The Appeal Button

Audio Companion

AI is being sold and adopted as a way to lower administrative burden in health insurance.

For insurers, that matters immediately. AI enters an operating system already built for claims, coding, review, denial, appeal, audit, payment, and reporting. The tool has somewhere to go.

For doctors, clinics, and hospitals, it matters too. Health-insurance companies can deny claims, require prior authorization before tests or treatment, pay slowly, dispute billing codes, and create paperwork that practices have to absorb.

Fighting those decisions is not just back-office cleanup. For many clinics, it is part of staying open: staff and systems chase payment, supply documentation, and push back when care they already provided is not paid.

An AI-powered appeal tool inside a clinic can help with that work. It can make it easier to contest denied claims and recover payment, which is one way practices keep seeing patients while the paperwork expands.

The harder question is whether the burden ever moves for the patient.

Patients can already use AI to draft appeal letters. That can help. But the useful case assumes the patient recognizes the denial as appealable, finds the records, trusts the tool, uploads the right documents, submits through the correct channel, tracks the deadline, and escalates if the answer is still no.

That chain is the burden.

The appeal letter is the artifact left behind after the burden has already landed on the person least equipped to carry it.

A patient gets a denial. Somewhere in the portal, the app, the clinic workflow, or the paper notice, there may or may not be a usable path that functions like a button:

Appeal my denial.

If the path exists, it may sit inside an insurer website the patient rarely visits, behind a password reset, a claims-detail screen that looks like a billing archive, or a menu label that does not say "appeal." It may run through pages where "submit," "message us," "request review," and "upload documents" all sound adjacent to the thing the patient is trying to do.

Finding the official doorway is already work. Reaching it may only get the patient to the starting line.

At this point, the button does not reduce the burden. It reveals the test.

At the entry point, the test is not simply whether the care was necessary. It is whether the patient has the operational capacity to find the doorway, recognize that the denial can be challenged, gather the right records, submit a coherent appeal, and keep going if the first answer is still no.

That makes the early button a filter before it becomes assistance. Patients who are comfortable with portals, paperwork, deadlines, and institutions are more likely to get through. Patients who are very sick, exhausted, unsupported, juggling work or caregiving, low on health literacy, or working in a second language are more likely to stop.

The denial then survives because the process found the patient's limit, not because the denial was necessarily right.

A real burden-reducing appeal button would not just point toward the process. It would move the work: identify and explain the denial, pull the relevant policy language and clinical records, ask the provider for missing medical-necessity support, assemble the packet, get authorization, submit through the right channel, track the clock, and escalate to external review when the rules allow.

That is burden transfer: not a better letter, but a workflow that keeps the patient from becoming a project manager before the system will reconsider its own decision.

The reason this does not exist everywhere is not that the language model is too weak.

The barrier is the system around the model.

The denial lives with the insurer. The clinical evidence lives with the provider. The plan rules may live in a policy document the patient has never read. The deadline is attached to the denial notice.

Then the procedural layer begins. Submission may mean a portal, fax number, mailing address, phone process, or delegated vendor. External review depends on plan type, state, urgency, and exhaustion of internal appeal rules.

Then the authority layer begins. Someone may need to authorize a representative, request records, or distinguish a billing denial from a medical-necessity denial, prior-authorization denial, or step-therapy dispute.

None of this is impossible. It is also not one button today.

Federal policy is moving in the right direction: CMS prior-authorization interoperability APIs, clearer denial reasons, faster response times, HIPAA access rights, and information-blocking rules. But rails are not a workflow. Some important API requirements are still being phased in. Drug prior authorizations are outside the CMS rule. HIPAA access can still take time, and access to records is not the same as explanation, strategy, submission, or follow-up.

That is the gap AI cannot cross by writing prettier paragraphs.

Some current tools and services show both the path and the limit.

Fight Health Insurance can generate appeals, explain denials, analyze policy language, point people toward state resources, and offer faxing support. Counterforce Health pairs AI appeal generation with expert support. Claimable focuses on specific treatment categories and says it can mail and fax appeals while supporting the patient through the process. Patient Advocate Foundation and Solace put human advocates around the same problem.

That is the hopeful fact: the patient side is no longer empty.

But these examples are category signals, not proof of default burden transfer. The state of the art still mostly begins after a patient, caregiver, clinic, or advocate recognizes the denial as contestable and brings the case to the tool. The tool may reduce the work. It does not yet reliably absorb the work by default.

Meaningful progress may come first in provider-portal, EMR-adjacent, or payer-provider workflows where the denied service, clinical record, order, policy rule, and payer response can be tied together. Even there, patients still need the practical capacity to maintain access, find the denial, and authorize the process. But for a narrow class of appeals, the work could begin to feel like one button.

Outside that integrated setting, the burden rises fast. The workflow may require phone calls to physician offices, hospitals, medical-record departments, insurers, and vendors; paper and online records spanning months or years; the right policy language; clinician clarification; and authorization, privacy, and representation rules.

