Building an AI-First Safety Net: Pair Team’s Vision for Precision Benefits in Medicaid

DISCLAIMER: The views and opinions expressed in this essay are my own and do not reflect the views of my employer, or any of its affiliates.

---

If you are interested in joining my generalist healthcare angel syndicate, reach out to trey@onhealthcare.tech or send me a DM. Accredited investors only.

---

TABLE OF CONTENTS

Abstract

Introduction: Reimagining Medicaid Through AI

The Scale of the Opportunity

What Digital Assistants Actually Mean

Precision Benefits: The Core Innovation

Understanding the Barriers to AI Adoption

The Regulatory Reality

Pair Team’s Implementation Strategy

The Path Forward

ABSTRACT

Medicaid serves over 70 million Americans and represents roughly $700 billion in annual spending, yet remains largely untouched by the AI revolution reshaping commercial healthcare. In a new policy brief published in NPJ Digital Medicine, Pair Team and researchers from Stanford, Berkeley, and Carnegie Mellon introduce “Precision Benefits,” a framework for using AI-powered digital assistants to deliver targeted support at the exact moment it can prevent health deterioration and social crisis.

The core insight: predictive analytics can identify people at risk, but without real-time automated intervention, those insights sit unused in dashboards. Pair Team is building the infrastructure to close this gap.

Original Research Paper by Pair Team: https://link.springer.com/article/10.1038/s41746-025-02195-9

Key components of the vision:

- AI-powered digital assistants operating 24/7 in multiple languages for care navigation, eligibility verification, and behavioral health support

- Real-time intervention triggered by predictive signals like missed appointments or prescription refills

- Integration across fragmented systems connecting healthcare providers, community organizations, and benefits agencies

- Evidence-based approaches with trials showing 51% reduction in depression and 31% reduction in anxiety from AI chatbots

The economic case is substantial: AI could save Medicaid $37-67 billion annually through administrative efficiency alone, before accounting for clinical and operational improvements. But Pair Team’s work goes beyond cost savings to reimagine how technology can deliver more human, proactive, and accessible care to vulnerable populations. This represents a fundamental shift from reactive crisis management to proactive support that catches people before they fall.

Introduction: Reimagining Medicaid Through AI

Medicaid faces a fundamental tension. It serves over 70 million Americans, many with complex medical and social needs, yet operates with chronic workforce shortages, fragmented service delivery, and administrative burden that consumes 15% of every healthcare dollar. Beneficiaries struggle with language barriers, behavioral health access, and unmet social needs. Meanwhile, the technology reshaping how people navigate every other part of their lives has barely touched Medicaid.

Pair Team, working with researchers at Stanford, Berkeley, and Carnegie Mellon, has published a policy brief in NPJ Digital Medicine that lays out a different path forward. The paper introduces “Precision Benefits,” a framework for using AI to deliver the right support to the right person at the right time, before avoidable health and social deterioration occurs. This isn’t about incremental efficiency gains. It’s about fundamentally reimagining how a safety net program can function when it has AI-powered infrastructure at its core.

The timing matters. AI’s ability to process language, reason, and respond has matured rapidly. Large language models can now conduct natural conversations across languages and contexts. Frameworks like retrieval-augmented generation (RAG) ground AI responses in verified data rather than hallucinations. The Model Context Protocol (MCP) lets AI systems connect to external databases and take real-world actions through APIs. Together, these technologies form the foundation for digital assistants that can actually help people navigate complex systems rather than frustrate them with rigid decision trees.

Pair Team’s work comes at a policy moment when both efficiency mandates and equity priorities align. H.R.1 requires states to improve Medicaid verification and reduce improper payments while fostering innovation. The economic pressure is real: with potential savings of $37-67 billion annually just from administrative automation, states have both budget incentive and political cover to try new approaches. The question isn’t whether AI will come to Medicaid, but whether it will be built responsibly and deployed in ways that actually serve beneficiaries rather than just cutting costs.

The Scale of the Opportunity

The numbers tell a stark story. Medicaid spending approaches $700 billion annually serving over 70 million Americans. Administrative overhead consumes roughly 15% of every healthcare dollar. Apply AI automation to that administrative burden and you’re looking at $37-67 billion in potential annual savings, scaling proportionally from estimates that AI could reduce overall U.S. healthcare spending by 5-10% or up to $360 billion.

But Pair Team’s brief makes clear that administrative efficiency is just the starting point. The real opportunity comes from clinical and operational improvements: reducing redundant testing, optimizing resource allocation, improving care coordination, preventing avoidable ER visits and hospitalizations. When you catch someone before they fall into crisis, the savings compound exponentially. A few hundred dollars of timely intervention can prevent thousands in emergency care costs.

