Health Insurers Use AI to Process Claims. Why Not Help Users Pick Better Plans?

Most of the public debate about AI use in health insurance is an argument about denials. People want to know whether a model is deciding prior authorizations, how fast it can refuse a treatment, and whether a human reviewed that refusal in any meaningful way, which are all fair questions.

These questions focus on the final step in a long process; the earlier step that helps shape the outcome gets almost none of the same attention: enrollment.

Choosing a health care plan sets up the consequences down the line – a claim is adjudicated against a plan the member selected earlier. That plan sets the deductible, coinsurance, network and formulary tier. Insurers then adjudicate claims under those terms, sometimes using automated systems.

But the focus of the debate is on the wrong end: approving or denying a claim due to medical necessity. Federal data shows that medical necessity denials account for a small fraction of what goes wrong, whereas enrollment data could explain much of the rest.

Consumers need help with enrollment. They often compare health plans based on a single number, such as the monthly premium or deductible, because it’s easier to understand than the full cost of coverage. While technology can analyze and compare plan benefits at scale, predicting an individual’s health care needs and out-of-pocket costs is far more difficult.

Yet this is where better information could make the biggest difference in choosing coverage.

Denials skirt the real issue

The analysis by KFF (formerly Kaiser Family Foundation) of federal marketplace data found that insurers on HealthCare.gov denied 19% of in-network claims in 2024. Within those denials, 5% were based on medical necessity and about 9% came down to a missing prior authorization or referral.

Much of the debate over AI in health insurance focuses on medical necessity reviews. Other denials involve administrative requirements, coverage exclusions and network restrictions. While not all of these problems can be traced to plan selection, some underscore the importance of helping consumers understand their coverage before enrolling.

Under an individual coverage health reimbursement arrangement, employees select their own plan using funds provided by their employers. In a sample of 20,000 members in our 2026 enrollment data, the average member had 74 plans available. They scrolled past 31 of them and actively compared about three. Use of the cost filter rose 64% year over year while filtering on Health Savings Account (HSA) eligibility accounted for 6% of filter use.

Those shoppers were not being careless. Everything that determines real exposure sits behind terminology most people have never had explained to them, so the premium ends up doing the work that deductibles, coinsurance, and out-of-pocket maximums should be doing. Sorting by cost is a rational response to an interface where cost is the only legible variable.

The consequences show up in the same survey: 39% of employees reported delaying care because of cost while 53% said they or someone close to them had struggled to pay a medical bill while insured.

A hard-to-understand document

Federal rules already require every plan to publish a Summary of Benefits and Coverage in a standardized format, so the layout and the sections are identical for every carrier in the country. That made the documents comparable to look at without making them comparable to understand.

But it can still confound users. A plan can advertise a zero-dollar deductible and carry a high copay. The terms are accurate and the headline number is the best one on the page, so a person will choose their plan based on what they can understand. These structures are common and fully disclosed. Nothing is hidden from the buyer except what it means practically.

The industry should be more candid about which half of this problem is hard. Reading plan terms at scale is the solved half. The data is structured and machine readable, and analyzing several thousand plan designs is not technically demanding. We do that work and so do plenty of others.

The unsolved half is the person. Telling someone what a plan will really cost them requires knowing what they are likely to use. At the point of enrollment, the industry mostly does not know.

Platforms know one’s salary rather than household income. A new member has no claims history on file, because it sits with the prior carrier. You can ask people directly, which is both permitted and useful, but recency bias runs the answers. Someone coming off a hard year overestimates what the next one holds. Someone who has not seen a doctor since 2023 assumes the pattern continues.

Recorded utilization predicts far better than recall and it is the one thing nobody hands over at the moment of the decision.

Then there is the knowledge that isn’t captured: Does an insurer cover care across states? It matters if your specialist practices in another state. Whether a carrier is demanding extra residency documentation this year changes what enrollment involves for that member. That knowledge sits with licensed agents who work the market every day, and right now they catch what the model misses. This describes where the technology currently stops rather than arguing against it.

Automating enrollment

Investments typically go toward the claims system for understandable reasons. Claims are high volume, the data already exists in usable form, and the return is immediate and measurable. Enrollment happens once a year per person and the inputs are thin.

Regulation widens the gap further. Most modern AI tooling was built for commerce and marketing where nobody has to think about whether a model can touch protected health information. Doing anything with PHI (Protected Health Information) can sometimes mean building the secure plumbing yourself before you write a line of code for what you set out to build.

Faced with that cost, teams build where the payoff is clearest. We understand the math; yet the stage with the least available data may have the greatest influence on consumers’ health care costs and access to care.

None of this makes the denial debate wrong. Automated review deserves the scrutiny it is getting and the governance questions around it are real. But a faster decision about a claim is still a decision about a claim. The claim was shaped by someone who spent little time sorting plans by the only number that made sense to them. Every argument about whether a model should be permitted to deny a treatment is happening one step downstream of the question of why that treatment fell outside the plan in the first place.

Fixing that problem does not require new science. Members can be asked a few things they know the answer to, such as whether they take a maintenance medications or have surgery on the calendar, which is a better input than asking anyone to estimate a year of medical spending.

With even that much data, a plan’s cost-sharing can be run forward into a projected annual cost, and that number could sit on the plan card next to the premium. Matching named doctors and hospitals against every network available where someone lives is already routine; it’s the part that works. None of that needs a model to diagnose anyone or decide what care they should get.

Portability is harder. The best predictor of what someone will use next year is what they used last year, and that record stays with the carrier they are leaving. For a long time, there was no routine way for a member to carry it across. But that’s changing. Federal rules beginning in 2027 will mandate that certain insurers exchange claims and other health information when patients change coverage, with patient permission. But the requirements do not cover all insurers, and the exchange generally occurs after enrollment, limiting its usefulness for consumers comparing plans beforehand.

If AI is going to do something structural in health insurance, the enrollment decision deserves more attention. The opportunity is to help people understand their coverage before they need to use it.