Skip to main content
    Data and architecture

    The Hidden Hard Problem in Healthcare AI: Acquiring Trustworthy Medicare Data from Fragmented Systems

    Before AI can reason over healthcare data, someone has to reliably acquire and govern it

    8 min readWilf Russell· CTO, Livmor.ai

    The conversation around AI in healthcare often starts with models.

    Large language models. Copilots. Agents. Prompt engineering. Retrieval systems.

    But long before any of that matters, there is a more foundational challenge healthcare platforms must solve first:

    How do you acquire trustworthy healthcare data from ecosystems that were never designed to behave coherently?

    At Livmor.ai, we've found this challenge is significantly more difficult than most people expect, particularly in Medicare, where fragmentation is not an edge case. It is the operating model.

    What appears externally as a simple "plan search" or "AI-powered Medicare experience" is actually the result of disciplined acquisition architectures continuously locating, retrieving, validating, and governing data across dozens of inconsistent systems, formats, and operational workflows.

    This is not a matter of "winging it" through fragmented data.

    It is an engineering and governance problem requiring deliberate acquisition strategies, provenance tracking, validation layers, and operational controls capable of producing trustworthy downstream experiences despite deeply inconsistent upstream ecosystems.

    Before AI can reason over healthcare data, someone has to reliably acquire and govern it.

    That turns out to be the hard part.

    The Illusion of "Available" Healthcare Data

    At a glance, Medicare data appears highly accessible.

    The Centers for Medicare & Medicaid Services publishes extensive datasets covering plans, formularies, providers, pharmacies, and enrollment structures. Carriers publish plan materials and Evidence of Coverage documents. State DOI systems expose Medigap filings and rate information. Interoperability frameworks from HL7 International provide APIs and structured exchange models.

    From the outside, it can appear that the challenge is simply connecting to APIs and storing records.

    The reality is far more operationally hostile.

    Different systems update on different schedules. Different organizations model similar concepts differently. Some data sources are machine-readable APIs; others are PDFs, spreadsheets, web portals, or interactive search tools never designed for automated extraction.

    Even the definition of "authoritative" becomes blurry in practice.

    During Annual Enrollment Period transitions, for example, CMS crosswalk files are frequently treated as canonical references for plan migration and mapping. Yet operationally, those crosswalks can diverge from what ultimately appears downstream on Medicare.gov.

    Not enough to create systemic alarm.

    But absolutely enough to erode trust if downstream platforms assume all upstream systems agree.

    To be clear, the industry is actively working to improve interoperability and acquisition consistency. Frameworks such as FHIR, Da Vinci PDex, CMS interoperability mandates, and payer API initiatives represent important progress toward more standardized healthcare exchange models.

    But operationally, fragmentation remains deeply embedded in the ecosystem.

    Standards improve transport and structure, but they do not eliminate asynchronous publication cycles, inconsistent operational implementations, incomplete datasets, state-level variability, or real-world divergence between systems representing ostensibly similar information.

    The existence of standards does not eliminate the need for disciplined acquisition and governance architectures.

    In many ways, it increases the importance of them.

    Once you encounter this at scale, you stop thinking about healthcare ingestion as ETL.

    You start thinking about it as continuous acquisition under uncertainty.

    Fragmentation Exists at Every Layer

    One of the biggest misconceptions in healthcare technology is that fragmentation is primarily a schema problem.

    It isn't.

    Fragmentation begins much earlier, at discovery and acquisition.

    A single Medicare acquisition workflow may involve CMS plan and formulary files, state DOI filings, carrier PDFs, provider network APIs, pharmacy datasets, browser automation against legacy portals, spreadsheets uploaded manually, and human validation workflows where no reliable machine-readable source exists at all.

    The acquisition strategy itself becomes part of the architecture.

    Different states expose data through entirely different operational models.

    New York Medigap data, for example, may arrive through structured PDF rate tables published by the Department of Financial Services. Florida exposes rate information through an interactive consumer web experience. California historically relied on Oracle APEX-backed workflows with different normalization characteristics entirely.

