CMS Seeks AI Tools to Enhance Medicare Plan Selection Process

On February 23, 2026, the Centers for Medicare & Medicaid Services issued a Request for Information seeking AI and machine learning tools to improve how Medicare beneficiaries navigate plan selection. The document is narrow in its immediate scope — CMS wants to improve Medicare.gov, the Plan Finder tool, and the 1-800-MEDICARE call center. But read in the context of the broader federal AI agenda and CMS’s concurrent innovation models, the RFI signals something considerably more significant: a deliberate attempt to embed predictive analytics and conversational AI into the structural core of Medicare administration.

Understanding what CMS is actually asking for, why it matters operationally, and what constraints could limit its ambitions requires moving past the surface-level procurement framing.

What CMS Is Actually Asking For

The RFI’s stated problem is straightforward. Medicare beneficiary decision-making currently relies on static comparison tables and dense plan documents. For the roughly 70 million Americans enrolled in Medicare — a population disproportionately affected by limited health literacy, language barriers, and cognitive challenges — these tools are demonstrably insufficient. During peak open enrollment periods, call center wait times spike as representatives walk beneficiaries through plan comparisons verbally and at length.

CMS is now soliciting tools across four functional categories. The first is personalized, claims-informed plan recommendations: AI systems that can use a beneficiary’s actual utilization history to suggest plans likely to minimize out-of-pocket costs and maximize coverage fit. The second is predictive analytics for plan matching — models that go beyond historical claims to incorporate demographic and preference signals in generating recommendations. The third is plain-language translation of Medicare documents and plan materials. The fourth, and perhaps most operationally ambitious, is real-time conversational AI: chatbots, virtual assistants, and AI voice advisors capable of handling routine call center inquiries without human intervention.

Respondents must demonstrate market readiness, organizational independence, and data ownership transparency. They must also specify their approach to privacy, security, and equity — requirements that reflect both the sensitivity of Medicare claims data and the well-documented risk of algorithmic bias in tools built on historically unequal healthcare utilization patterns.

The comment window closes March 31, 2026.

The Regulatory and Institutional Context

This RFI does not emerge in a vacuum. It is one component of an accelerating federal AI strategy that has been building since the Trump administration’s February 2025 OMB Memorandum M-25-21, which directed agencies to accelerate AI adoption. HHS responded in December 2025 with its own AI Strategy, framing a “OneHHS” approach to embedding AI across departmental operations.

The scale of this shift is quantifiable. A 2025 Government Accountability Office report found that AI use cases across eleven federal agencies nearly doubled from 571 in 2023 to 1,110 in 2024. HHS recorded the largest single-agency increase — from 157 to 271 total use cases — with generative AI deployments within HHS rising from 7 to 116 in the same period. CMS’s RFI is, in this light, consistent with a broader institutional posture rather than an isolated initiative.

Concurrently, CMS has launched a series of innovation models with explicit AI components. The WISeR (Wasteful and Inappropriate Service Reduction) model, active since January 2026 in six states, deploys AI alongside clinical review to identify low-value services in Original Medicare prior authorization processes. The ACCESS model, beginning July 2026, uses AI diagnostics to identify chronic disease patients suited for technology-enabled care management and ties payments to measurable outcomes rather than service volume. The Health Technology Ecosystem initiative, which has attracted more than 600 organizations since its July 2025 launch, frames AI as the catalyst for breaking down data silos that have long impeded interoperability.

The February RFI on plan selection fits within this architecture as the beneficiary-facing layer — the consumer interface through which Medicare’s complexity is filtered for the individual enrollee.

Why Plan Selection Is a High-Stakes AI Application

The operational logic for AI in Medicare plan selection is compelling. Open enrollment decisions carry year-long financial consequences. Research consistently shows that beneficiaries systematically choose plans that do not minimize their expected costs, in part because the complexity of premium, deductible, and coverage comparisons exceeds what most individuals can process unaided. Claims-informed recommendation engines have the potential to correct this at scale, using actual utilization history to project real cost exposure under different plan scenarios.

