As open enrollment approaches, millions of consumers are already asking AI tools which health plan to pick. A new benchmark shows those recommendations are getting more accurate—with implications for risk pools and the rules governing insurance advice.
By Khris Dai
“I cannot decide which plan to choose. I am 33, don’t really have any health issues. I like to see a chiropractor, get annual checkups, and occasionally fill a prescription. Which plan should I choose?” or “I was relatively healthy until this year, getting pneumonia twice and had an ER visit. Now my doctor wants to run tests when I get healthy. I don’t currently take scheduled prescriptions, but when I get them prescribed, generics only cost me $10. Which plan do you recommend?”
As open enrollment approaches, consumers are increasingly asking artificial intelligence (AI) tools sophisticated questions like these. OpenAI reports between 1.6 million and 1.9 million ChatGPT messages a week about health insurance, including comparing plans, understanding prices and coverage, and navigating claims and billing.[i] AI tools can work through the decision with the consumer: They can translate deductibles and coinsurance, estimate annual spending, compare premiums and out-of-pocket costs, test different utilization scenarios, and explain the tradeoffs conversationally based on the consumer’s own situation.
That capability could be particularly consequential during open enrollment, when consumers are choosing coverage for the year ahead. When consumers have an insurance advisor in their pocket, actuaries need to understand how quickly that capability is improving, what makes the recommendations more accurate, and how better matching could affect risk pools. The broader implications for plan information and emerging AI policies also deserve attention.
A Threshold Worth Noticing
Mainstream generative AI arrived in late 2022 with the release of ChatGPT. By late 2025, reasoning models had become widely available on consumer chatbots, making them better at complex analytical tasks.
That progression matters for health plan selection, where consumers must weigh multiple variables at once. To understand how well AI handles that complexity, I recently benchmarked models from two leading AI providers, OpenAI and Anthropic. The study[ii] evaluated five ACA-compliant plans across 12 claims-anchored consumer profiles, generating 1,564 responses across four progressively richer levels of consumer-provided health information. Recommendations were then scored against a claims-based pricing engine. The study focused on what objective the models appeared to optimize, how accurately they estimated costs and recommended plans, and how consistent those recommendations were when the same question was asked repeatedly.
A recommendation was considered near-optimal if the selected plan was within $250 per year of the benchmark-optimal plan. The best current models available to free-tier consumers selected a near-optimal plan 57% of the time, compared to 36% for the prior-generation model. Their mean regret, or the average additional annual cost (premium plus out-of-pocket) relative to the benchmark-optimal plan, was also substantially lower: $836 versus $1,890.
Reasoning mattered as well. The same model selected a near-optimal plan 56% of the time with high reasoning effort, versus 33% without it, while mean regret was $940 versus $1,780.
Insurance Shopping Is Becoming Conversational
Traditional enrollment platforms collect a limited set of structured information to show consumers available plans and premiums. HealthCare.gov goes a step further, asking consumers to characterize their expected health care use as low, medium, or high, and using that information to estimate out-of-pocket costs. AI changes the interaction again: Consumers can describe conditions, medications, recent care, and expected needs in natural language. This shift is already reaching the enrollment process: KFF notes that AI tools are increasingly being promoted to both companies and individuals to help consumers navigate health insurance.[iii]
The study was designed to mirror this progression. It tested four progressively richer levels of information: (1) demographics only; (2) expected utilization, similar to HealthCare.gov’s low/medium/high categories; (3) conditions and medications; and (4) more detailed clinical information, including expected visits, care episodes, and named medications. The last two levels reflect the additional information consumers can disclose conversationally to AI, with the fourth testing whether more detail improves the recommendation.
More information helped, but only to a point. Adding conditions and medications increased correct recommendations to 61% across four current models, with the strongest model reaching 73%. Adding more detailed clinical information produced no further improvement and, in some cases, reduced accuracy. For example, among medium utilizers (those needing regular care), accuracy fell from 72% to 28%, and models tended to underestimate out-of-pocket costs.
One reason more detailed clinical information does not necessarily improve accuracy is that models must translate how consumers describe their care into the full mix of services that drives actual spending. A consumer may say, for example, that they see a specialist 10 times, while the resulting claims may also include laboratory tests, imaging, and other separately billed services. Models often simplify that mapping and rely on generic price assumptions. By contrast, when the study provided the AI models with market-level prices for the expected services, the strongest model’s accuracy increased to 90%.
