insurance details captured at registration, the eligibility responses received, and who captured them and when, needed to trace a denial back to a front-end step.
Prior authorisation records high sensitivity: Authorisation requests, approvals, denials, approved units and dates, and the service they cover, matched to the scheduled and delivered service.
Insurer rules and contracts medium sensitivity: Each insurer's current list of services requiring authorisation, coverage policies, filing deadlines, and contracted rates, maintained by the revenue cycle team.
Clinical documentation high sensitivity: Admission, history and physical, and discharge notes and orders, needed for medical necessity responses and for any model that reads notes. The most sensitive data in the problem.

Evaluation metrics

Initial denial rate by insurer and by root-cause category, measured on submitted claims, and tracked against the same insurer's rate in the prior period.
Share of denials that are preventable by a front-end or pre-submission check, and whether it falls after each fix.
Net revenue recovered on denials, and the value written off because a denial was never worked or missed its appeal deadline.
Days in accounts receivable and cost to collect, so that a gain in first-pass payment is not bought with staff time elsewhere.
For any risk score, the share of flagged claims that a biller found and corrected an error on before submission, at the number of claims billers can actually review.

Human oversight

Patient access staff confirm eligibility and authorisation before the service, and billers decide every correction and resubmission. Clinicians and clinical documentation staff decide every response about medical necessity or clinical documentation. A risk score or a language model only ranks claims for review or drafts material for a person to check, and it never changes a code, submits a claim, or sends an appeal on its own.

Common pitfalls

Buying a prediction tool before fixing eligibility and authorisation checks, so the model learns to predict the front-end errors a deterministic check would have stopped.
Training a model on reason codes as if they were causes, when insurers return vague codes and a large share fall into an other category.
Measuring success as denials overturned on appeal, which rewards working denials rather than preventing them.
Letting a model or an automated rule change diagnosis or procedure codes to avoid a denial, which turns a billing tool into a coding compliance risk.
Training on remittance history from a period before a major insurer rule change, such as a new authorisation list, so the model scores current claims against rules that no longer apply.
Letting the denials team rework the same denial every month without a route back to the front-end or coding team that caused it.

When not to use AI

Eligibility verification, prior authorisation tracking, and claim edits are incomplete or manual, so the deterministic checks will remove more denials than any model.
The organisation is a small practice with too few denials per insurer for a model to learn from, where a clearinghouse scrubber and a weekly denial review cover most of the value.
The question is whether care was medically necessary, which is a clinical judgement for a clinician, not a prediction.
Remittance data does not record the reason codes or the outcome of resubmissions and appeals, so there is no reliable label to train on.
A recent insurer rule change has made the historical denial pattern a poor guide to the current one.

Sources

Agent context

A distilled brief for an AI agent or an engineer scoping this work. It carries the constraints and pitfalls that change an implementation decision, not the whole page.

Get the agent context as Markdown

Objective

Reduce the initial denial rate on a provider's claims by stopping preventable errors before submission and routing every denial to the team that owns its root cause, without slowing claim release or changing clinical coding to suit an insurer.

Constraints

All claims, remittance, registration, and clinical data is protected health information. Use it only for payment and operations, keep it inside the provider's controlled environment, and pass it to any outside service only under a business associate agreement and only the fields that service needs.
Build eligibility, prior authorisation, and claim edit checks before any model. Do not train a model to predict errors that a deterministic check can stop.
Never change a diagnosis or procedure code, add clinical documentation, submit a claim, or send an appeal automatically. A biller, coder, or clinician decides each one.
Medical necessity is a clinical judgement. A model may flag an encounter for review but must not be presented as deciding whether care was necessary.
Store and train on each claim as it was submitted, not as later corrected, and do not use fields filled in after submission as features.
Treat an insurer's change to its authorisation list or coverage policy as a break in the history. Do not train across it without checking performance on claims after the change.
Do not treat claim adjustment reason codes as root causes. Record the root cause assigned by the person who worked the denial.
Size any ranked review list to the number of claims billers or reviewers can actually work each day.

Inputs

Registration records with insurer, plan, member ID, and the eligibility response and its date.
Each insurer's current list of services requiring prior authorisation, coverage policies, and filing and appeal deadlines.
Authorisation records with number, approved service, units, and date range.
Claims as submitted, with procedure and diagnosis codes, modifiers, place of service, providers, charges, and authorisation number.
Remittance per claim line, with claim adjustment reason codes, remark codes, amounts, and the outcome of any resubmission or appeal.
For large health systems only, admission and discharge notes linked to inpatient encounters.

Outputs

Holds on claims that fail an eligibility, authorisation, or claim edit check, each with the rule that fired and what to fix.
Denial work queues sorted by insurer, value, and appeal deadline, with a recorded root cause and owning team per denial.
A monthly root-cause report per insurer for the patient access, coding, and documentation leads.
Where justified, a ranked list of claims most likely to be denied among those that passed every check, sized to biller capacity.

Success criteria

Initial denial rate falls for eligibility, authorisation, and missing claim data denials, per insurer, against the starting point.
The value written off because a denial was never worked or missed its appeal deadline falls.
Days from service to claim submission do not rise because of holds.
Recurring root causes fall after they are fed back to the owning team.
Any risk score finds correctable errors on a meaningful share of the claims it flags, at real biller capacity, measured on claims after the most recent insurer rule change.

Last updated 2026-10-02