AI quietly denies more claims

Insurance denials are becoming a software story. Not because a robot suddenly took over the whole system, but because more of the boring middle — intake, coding, prior authorization, length-of-stay reviews, routing — now runs through automated tools. That changes the texture of a denial. It can arrive faster, feel more final, and be harder for patients or even staff to unpack. The recent pushback is not really “AI bad.” It’s that opaque automation is getting embedded in decisions that affect care and money, while the explanation and appeal machinery still looks old. ### Where is this showing up? The clearest examples are in health insurance, especially prior authorization and post-acute care reviews. These are the checkpoints insurers use before approving treatment, rehab, or continued stays in facilities. More of that work is now digital by design, and CMS has been pushing the industry toward electronic prior authorization through new API and reporting requirements. Faster pipes can help. But faster pipes also make it easier to scale denials if the underlying logic is too aggressive. (statnews.com) ### Why are people worried now? Because the evidence stopped being hypothetical. The American Medical Association said in February 2025 that 61% of physicians worry unregulated AI and predictive tools are increasing denials of medically necessary care. A Senate investigation published in October 2024 went further, saying UnitedHealth, Humana, and CVS used technology to help deny post-acute coverage requests for Medicare Advantage patients, with roughly a quarter of such requests denied. (cms.gov) That turns a vague fear into a measurable pattern. ### What’s the UnitedHealth case about? It’s the best-known test case because it puts a specific algorithm under a microscope. Patients sued UnitedHealth over claims that an algorithmic tool was used to cut off rehab and nursing-facility coverage too early for some Medicare Advantage members. Reporting around the case described pressure on clinical staff to follow algorithmic recommendations, and in February 2025 a judge allowed the lawsuit to move forward. (ama-assn.org) That does not prove every denial was wrongful. But it does show courts think the allegations are serious enough to examine in discovery. ### Is the problem AI itself? Mostly, no. The deeper problem is black-box administration. A model can rank risk, estimate likely rehab days, or flag missing documentation. None of that is automatically improper. The catch is when those outputs become de facto decisions — especially if frontline staff cannot meaningfully override them, and patients cannot see the rationale in plain language. At that point, “automation” stops being clerical help and starts acting like policy. (statnews.com) That is the part regulators and doctors are reacting to. ### Why does opacity matter so much? Because appeals depend on knowing what you are appealing. If a denial comes from a chain of coding rules, utilization benchmarks, and prediction scores, a patient may only see a generic notice. A doctor may not know whether the issue was missing records, a plan rule, or a probability threshold buried in software. It’s like being told you failed a test without seeing the questions. You can challenge it, but you’re guessing. (ama-assn.org) CMS’s newer rules try to force more visibility through metrics and electronic exchange, but those rules do not automatically make every algorithm explainable. ### Does this stop at insurance claims? Not at all. The same design risk shows up in scheduling, intake, triage, and other administrative workflows. If an automated system decides who gets routed first, which paperwork counts as complete, or whether a request gets kicked back, then the user still experiences a denial — just earlier in the funnel. It may not look like an insurance claim rejection, but the effect can be similar: delay, confusion, and no obvious human owner of the decision. (cms.gov) That’s why the governance question is bigger than one insurer or one lawsuit. ### So what would better guardrails look like? Three basic things. Clear reasons people can actually read. A real human override, not a ceremonial one. And public reporting that shows denial rates, turnaround times, and overturn rates after appeal. CMS is already requiring some prior authorization metrics to be posted beginning in 2026, which is a start. But the bigger fix is cultural — organizations have to treat automated recommendations as inputs, not verdicts. (cms.gov) ### Bottom line AI is not quietly denying claims all by itself. Organizations are using automation to make high-volume decisions faster, and the oversight has not caught up. That gap is where people get hurt — not in the existence of software, but in software nobody can question. (statnews.com) (cms.gov)