As models get better at arithmetic and reasoning through health information, the harder part becomes translating a consumer’s story into an expected mix of services and attaching realistic prices to those services—tasks that depend on reliable clinical, utilization, and pricing data.
Better Individual Choices Can Change Risk Pools
The most advanced models overwhelmingly treat minimizing expected annual cost as the primary decision objective—80% to 86% of primary recommendations in the study were framed that way. Protection against an unexpected high-cost year was usually secondary.
For an individual consumer, minimizing expected annual cost can make sense. But it is not a universal definition of a better decision. A family with limited liquidity may prefer a higher premium and lower exposure to a bad year; another consumer may value network access, continuity of care, or account tax treatment.
Evidence from health-plan choice suggests, however, that consumers place relatively little value on incremental risk protection. Handel, Kolstad, and Spinnewijn estimate relatively low risk premia for choosing a PPO over a high-deductible health plan (HDHP), reflecting both low estimated risk aversion and the downside protection already provided by the HDHP’s out-of-pocket maximum.[iv]
As it becomes easier for consumers to choose based on their expected health care costs, that limited willingness to pay for additional coverage can make healthier consumers more likely to move toward the lower-cost plan, while higher utilizers gravitate toward richer coverage. This can lower total cost for individual consumers but also reduce cross-subsidies among plans and change the composition of the risk pool for the insurer—a classic adverse-selection pattern.
Research on health-plan choice[v] has long documented that information frictions, complexity, and switching costs keep consumers from selecting the plan that best fits them, thereby mitigating adverse selection. Now AI can reduce some of those frictions at scale by helping consumers interpret plan information, incorporate their expected health care needs, and compare options more easily.
That potential shift in consumer choice also raises the question of how insurers and other market participants respond. In the ACA market, risk adjustment is designed to compensate plans that attract higher-risk enrollees and can mitigate some of the financial effects of adverse selection. Handel, Kolstad, and Spinnewijn similarly find that risk adjustment can substantially offset the adverse-selection effects of reducing choice frictions.
Employer-sponsored markets have different levers. Employers determine both the plans offered and how much employees pay for each option, so contribution structures and plan options may play a greater role in shaping incentives to select richer or leaner coverage.
Consumer AI may not only improve plan selection. At sufficient scale, it could also change the market dynamics that plan sponsors, insurers, and actuaries will need to account for.
Policy Landscape and Implications
Insurance regulators are developing frameworks for the use of AI. The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in 2023. As of April 2026, 25 jurisdictions had adopted the bulletin, while four others had insurance-specific AI regulation or guidance.[vi]
The bulletin expects insurers using AI to maintain a written governance program, validate and monitor model performance, document data quality and model objectives, and oversee third-party systems. It also makes clear that decisions supported by AI remain subject to existing insurance laws and consumer protections. States have taken somewhat different approaches, but the common focus is AI used by or on behalf of insurers.
Those frameworks primarily address how insurers use and govern AI. AI-assisted enrollment brings a related issue into focus: the quality of the information plans disclose. Existing policy already requires health plans to disclose different types of information in standardized formats. The Summary of Benefits and Coverage (SBC) standardizes consumer-facing information about benefits and cost sharing. Separately, the Transparency in Coverage rules require plans to publish negotiated prices in machine-readable files. Together, these requirements provide much of the information an AI system might use to estimate a consumer’s expected costs, but they were not necessarily designed with AI interpretation in mind.
AI raises the bar for the quality of that information. In the study, one plan’s SBC appeared to indicate that outpatient mental-health visits were subject to a $50 copay from the first visit. The insurer’s filed plan design, however, required the deductible to be met first, meaning the member would pay the full negotiated cost of each visit until reaching the deductible. For a therapy-heavy member, that difference materially changed the expected out-of-pocket costs. The problem is not limited to benefit descriptions. CMS has identified related challenges with existing price-transparency files, including data ambiguity and insufficient contextual information.[vii]
As AI increasingly processes insurance information on consumers’ behalf, disclosure policy may need to consider not only whether information is available and understandable to consumers, but also whether it is sufficiently clear, consistent, and contextualized to support reliable machine interpretation.