The access gaps are massive and well-documented. About 25% of Medicaid beneficiaries speak a language other than English at home. Wait times for behavioral health appointments can exceed 30 days in many areas, with some regions facing even longer delays. Community health workers and care coordinators are overwhelmed and under-resourced. Transportation to appointments remains a persistent barrier. These aren’t problems you solve by hiring more staff because the workforce shortages are structural, especially in long-term services and supports (LTSS) and home and community-based services (HCBS).

What makes Medicaid particularly compelling as a testing ground is that it serves populations with the highest needs and the most fragmented care. Multiple chronic conditions are the norm rather than the exception. Social determinants of health like housing instability, food insecurity, and lack of transportation directly impact health outcomes. Language barriers complicate every interaction. Eligibility requirements are complex and vary state by state. If you can build AI systems that work in this environment, delivering personalized support to people navigating all these challenges simultaneously, then those systems will work anywhere.

What Digital Assistants Actually Mean

Pair Team uses the term “AI-powered digital assistants” deliberately to distinguish what they’re building from basic chatbots. These are agentic systems designed to accomplish tasks on behalf of users with a degree of autonomy. They can initiate workflows, interact with multiple systems, and support care navigation based on user input or internal reasoning. The difference between a rules-based chatbot and what Pair Team envisions is the difference between a phone tree and a human assistant who actually understands your situation and can help you solve problems.

The technical foundation is what makes this possible now when it wasn’t a few years ago. Large language models provide natural conversation and general-purpose reasoning that can handle ambiguity and context. RAG connects those models to verified knowledge sources and maintains memory across interactions so the assistant remembers previous conversations. MCP provides the connective tissue to external systems, letting AI assistants check eligibility, schedule appointments, coordinate transportation, or trigger alerts to care teams by calling the appropriate APIs. Together, these technologies enable digital assistants that don’t just answer questions but actually help people navigate complex systems.

Picture what this looks like in practice. A Medicaid member texts in Spanish asking about their upcoming cardiology appointment. The digital assistant confirms the appointment in the provider’s EHR, checks whether transportation is arranged, sees that it’s not, coordinates with the state’s non-emergency medical transportation broker to schedule a ride, sends confirmation back to the member with pickup details, and adds a reminder to their calendar. All automated, all in Spanish, no human intervention unless something breaks. The member gets a seamless experience even though the backend involves integration across multiple disconnected systems.

Or consider a more complex scenario that Pair Team highlights. A community health worker at a food bank is helping someone who can’t afford their diabetes medication. The worker asks the digital assistant (via voice) to check Medicaid enrollment status, discovers the person is eligible but not enrolled, walks through a streamlined application right there, schedules a primary care appointment for next week, arranges transportation, and flags the case for pharmacist outreach about medication assistance programs. The assistant handles all the backend integration across eligibility systems, provider scheduling, transportation coordination, and care management platforms. What would normally take multiple phone calls, long hold times, and several days of back and forth happens in one conversation.

This is what Pair Team calls the “embedded staff” model. Instead of digital assistants being a member-facing portal that nobody uses, they become infrastructure that enhances everyone who touches Medicaid. The food bank employee, the 211 crisis line operator, the school nurse, the housing case manager. They all get augmented by AI that knows how to navigate the complexity of healthcare and social services on behalf of the people they’re trying to help.

The multilingual, 24/7 availability aspect can’t be overstated. An AI that can conduct fluent conversations in Spanish, Mandarin, Vietnamese, Arabic, and dozens of other languages removes barriers that interpretation services can’t fully solve. It’s not just translation, it’s native conversation that understands cultural context and adapts tone appropriately. And because it’s always available, members can get help at 2am when they remember a question or 6pm after work when they finally have time to deal with health stuff.

Precision Benefits: The Core Innovation

Here’s where Pair Team’s vision gets really interesting. The concept of Precision Benefits represents a fundamental shift in how safety net programs can function.

Healthcare has had predictive analytics for years. Every health plan has some kind of risk stratification model that identifies members likely to have an ER visit, hospital admission, or chronic disease exacerbation. The problem is these predictions mostly just sit in reports. Maybe a care manager gets a list of high-risk members to call, but they’re overwhelmed, under-resourced, and calling people who may or may not answer. By the time human outreach happens, if it happens at all, the window for intervention has often passed. The insight was there, but nothing happened with it.

Precision Benefits flips this model completely. Instead of prediction leading to a manual workflow that may or may not execute, prediction triggers automated, personalized intervention at the exact moment it can make a difference. The AI doesn’t wait for a care manager to review a dashboard and add a task to their queue. It acts immediately because it can, it’s software running 24/7 with zero marginal cost per interaction. This is the key insight that makes Pair Team’s approach different from traditional care management.