An AI agent would need authority to act on the patient's behalf: to call about PHI, request records, and speak with an insurer. There may be answers, but they are not straightforward or imminent, and they are not the same as a workflow inside systems that already hold the clinical facts, payer response, and submission channel.

The clear metric here is the burden gap: how many people are told no, how many push back, and how often the answer changes when they do.

The target is not a perfect one-to-one appeal filed for every denial. Some denials are correct. Some are duplicates. Some claims should not be paid, especially when the diagnosis is not supported by documentation, the proposed treatment does not match the diagnosis, or the treatment may be more harmful than not providing it.

A system where every denial becomes a full dispute may be another kind of failure: an automated appeal factory where weak, duplicate, or inappropriate claims consume the review capacity that should be focused on necessary care.

The warning sign is the combination: denials are common, appeals are rare, and, in at least some categories, the appeals that do happen often change the answer. In that pattern, denials are effective in precisely the wrong way. They do not merely filter inappropriate care. They can also leave necessary or appropriate care unchallenged because no one has the stamina to challenge the decision.

Silence is not neutral. It is part of how the system works.

The public data does not give one clean picture across the whole health-insurance system. It gives pieces of the shape from different parts of the system, so the data should not be compressed into one single metric.

In ACA marketplace plans on HealthCare.gov, KFF found that consumers appealed fewer than 1% of denied in-network claims in 2024.

In Medicare Advantage prior authorization, KFF found that 11.5% of denied requests were appealed in 2024, and more than eight in 10 appealed denials were partially or fully overturned. That is the more striking example. It may mean many initial denials were too aggressive. It may also mean the appeal supplied missing documentation. Either way, many people who fought got a different answer.

KFF's consumer survey showed the patient side: most people with denied claims did not know they had appeal rights, and most did not file formal appeals.

When denials are common and appeal rates are low, despite appeals being effective in at least some categories, the door is hidden, heavy, or not worth reaching. A low appeal rate is not just evidence that people chose not to fight. It can also mean they did not know, could not find the path, ran out of stamina, or rationally accepted an accurate, low-stakes, or duplicate denial.

If AI is working for patients, the primary signal is a shrinking burden gap: more people know they can easily appeal, more denials for necessary or appropriate care are challenged, and those denials have less room to survive simply because no one has the energy to fight them.

Fewer appeals is the second-order goal, after the system learns that bad denials will no longer disappear into exhaustion and starts making cleaner first decisions.

Gold-carding (exempting clinicians or practices with strong approval records from some prior-authorization requirements), real-time authorization, and prior-authorization reduction still matter. These reforms are probably where the best version of the system eventually has to go. But they are not the patient's immediate problem when a denial letter arrives.

The immediate problem is the assigned task: understand this, gather evidence, translate need into process language, hit the deadline, track the answer, and keep going if the first answer is no.

Collapse that burden, and downstream reforms become easier to imagine. If contestable denials can be challenged with much less patient effort, weak or under-explained denials lose one quiet advantage.

If clean cases are appealed automatically and overturned predictably, pressure moves upstream: approve them earlier, exempt clinicians whose requests are almost always approved, narrow the code lists, and explain denials clearly enough that a machine, a clinician, a patient, and an external reviewer can all see the same dispute.

The appeal button is not the final reform.

It is the lever that changes what the system can get away with before the final reform arrives.

So yes, AI lowering administrative burden for insurers and physicians matters.

It matters because one party's efficiency can become another party's homework. One institution's throughput can become one patient's delay. One side's automation can make the system look modern while the person at the edge still carries the same stack of recognition, evidence, translation, deadline, and stamina burdens.

The test is not whether AI can produce a better appeal letter.

The test is whether the person with the least power has less to discover, less to gather, less to prove, less to track, and less time spent waiting while the institution decides whether the care counts.

The appeal matters.

The burden-reducing button matters more.

Not because buttons are magic.

Because a real button means the burden finally moved.

What this is: an Explainer about AI, health-insurance administrative burden, and the gap between being denied and being practically able to contest a denial. It is not legal advice, medical advice, insurance advice, patient advocacy instructions, policy reporting, or a product forecast.

Confidence: Medium on the burden-gap frame. Stronger on the appeal-rate / awareness gap supported by KFF; medium-low on near-term feasibility because the rails exist in pieces but the cross-system workflow is not broadly implemented.

What would change our mind: evidence now or over the next 12-24 months that true patient-side appeal initiation is embedded at scale; low appeal rates among contestable adverse determinations mostly reflecting accurate, low-stakes, duplicate, or patient-accepted denials rather than friction; appeal rates rising while overturn rates fall because weak or under-documented denials are resolved earlier; or API/access rules producing default workflows instead of better data availability.

Process transparency: AI tools draft and critique; the human author owns final wording, published claims, and errors. Process review is not evidence that the claims are true; it is only a description of how the draft was prepared.

Sources and anchors