From 30+ Minutes to Seconds: What AI Abstraction Actually Changes About Your HEDIS Season

Manual abstraction is costly and runs on a fixed calendar

Spend a summer inside a health plan’s quality department and the rhythm becomes familiar: the fax machines hum, the spreadsheets multiply, and somewhere a certified HEDIS abstractor is on their third coffee, squinting at a scanned PDF of a provider note from eight months ago, deciding whether a counseling checkbox counts for the WCC measure. Multiply that single chart by 411 per measure, the hybrid-methodology floor (CMS QRS Technical Specifications 2026), then by however many HEDIS measures the plan reports, and then by the number of plans still using hybrid sampling. The industry talks about “burden” a great deal, but it rarely puts a dollar sign on it.

The burden of manual chart review is stark. A trained abstractor spends up to 30 minutes on a single chart for a complex hybrid measure, and the cost of that specialized clinical expertise - interpreting unstructured notes and applying HEDIS specifications - runs roughly 10× what an AI-assisted review costs per chart. For a mid-sized plan running hybrid methodology across 15 measures at 411 charts each, the arithmetic is 6,165 charts at a manual cost can add up to a six-figure abstraction pass before overread, before inter-rater reliability checks, and before the inevitable rework when an auditor flags something.

The money is the smaller half of the problem. The larger cost is time, because HEDIS season runs on a fixed calendar: data lock, abstraction window, submission deadline. Every day a chart sits in a queue is a day the plan cannot spend on provider outreach, member engagement, or chasing the one missing lab value that would close a gap. Plans routinely hire seasonal staff to absorb the surge, training them, running them through inter-rater reliability testing, and hoping they stay through the whole cycle. It is a staffing model built on a prayer, and it is the model the technology described below is in the process of replacing.

Healthcare-specific language models match expert abstractions on quality measures

The term “AI chart review” gets used loosely, so it is worth saying precisely what it means for quality measures. The data that HEDIS evidence lives in or fails to live in is often unstructured clinical text: provider notes, lab reports, specialist consult letters, discharge summaries. A diabetes eye exam may be documented in a single sentence buried in a PCP note; a BMI percentile with counseling may be spread across two encounters in two different EHRs. Healthcare-specific language models, fine-tuned on real clinical corpora and validated against domain-expert annotations, can read these documents and extract the specific data elements that HEDIS measures require: dates, values, codes, and the clinical context that decides whether something counts.

The evidence that this works at production grade is no longer only vendor marketing. A 2024 pilot study tested whether large language models could abstract complex quality measures from EHR data, in this case hospital-level measures such as sepsis bundles, and found that they can perform accurate abstractions that match expert human reviewers (Boussina et al, NEJM AI 2024). The same clinical NLP approach translates directly to HEDIS measure abstraction, because the document types, provider notes and lab reports, and the extraction tasks, finding specific clinical data points in unstructured text, are structurally identical. An LLM trained to classify reasons for missed follow-up after an abnormal FIT achieved 89% accuracy against physician review and surfaced addressable barriers such as refusal, comorbidities, and transportation (Williams et al., Jt Comm J Qual Patient Saf 2026), and NCQA is actively developing a HEDIS measure for follow-up after abnormal CRC screening, which turns this classification task into a directly auditable workflow. A broader review of LLMs in healthcare quality management confirmed that automated measure abstraction has moved from proof-of-concept studies into operational deployment, with multiple health systems and vendors running production pipelines (Frontiers in Digital Health [2026]).

What this looks like in practice is a reassignment of the abstractor’s role. Instead of spending 30 minutes reading a chart and typing its evidence into a HEDIS software tool, an AI engine ingests the chart, identifies every relevant data point across the measure specifications, flags the ones that close a gap, and surfaces only the edge cases for human review. The abstractor becomes a validator, not a data-entry clerk, which is the same shift in posture that the Quality Health chart review platform is built around.

The move from data entry to validation reshapes coverage, consistency, and cadence

The most immediate change is speed, with more than a 90% reduction in manual review time possible. Even read conservatively, a 30-minute chart that becomes a 10 second validation lets an abstraction team process six to ten times its current volume in the same window. Speed, though, is not the most interesting change. Three others matter more.