    Some states provide downloadable spreadsheets. Others expose semi-structured filings. Others effectively require browser automation and extraction logic against systems never intended for machine consumption.

    That is why the platform evolved toward a parser registry architecture rather than a single ingestion model. Each acquisition strategy must adapt to the operational realities of the source while still producing a normalized downstream representation.

    This is not simply API integration.

    It is acquisition engineering.

    And despite all of that variation, the downstream experience still needs to feel coherent and trustworthy to consumers asking questions like "Is my doctor covered?", "What happens to my formulary next year?", "Why did this premium change?", or "Which plan best fits my medications and providers?"

    That coherence only exists if the underlying acquisition systems are disciplined enough to absorb fragmentation without exposing it to the consumer experience.

    Parser registry
    One acquisition strategy per source, one normalized output
    New York
    Structured PDF rate tables
    Community rated, by region and plan letter
    Florida
    An interactive consumer web tool
    Attained-age rates, county by county
    California
    A form-driven web application
    Annual rates and plan variants
    Other states
    Downloadable spreadsheets
    One general parser for the standard rate layout
    No parser
    Manual acquisition
    Entered and validated by a person, recorded as a patch
    Figure 1: State-level healthcare data fragmentation requires adaptive parser strategies spanning PDFs, web systems, spreadsheets, APIs, and manual acquisition workflows.

    The Architecture Problem Nobody Sees

    Consumers experience healthcare platforms through relatively simple interactions: searching for plans, comparing premiums, validating provider participation, checking medication coverage, or asking AI-powered questions about benefits and costs.

    Behind those experiences sits an acquisition architecture continuously locating, retrieving, partitioning, validating, and tracking healthcare data across fragmented ecosystems.

    At operational scale, this becomes substantial quickly.

    A single Medicare Advantage cycle may involve thousands of plans, tens of thousands of formulary variations, hundreds of thousands to millions of provider-network relationships, and continuously evolving pharmacy and eligibility datasets.

    And unlike many domains, healthcare data cannot simply be treated as append-only. Relationships drift constantly across plan years, provider affiliations, formularies, pharmacy networks, and regional market changes.

    The acquisition challenge is therefore not only scale.

    It is continuous change management under regulatory and operational constraints.

    At Livmor, the platform evolved into a layered acquisition pipeline responsible for source discovery, parser orchestration, staged acquisition, validation, partition-aware ingestion, provenance tracking, and downstream production promotion.

    The architecture treats each stage as an independent trust boundary rather than a single ingestion flow.

    That distinction matters because healthcare acquisition is not simply about collecting records.

    It is about continuously assessing how trustworthy those records are at every stage of the pipeline.

    Sources
    Where the data lives
    State filings, rate tables, web tools, spreadsheets
    Extract
    Retrieve and record
    The raw file is kept as retrieved, with a manifest of where and when it came from
    Stage
    Parse and validate
    Records are staged apart from production and checked before anything depends on them
    Curate
    Govern and correct
    Rules and human-managed patches are applied, each one audited and reported
    Promote
    Publish atomically
    A validated snapshot replaces the production set in one step
    Figure 2: A layered acquisition architecture where healthcare data is progressively retrieved, validated, staged, and prepared for downstream governance and curation.

    The Cadence Problem: Asynchronous Truth

    One of the hardest operational realities in healthcare acquisition is that nothing updates at the same time.

    CMS releases follow regulatory schedules. Carrier updates occur continuously. Provider directories drift constantly. Formularies evolve throughout the year. State filing systems operate independently.

    This creates an environment where "latest" does not always mean "correct."

    A carrier may publish updated plan information before CMS has finalized or exposed the corresponding downstream confirmation data. A provider directory may update before affiliated network relationships have propagated through adjacent systems. During AEP windows, plans can temporarily exist in partially synchronized states across carriers, CMS artifacts, and consumer-facing experiences.

