Among acute-care visits by Medicaid beneficiaries, 39.6% are for ambulatory care-sensitive conditions, the conditions that timely outpatient care can usually prevent. Most of those visits happen in emergency departments. That share ranges from 35.1% in Pennsylvania to 44.2% in Mississippi, and within individual states, acute-care utilization for these conditions varies more than 13-fold between counties.

Variation of that size is difficult to explain by differences in population health alone. The analysis found that proximity to urgent care facilities and the density of rural health clinics were among the largest explanatory variables, which points at local care infrastructure as a driver of where these visits happen. The community engagement requirement takes effect January 1, 2027, and is expected to reduce service lines, appointment availability, and coverage further.

Why hiring does not close the access gap

HRSA projects a shortage of 124,180 physicians by 2027, rising to 187,130 by 2037, with the largest deficits in primary care and non-metropolitan areas. More than one million registered nurses are expected to retire by 2030. Rural areas absorb the most of this: 417 rural hospitals are at risk of closure, and 41.2% currently operate at a loss.

Primary care capacity is declining faster than it is being rebuilt, so hiring alone will not restore access for Medicaid beneficiaries. Automation is the response most organizations are reaching for, and it does increase the number of patients a team can reach in an hour. The open question is narrower: which parts of clinical decision-making can be automated safely.

Why general-purpose AI fails this population

A 2026 evaluation in Nature Medicine ran ChatGPT Health, OpenAI's consumer health tool, through 60 clinician-authored vignettes across 21 clinical domains. It undertriaged 52% of gold-standard emergencies, routing patients with conditions such as diabetic ketoacidosis toward a 24-to-48-hour appointment instead of the emergency department. When a family member minimized the symptom, the model shifted toward less urgent care, with an odds ratio of 11.7 (95% CI 3.7 to 36.6). These results describe ChatGPT Health in this evaluation and should not be generalized to all frontier models.

A separate benchmark evaluated 31 language models on 100 real primary care cases and found potential for severe harm in up to 22.2% of cases, at the higher end of the range. Of the severely harmful errors, 76.6% came from omission, the model failing to recommend a necessary action rather than recommending a wrong one. Safety correlated only moderately with existing medical-knowledge benchmarks, at r = 0.61 to 0.64.

Omission is the failure mode that matters most for Medicaid populations, and it is the one accuracy scores are worst at surfacing. When a model identifies a condition but does not apply the guideline that governs it, the output contains no false statement to flag. A clinician reviewing a recommendation that reads as reasonable has no reliable way to see what is missing. Multiple independent research groups have concluded that hallucination is an inherent property of large language models, so safety cannot rest on the model alone.

What separates a safe system from an unsafe one

If safety cannot rest on the model, it has to be enforced by a layer outside it. That reframes the evaluation question: which layer of the system is responsible for the clinical decision, and can that responsibility be demonstrated rather than asserted. Both are now measurable.

Waymark researchers compared methods for catching clinical safety violations across an 832-item benchmark spanning eight whole-person clinical domains:

• Formal verification identified all 612 violations with no false alarms, returning either a proof or a replayable counterexample for each case.
• Random unit testing detected 9% and returned 43 false safety assurances.
• A medically fine-tuned language model detected 69.9%, falling to 17.7% on violations spanning multiple encounters.
• Blinded board-certified physicians detected 81.3% in a matched sample.
• A frontier reasoning model also identified all 612 violations, at roughly a thousand times the computing cost.

The finding concerns the kind of evidence each method returns and what it costs to obtain, alongside the detection rates themselves. Formal verification returns a proof a reviewer can inspect. The other methods return a score that has to be trusted. The benchmark and verifier are publicly available for any team to run against its own rule sets.

For a Medicaid population whose care runs across many encounters and many hands, evaluation has to be based on results across encounters. Safety is a property of the care process, not of any single interaction.

What a purpose-built architecture looks like

Waymark Compass separates these layers so that the separation can be tested. Generative AI handles communication. Causal machine learning handles decisions. Deterministic rule engines enforce the safety constraints, and clinicians supervise the outcome. The safety-critical decision paths sit in a different architectural structure from the conversational layer, which is the arrangement that makes the separation verifiable rather than assumed.

Demonstrating that separation requires control over the clinical decision layer. A vendor deploying a licensed model it did not build cannot show where the boundary sits, or that it holds under load, because the demonstration requires architectural access the license does not confer.

What to ask ahead of a demo

Four questions establish where a vendor's safety claim actually rests, and each has a verifiable answer:

• Which part of the system makes the clinical decision?
• Which population was the model trained on, and how does that population differ from the one the health system serves?
• Does the safety claim rest on a proof, a sample, or a benchmark score?
• What happens when the model is uncertain, and who is accountable for the escalation?

A vendor who cannot answer these questions has not necessarily built an unsafe system. It may mean only that separation was never set as a design requirement, and so was never measured. The benchmark is public, and Waymark Compass is one of the systems it can be run against. For health systems expanding access under real workforce constraints, a standard that can be tested is worth more than a claim that has to be trusted.