The first is coverage. When abstraction is slow and expensive, plans sample, and the HEDIS hybrid methodology sets a 411-chart floor that plans rarely climb far above because every additional chart costs real money (CMS QRS Technical Specifications 2026). AI abstraction breaks that math: as the marginal cost per chart approaches zero, a plan can review its entire eligible population rather than the sample, which means more charts reviewed, more gaps found, and more gaps closed.

The second is consistency. Manual abstraction has an inter-rater reliability problem, in which different abstractors read the same clinical scenario differently and the same abstractor reads it differently on Tuesday than on Thursday. A model does not tire, does not have off days, and applies the specification rules identically every time. For any plan that has been through a painful HEDIS audit where inter-rater disagreement was flagged, the value of that consistency shows up directly in the audit letter.

The third is year-round abstraction. HEDIS is no longer really a season; it is becoming a continuous quality monitoring exercise. NCQA is pushing hard toward digital quality measures, and CMS is reinforcing data-driven reporting expectations (NCQA 2025). Plans that can already afford to abstract year-round do so; AI makes year-round abstraction affordable for every plan, which quietly dissolves the seasonal staffing model described above.

Finding what shouldn’t be there is the RADV half of the problem

The standard chart review workflow runs in one direction: the reviewer hunts for missing evidence. Did the member get an eye exam? Was the BMI percentile documented? Is there a mammogram result in the chart? The work is gap closure, and the rate goes up when the gaps fill.

There is a parallel problem that is arguably more dangerous, and that most AI chart review conversations miss entirely: codes that are in the record but should not be. A diagnosis code copied forward from a prior visit without re-verification; a risk-adjustment HCC that lacks the documentation to support it; a measure exclusion applied incorrectly. This is the RADV (Risk Adjustment Data Validation) angle, and it just acquired a new regulatory edge. In January 2026, the CMS 2027 Medicare Advantage Advance Notice signaled a clear shift toward bidirectional validation, in which health plans are expected not only to find missing codes but also to identify and remove unsupported ones. Industry analysts have noted this marks a decisive turn: plans that cannot do both are walking into audit exposure (CMS 2027 Advance Notice; Shashank 2026).

A properly architected AI abstraction engine reads charts in both directions at once. It flags the missing A1c that would close the diabetes eye exam gap, and it flags the HCC code that appeared in the problem list but has no supporting documentation in the encounter note. That is the difference between “we found more gaps” and “we are actually RADV-ready.” Most competing solutions in the HEDIS AI space were built for the old world of finding gaps, closing gaps, and reporting the rate. The bidirectional approach is not a feature layered onto that architecture; it is a fundamentally different architecture decision, and the half of the problem it does not solve is the half that shows up in an audit.

Digital measures move the intelligence upstream, they do not retire it

A plan thinking it will keep doing manual abstraction for a few more years should read what NCQA has published. The organization has stated it is shifting toward digital-only HEDIS reporting, powered by clinical data from EHRs through FHIR APIs (NCQA 2025). The MY 2026 specification expanded ECDS (Electronic Clinical Data Systems) reporting paths, and NCQA has signaled that digital-only HEDIS is the destination by 2030 - making the timeline a matter of when, not if. This is not a technology trend; it is a compliance timeline.

Digital measures do not eliminate the need for AI abstraction. They change where the intelligence lives. Even with perfect FHIR pipelines, the work is still extracting structured data from unstructured text. The difference is that the AI sits in the data pipeline rather than in the abstraction tool, processing clinical notes as they flow in rather than reading them retrospectively. The same NLP, the same clinical language models, simply sit earlier in the workflow. The plans investing in AI abstraction now are not just making this year’s HEDIS season easier; they are building the infrastructure that will feed digital quality reporting in 2028, 2029, and 2030. The plans that wait until digital measures are mandatory will be building that infrastructure under a deadline, under audit pressure, while their competitors have been refining their models for three years.

Five questions that separate a gap tool from an audit-defense tool

For a quality team evaluating solutions, and based on the volume of RFPs moving through the quality space right now many are, the discrimination that actually matters is the one above: gap closure versus audit defense. Five questions sort the market.