    Healthcare acquisition systems therefore need to balance freshness, trustworthiness, source stability, operational confidence, and downstream consumer impact simultaneously.

    Blindly preferring "newest available data" can actually reduce trust rather than improve it.

    This becomes especially important during AEP windows, where source systems often evolve asynchronously while consumers are actively making plan decisions.

    Without disciplined acquisition controls, healthcare platforms become overwrite engines rather than trustworthy systems.

    Acquisition at Scale Creates Infrastructure Problems

    Healthcare acquisition is not lightweight.

    Large CMS acquisition runs can involve millions of records, long-running COPY operations, parser execution, document extraction workflows, browser automation, validation pipelines, and partition-aware orchestration operating simultaneously across many datasets.

    Running those workloads directly against production-serving infrastructure creates operational risk:

    • degraded query responsiveness,
    • I/O contention,
    • lock escalation,
    • and reduced consumer experience quality.

    To avoid that, Livmor separated acquisition infrastructure from production query systems entirely.

    The platform uses dedicated acquisition environments combined with blue-green promotion patterns and atomic promotion workflows.

    That separation allows acquisition pipelines to scale independently while protecting downstream user experiences from ingestion-related volatility.

    More importantly, it allows data acquisition and validation workflows to evolve continuously without destabilizing production-serving systems.

    Provenance Starts at Acquisition

    One of the most important architectural lessons we learned is that provenance cannot be added later.

    It must begin at acquisition time.

    Every acquisition workflow needs to know where the source originated, when it was retrieved, which parser processed it, what checksum validated the artifact, which partition produced the output, and which downstream datasets inherited the result.

    This lineage becomes critically important once discrepancies inevitably emerge across systems.

    Without acquisition provenance, downstream explainability becomes nearly impossible.

    And once AI systems are layered on top of healthcare workflows, explainability is no longer optional.

    Trustworthy AI systems require trustworthy acquisition systems beneath them.

    Why This Matters for AI

    AI systems inherit the strengths and weaknesses of the acquisition systems beneath them.

    If provider acquisition pipelines drift out of date, AI may confidently recommend physicians who are no longer in-network. If formulary acquisition lags carrier changes, downstream AI experiences may generate incorrect drug coverage guidance. If crosswalk mappings diverge during AEP windows, recommendation systems may incorrectly infer plan continuity or migration paths.

    These are not merely "hallucinations" in the traditional LLM sense.

    They are trust failures originating upstream in acquisition and governance layers.

    That distinction matters because many healthcare AI conversations focus heavily on model behavior while underestimating the operational fragility of the underlying data ecosystem.

    At Livmor, one of the core architectural realizations was that trustworthy AI experiences require trustworthy acquisition systems first.

    Before a platform can answer healthcare questions intelligently, it must continuously locate, retrieve, validate, and govern a coherent representation of an ecosystem that was never designed to behave coherently in the first place.

    That is the real systems challenge.

    The Strategic Reality

    Healthcare AI is increasingly framed as a model problem.

    But the organizations building durable platforms are discovering something different:

    The long-term advantage is not just intelligence.

    It is operational data discipline.

    The ability to continuously locate, retrieve, validate, track, and govern fragmented healthcare information at scale becomes the foundation upon which every downstream experience depends.

    Without that discipline, even the most sophisticated AI systems eventually lose trust.

    Final Thought

    The healthcare industry keeps asking whether AI can understand healthcare.

    A more important question may be this:

    Can we build healthcare acquisition systems capable of continuously locating, retrieving, validating, and governing trustworthy data from an ecosystem fundamentally designed around fragmentation?

    At Livmor, we believe the organizations that solve that challenge will define the next generation of trustworthy healthcare AI experiences.

    More from Wilf Russell

    Stay ahead of what's changing in Medicare

    Get platform updates, industry insights, and LivmorIQ news delivered to your inbox.

    No commitment. We'll reach out before June 30 to complete your setup.