The example in the paper is perfect because it’s so common. Single mother, working hourly job, juggling caregiving responsibilities, managing diabetes and hypertension. Her own mother gets sick, she misses work shifts to help, income drops, she starts skipping her own medical appointments and rationing medication to make rent. Blood sugar climbs, blood pressure spikes, nobody notices the warning signs until she lands in the ER. Health destabilized, job at risk, eviction looming. This cascade happens all the time in Medicaid.

But here’s the thing. That missed prescription refill? That no-show appointment? Those are observable signals that predictive models can flag in real time. Without automated intervention, they’re invisible until the crisis hits. With Precision Benefits, the AI catches the missed refill immediately, reaches out via the member’s preferred channel (text, voice, whatever), understands through conversation that money’s tight, offers to connect them with emergency assistance programs, schedules a telehealth visit to adjust medications if needed, and follows up to make sure they picked up the prescription. Total interaction cost is basically nothing. Potential savings from preventing an ER visit and all the downstream consequences? Easily thousands of dollars, plus the immeasurable human impact of keeping someone stable and housed.

This is what Pair Team means by delivering the right resource, in the right dose, at the right time, before breakdowns become emergencies. It’s precision medicine but applied to the entire continuum of health and social needs. Just as clinicians adjust medications based on patient response, Medicaid can use AI to titrate support, shifting from reactive crisis management to responsive, proactive care.

For prevention, these tools can flag early risks like missed screenings, rising BMI trends, or social isolation patterns, then offer nudges or referrals to local programs before problems develop. For chronic disease management, AI can support daily stability: reminding someone with asthma to refill inhalers before they run out, track symptoms to identify patterns, stay indoors during high pollution days, and connect with their doctor if control is worsening. The system adapts to individual needs and circumstances, providing more intensive support when someone’s struggling and lighter touch when they’re stable.

Understanding the Barriers to AI Adoption

Pair Team’s brief doesn’t sugarcoat the challenges, and it’s worth understanding why AI adoption in Medicaid has been slow despite obvious opportunities. The barriers are real and solving them requires more than just good technology.

First, Medicaid has historically been hesitant to fund emerging innovations even when cost-saving potential is clear. States are risk-averse, budgets are perpetually tight, and political incentives favor avoiding mistakes over taking chances on new approaches. Even when pilots show clear ROI, scaling is slow because procurement is fragmented and every state wants to customize everything. The lack of clear reimbursement pathways and financial incentives discourages investment in AI tools. This is compounded by Medicaid’s regulatory complexity and state-by-state variation in delivery models, from managed care to fee-for-service, requiring coordination across multiple levels of government.

Second, fairness and safety concerns are paramount. AI models trained on commercial insurance data will not generalize well to Medicaid’s socioeconomically and racially diverse populations. The demographics are different, the social determinants are different, the disease burden is different. Using training data that doesn’t represent the population you’re serving risks biased predictions and unequal care. Generative AI introduces additional risks like hallucination (making up information), inappropriate tone, and overgeneralization if not properly constrained.

Pair Team addresses these concerns head-on with specific mitigation strategies. RAG grounds outputs in verified data rather than allowing the model to generate unconstrained responses. Synthetic data can augment underrepresented populations in training sets without privacy risks. Reinforcement learning from human or expert feedback aligns model behavior with fairness and empathy standards. Rigorous pre-deployment testing using LLM-as-a-judge evaluation frameworks can ensure transparency, reliability, and fairness before integration into care. These aren’t theoretical solutions, they’re practical approaches that can be implemented today.

Third, workforce displacement concerns deserve attention even though the framing is often wrong. Medicaid already faces a structural mismatch between service demand and workforce supply, especially in LTSS and HCBS where workforce shortages are severe and getting worse as the population ages. AI should be viewed not as replacing human workers but as augmenting human capacity. By automating routine tasks, triaging cases, and freeing professionals to focus on hands-on, relationship-based care, AI can actually improve job quality while expanding access. A care coordinator who spends half their day on paperwork and phone tag can spend that time building relationships with members instead if the administrative tasks are automated.

Fourth, technology capacity poses a real constraint. Many Medicaid agencies lack the in-house expertise and infrastructure to deploy and govern AI tools effectively. Fragmented data systems and limited interoperability hinder integration with healthcare and social service networks. Different states use different eligibility systems, different managed care organizations use different care management platforms, providers use different EHRs. Getting these systems to talk to each other is hard, expensive, and often requires negotiating data sharing agreements that raise privacy concerns. Trust is also a barrier, with justified fears that data sharing could lead to surveillance or be used in ways that harm beneficiaries.

Addressing these challenges requires investment in secure, interoperable systems alongside transparent governance and genuine community engagement. You can’t just deploy AI tools and hope people trust them. You have to build trust through transparency about how the systems work, clear policies about data use and privacy, meaningful oversight mechanisms, and demonstrated commitment to fairness and accountability.