Clinical model quality, not just “AI”. A general-purpose LLM that can write emails will not abstract HEDIS measures accurately. The models the work requires are trained on clinical corpora, fine-tuned for medical NLP tasks, and validated against domain-expert annotations. Ask vendors what their training data is, what benchmarks they have tested against, and whether they can show abstraction accuracy rates broken down by measure.

Bidirectional review capability. A platform that only hunts for missing gaps is a gap-closure tool; a platform that also identifies unsupported codes is an audit-defense tool. These are not the same thing, and the difference should be asked for explicitly: does the system flag codes that lack sufficient documentation support?

HIPAA compliance architecture. The technical implementation matters as much as the policy. Where does the data live during processing? Is the model running in a dedicated, isolated environment? What is the business-associate chain? Health plans are understandably skittish about sending PHI through third-party AI pipelines, and the right answer is a HIPAA-compliant, dedicated-instance architecture in which the data never leaves the plan’s control. The AI Governance layer the platform enforces is built for exactly this question.

Audit trail and transparency. NCQA has started specifying requirements for technology-driven abstraction validation in HEDIS audits (Advent Advisory Group [2025]). A vendor that cannot show exactly which clinical sentence produced which data element, with a human-readable audit trail, will fail the auditor’s first question. Black-box AI does not fly in a HEDIS audit.

Integration with existing HEDIS software. Quality teams already run an abstraction platform, a QI workflow manager, and a Stars analytics dashboard. An AI abstraction engine that requires ripping all of that out and starting over is a non-starter. The solutions worth shortlisting integrate with what the team already uses.

Closing the gap is half the problem; defending the code is the other

HEDIS abstraction has been a manual, expensive, seasonal grind for decades, and that era is ending faster than most quality leaders realize. The technology exists, the economics work at roughly $50 per chart versus $5 with a roughly 10x ROI on a platform such as Quality Health’s, and NCQA is building the regulatory framework for it. A plan that selects a solution which only looks in one direction, only finding what is missing, is solving the problem of 2020 rather than the problem of 2027. The audit environment is changing, and bidirectional review is the difference between closing gaps and defending the codes.

The hardest part is not the technology. It is rewiring the muscle memory of an industry that has done HEDIS the same way since the 1990s. But the math does not lie: 30 minutes to 3 minutes, a seasonal scramble becoming a year-round quality engine, and a sampling floor becoming a full-population review. That is not a marginal improvement. It is a category change.

Sources

• Saraswathula A, Merck SJ, Bai G, et al. “The Volume and Cost of Quality Metric Reporting.” JAMA. 2023; 329(21):1840-1847. PMID: 37278813.
• American College of Emergency Physicians (ACEP). “Quality of Care and the Outcomes Management Movement.” Accessed via acep.org.
• SMART Health IT / EHR Good Neighbor. “MultiCare: Reducing Cost and Burden in Quality Reporting with FHIR.” Case study. Accessed via good-neighbor.smarthealthit.org.
• NCQA. “AI-Powered Data Transformation Accelerating Digital Quality.” NCQA Blog, October 24, 2025. Accessed via ncqa.org.
• NCQA. “HEDIS MY 2026: What’s New, What’s Changed, What’s Retired.” NCQA Blog, August 1, 2025. Accessed via ncqa.org.
• Centers for Medicare & Medicaid Services (CMS). “2026 Quality Rating System Measure Technical Specifications.” Accessed via cms.gov.
• Boussina A, et al. “Large Language Models for More Efficient Reporting of Hospital Quality Measures.” NEJM AI. October 24, 2024. PMC11658346.
• Williams et al. “LLM Classification of Reasons for Missed Follow-Up Colonoscopy.” Jt Comm J Qual Patient Saf. 2026. Accessed via jointcommissionjournal.com.
• “Large Language Models in Healthcare Quality Management: A European Perspective on Process Automation and Compliance.” Frontiers in Digital Health. March 26, 2026.
• Advent Advisory Group. “HEDIS MY 2025 Audit Changes.” November 13, 2025. Accessed via adventadvisorygroup.com.
• CMS. “2027 Medicare Advantage Advance Notice.” January 2026, referenced for the RADV and bidirectional validation direction; with Dentons On Call, “CMS Issues Proposed Changes to MA Risk Adjustment Process and RADV Audit Update,” January 27, 2026.