Over half of Medicare beneficiaries are now enrolled in Medicare Advantage rather than Original Medicare — a threshold the CMS Innovation Center highlighted in its MA improvement RFI earlier this cycle. This concentration means that plan selection decisions have direct implications for MA plan enrollment volumes, risk pool composition, and ultimately the bids that plans submit to CMS. AI tools that systematically redirect beneficiaries toward plans with lower expected costs could accelerate a rebalancing of the MA market, with particular pressure on plans that have historically benefited from enrollment driven by information asymmetry rather than value.

For plan operators, the downstream implications of CMS-deployed recommendation AI warrant close attention. If predictive matching tools reduce the friction of switching plans, churn rates could increase for plans whose networks, formularies, or cost structures are unfavorable to high-utilization members. Combined with the above-projection MA bids CBO noted for 2026 — already contributing to the HI Trust Fund’s accelerated depletion — this represents a compounding pressure on MA plan economics.

The Equity and Bias Problem Is Not a Side Issue

CMS’s requirement that respondents articulate their approach to fairness and equity is not boilerplate. It reflects a genuine and technically difficult challenge. AI tools trained on Medicare claims data will inherit the patterns embedded in that data — including patterns shaped by decades of unequal access to care, underdiagnosis in certain populations, and differential service utilization by race, geography, and socioeconomic status.

A recommendation engine that optimizes for predicted cost minimization, without accounting for these structural distortions, risks systematically steering low-income or minority beneficiaries toward lower-premium plans with restricted networks — not because those plans are optimal for them, but because their historical claims make them appear lower-risk. This is a known failure mode in healthcare AI that the industry has been grappling with since algorithmic decision-support tools entered clinical workflows.

The RFI’s equity requirements are necessary but insufficiently specified at this stage. A meaningful vendor response would need to address not only how bias is measured within the tool, but how it is monitored in production — particularly given that beneficiary populations and plan compositions change annually with open enrollment resets. CMS would be well-served to establish explicit, testable equity benchmarks as part of the eventual procurement framework rather than treating fairness as a pass/fail attestation.

The Structural Tension: Deregulatory Posture, Complex Deployment

There is a tension embedded in this initiative that deserves acknowledgment. The broader federal AI strategy is explicitly deregulatory in orientation. The administration has framed AI governance around removing barriers to private sector innovation rather than prescribing detailed oversight frameworks. ASTP/ONC, notably, withdrew non-finalized provisions of the HTI-2 Proposed Rule in late December 2025, signaling a preference for lighter-touch regulation in health technology.

Yet deploying AI at the scale CMS envisions — personalized plan recommendations for tens of millions of beneficiaries using individually identifiable claims data — involves precisely the kind of high-stakes, data-intensive application where governance frameworks are most consequential. Conversational AI tools operating as virtual insurance advisors at the Medicare call center carry liability and accuracy obligations that differ categorically from, say, an internal workflow automation tool.

This tension is not unique to CMS. It reflects the broader challenge facing healthcare AI deployment across the industry: the tools most likely to generate meaningful outcomes at population scale are also the ones that require the most robust governance, and governance takes time and institutional capacity to build.

What Comes Next

The RFI closes March 31. From there, CMS may issue a formal solicitation, pilot one or more tools in a defined geographic or enrollment context, or incorporate vendor responses into future rulemaking. The Health Technology Ecosystem framework suggests CMS prefers a challenge-based, voluntary model to regulatory mandates — which may mean that AI plan selection tools are initially tested through Innovation Center mechanisms before any program-wide rollout.

For vendors, the RFI is an early signal worth taking seriously. The market for Medicare-facing AI decision support is substantial, and CMS as a direct purchaser represents a distribution channel of unparalleled reach. Respondents who can demonstrate not only technical capability but also a credible, auditable approach to equity and beneficiary protection will be better positioned as CMS moves from information-gathering to procurement.

For the broader healthcare system, this RFI represents a meaningful inflection point. Medicare plan selection has long been a domain where information asymmetry disadvantaged beneficiaries and insulated plans from competitive pressure. AI-assisted navigation, done well, could shift that balance. Done poorly, it could replicate existing inequities at machine speed and government scale. The difference lies almost entirely in the rigor of implementation — and that rigor, as of late February 2026, remains to be defined.

Sources