Millions of beneficiaries across 44 state programs will have their eligibility re-examined by January 2027. Under the One Big Beautiful Bill Act (H.R.1), adults who became eligible for Medicaid under the Affordable Care Act must complete qualifying work activities or demonstrate an exemption, and states are now required to check Medicaid eligibility at least twice a year.

Facing compressed timelines and limited staff, states are looking to AI to automate the process; six agencies have already committed to using AI tools, and 21 more are considering it.

SHP’s Michelle Mello, JD, PhD, a professor of law and health policy, and Himaja Nagireddy, MS, a second-year doctoral student, write in this JAMA Health Forum article that AI could help states manage this looming workload if deployed carefully. If rolled out too broadly or too quickly, however, AI could make the law's effects on beneficiaries even worse. 

They suggest three strategies to help reduce this risk.

1.  Phasing AI Deployment by Task Complexity

AI should be deployed only for Medicaid redetermination tasks that AI tends to perform well. Algorithms excel at objective classifications drawn from structured, reliable data, such as verifying county unemployment rates or assessing Medicare Part B enrollment. 

AI can also synthesize multiple sources using predefined rules, even with imperfect data. Large language models (LLMs) are adept at extracting specific facts from messy, unstructured text—such as a scanned paystub or a clinician’s handwritten notes—while flagging uncertain cases for human review. Used well, AI can auto-approve eligibility at scale, freeing staff for more complex cases.

“Tasks that involve more subjective judgment are riskier to assign to AI,” Mello and Nagireddy caution. A leading example is the federal government’s complicated standard for “medical frailty” determinations, which evaluate whether a person has “‘serious or complex’ medical conditions, disabilities, substance use disorders, or other health conditions that ‘significantly impair’ paid and volunteer work.” 

2. Fostering Competition Among AI Vendors

Despite strong AI developer interest, 30 states plan to stick with existing vendors, citing time pressure, untested products, and cost concerns. Sixteen of these states report using a company whose past Medicaid eligibility algorithms were error-ridden enough to prompt complaints to the Federal Trade Commission. Mello and Nagireddy point out that when automated systems erred during the “unwinding” of pandemic-era Medicaid enrollment protections, the failures “adversely affected thousands of beneficiaries and cost states millions in vendor and legal fees.”

Federal policymakers could promote competition by creating sandboxes for smaller companies to train and test models on Medicaid data, and an online clearinghouse comparing products on functionality, performance, and price. The Centers for Medicare and Medicaid Services has already published expressions of interest from nearly 50 companies seeking to build work-requirement implementation tools.

States could also collaborate on vendor evaluation and require monthly dashboards tracking error rates, appeal volume, and procedural disenrollments. These steps would help agencies match vendors to tasks, weigh tradeoffs between error types, and hold vendors accountable.

3. Addressing Skewed Incentives

Setting the cutoff score that a predictive model uses to decide who counts as eligible for Medicaid is a value-driven judgment. It involves balancing two kinds of error: wrongly denying coverage to eligible people, and wrongly granting it to ineligible ones.

Federal rules push states in one direction. Starting in 2029, state Medicaid programs that exceed a 3% eligibility error rate face financial penalties. Yet only one type of eligibility error is scrutinized: keeping ineligible people enrolled. Combined with the current administration’s emphasis on fraud, states are pushed toward stringent enforcement of work requirements, raising the risk that eligible people lose coverage they are entitled to.

Mello and Nagireddy argue: "Given the harm to vulnerable beneficiaries, AI models should minimize the risk of wrongful disenrollment. State programs—not AI vendors—should decide the thresholds and protocols for escalating uncertain cases."

Conclusion: 

Medicaid eligibility determinations are already a growing source of administrative strain. The authors conclude: “A faint silver lining to H.R.1’s new requirements may be the technological innovation it inspires. State programs’ investment in new tools could offer a practical, lasting solution—if they deploy AI where it can help, not harm.”