AI Governance and Consumer Protection
Consumer decision support is not new to insurance regulation. In the ACA marketplace, web brokers that recommend qualified health plans must explain the rationale for their recommendations and the methodology behind their default plan display. They also cannot favor plans based on compensation from insurers, must allow consumers to view all available qualified health plans, and are subject to standards requiring correct and non-misleading information.[viii] Agents and brokers are also subject to state producer-licensing requirements.
The employer-sponsored market is less uniform. Employers decide which plans to offer and how contributions are structured, while ERISA responsibilities depend on the function being performed. A third-party administrator or enrollment vendor performing administrative tasks generally is not an ERISA fiduciary, but fiduciary responsibilities can arise when an entity exercises discretion in administering the plan.[ix]
Some states are beginning to regulate AI decision systems more directly. Colorado’s 2026 Automated Decision-Making Technology Act, effective in 2027, expressly covers AI used in insurance decisions such as underwriting, pricing, coverage, and claims, while excluding differentiated product recommendations from its definition of consequential decisions.[x] Colorado separately enacted requirements for conversational AI services[xi], illustrating how quickly the state-level framework is evolving.
These rules provide substantial protections when AI operates within traditional insurance channels. Consumer-facing AI creates a different regulatory issue: A consumer can independently ask a chatbot which plan to choose and receive what functionally resembles individualized insurance advice without necessarily going through the marketplace, employer, insurer, or broker that traditionally provided decision support.
The study illustrates that AI responses are probabilistic, so asking the same question multiple times can produce different recommendations. Additionally, the AI recommendation may embed assumptions that are not obvious to the consumer. In the study, for example, models overwhelmingly treated minimizing expected annual cost as the primary decision objective, even though consumers may value downside protection, network access, or other considerations differently.
It is too early to know whether this use of consumer AI calls for a policy response. But as the tools become more capable, it is worth monitoring how consumers use them, what drives their recommendations, and how they interact with the existing rules around insurance advice and enrollment.
Looking Ahead
The first generation of insurance technology put policies online. The next may put increasingly sophisticated insurance reasoning into a conversation. The 2027 open enrollment season may be the first in which a large number of consumers have access to an AI insurance advisor capable of making meaningful, personalized plan recommendations.
What happens next is worth watching. For consumers, these tools could make a complex product easier to understand and compare. For insurers and employers, they could change how people respond to plan design and contribution choices. For actuaries, they may become another factor shaping enrollment and risk selection. And for policymakers, the question is how existing frameworks apply and what, if anything, warrants additional attention as consumer AI becomes more widely used.
The coming enrollment season should begin to show how consumers use these tools, whether the recommendations influence their choices, and the extent to which those choices shape enrollment patterns and risk selection.
Khris Dai, MAAA, FSA, is founder & CEO of Visuary AI. She is a member of the Academy’s Behavioral Health Task Force.
Endnotes
[i] OpenAI. (2026, January). AI as a Healthcare Ally: How Americans are navigating the system with ChatGPT.
[ii] Consumer AI as Health Plan Decision Support: A Benchmark of LLM Plan Recommendations against a Claims-based Pricing Engine.
[iii] Navigating the Maze: A Look at Health Insurance Complexities and Consumer Protections, KFF.
[iv] Benjamin R. Handel, Jonathan T. Kolstad, and Johannes Spinnewijn, “Information Frictions and Adverse Selection: Policy Interventions in Health Insurance Markets,” The Review of Economics and Statistics 101, no. 2 (2019): 326–340.
[v] Modeling Effects of Enrollee Choice, Society of Actuaries.
[vi] “Implementation of NAIC Model Bulletin: Use of Artificial Intelligence Systems by Insurers,” status as of April 1, 2026.
[vii] Centers for Medicare & Medicaid Services, “Transparency in Coverage Proposed Rule (CMS-9882-P),” Dec. 19, 2025.
[viii] 45 C.F.R. § 155.220(c)(3); Centers for Medicare & Medicaid Services, guidance for Marketplace web-brokers.
[ix] U.S. Department of Labor, Understanding Your Fiduciary Responsibilities Under a Group Health Plan.
[x] Colorado S.B. 26-189, Automated Decision-Making Technology (2026).
[xi] Colorado General Assembly, H.B. 26-1263, Conversational Artificial Intelligence Service Operator Requirements (2026).