The Regulatory Reality

One of the most valuable contributions in Pair Team’s brief is clarity about FDA regulation, which often creates anxiety for healthcare AI developers. The reality is more favorable than many assume, especially for the use cases Pair Team is focused on.

The key distinction is whether the AI is making patient-specific diagnostic or treatment decisions that replace clinical judgment, or whether it’s supporting navigation, coordination, and logistics. Digital assistants that handle appointment scheduling, eligibility verification, care navigation, reminders, and basic member education fall outside FDA jurisdiction under the 21st Century Cures Act because they don’t diagnose or treat medical conditions.

Even more sophisticated tools can qualify as non-device Clinical Decision Support (CDS) if they meet four criteria: they’re healthcare provider-facing (not direct to consumer making medical claims), they show their inputs and reasoning transparently, they support rather than replace clinical judgment, and they allow clinicians to independently review and verify recommendations. This is an important safe harbor that Pair Team’s brief highlights.

So if you’re building an AI that reminds a member with asthma to refill their inhaler, tracks symptoms through simple questions, and suggests they message their doctor if things are getting worse, that’s not a medical device. If you’re building an AI that reviews home monitoring data and tells a clinician “this patient’s asthma control is poor based on these usage patterns, consider stepping up therapy,” that could be non-device CDS if designed to meet the four criteria. If you’re building an AI that automatically adjusts medication dosing without physician review, that’s definitely a device and requires premarket review.

The smart strategy Pair Team implicitly recommends is to intentionally scope products to non-device functions initially. Get to market, prove value, build trust. Then layer on clinical decision support features designed to meet CDS criteria. If future versions need patient-specific diagnostic or therapeutic capabilities that cross into device territory, FDA has Predetermined Change Control Plans (PCCPs) that allow companies to define in advance how models will evolve, what validation methods will be used, and how performance will be monitored. Submit the PCCP with initial review and you can iterate the model without repeated submissions, which is crucial for AI systems that need to adapt and improve over time.

This regulatory pathway is actually quite workable for the vision Pair Team lays out. Most of what Precision Benefits requires (care navigation, eligibility support, transportation coordination, connection to community resources, appointment reminders, medication refill prompts) falls clearly outside device regulation. The parts that touch clinical care can be designed as non-device CDS with appropriate safeguards. This is a much clearer and faster path than trying to get device approval for every feature.

There are also state-level AI governance considerations. Some states are passing laws requiring bias audits, transparency reports, and human review for automated decisions that affect people’s access to services. These requirements are generally aligned with best practices anyway. If you’re building responsibly with bias testing, human oversight, and clear explanations, compliance shouldn’t be a heavy lift. In fact, Pair Team’s brief proposes that Medicaid AI systems should include clear fallback to human support pathways, transparent explanation of reasoning, user feedback loops (like “I’m not satisfied, take me to a human”), and ongoing trust monitoring through surveys and metrics.

Pair Team’s Implementation Strategy

The brief lays out a comprehensive strategy for how AI adoption in Medicaid should actually happen, with specific recommendations for federal and state action. This is where Pair Team moves from vision to practical implementation.

At the federal level, the Centers for Medicare and Medicaid Services (CMS) needs to take a leadership role. Pair Team proposes CMS launch an AI Innovation Challenge to fund pilot projects that enhance administrative efficiency, streamline eligibility, and improve care coordination. This would provide the financial resources and political legitimacy for states to experiment without bearing all the risk themselves. A State Medicaid Director’s Letter outlining best practices would offer implementation guidance, addressing practical questions like whether payments for digital assistants should count under Medical Loss Ratio versus Administrative Loss Ratio, how to structure contracts with AI vendors, and how to evaluate effectiveness.

The brief calls for states to create AI Centers of Excellence, modeled in part after the Centers of Excellence in Regulatory Science and Innovation (CERSI) that exist for medical devices and pharmaceuticals. These would serve as dedicated hubs bringing together policymakers, technology developers, healthcare providers, and health plans to evaluate AI in real-world Medicaid settings. Priorities could include predictive analytics for early detection of health and social risk, AI-assisted care navigation and outreach, and public-private partnerships for scaling innovation. These centers would also serve the crucial function of developing shared standards and evaluation frameworks so every state doesn’t have to reinvent the wheel.

Pair Team emphasizes that AI tools should undergo rigorous bias audits and benchmarking to ensure fairness in access and outcomes. Training models on Medicaid-representative data improves accuracy, but for large language models, fairness also depends on safeguards like human-in-the-loop review for sensitive decisions, output constraints that prevent harmful or biased responses, and prospective monitoring for differential impacts across populations. The brief proposes pilot user testing that emphasizes acceptability and trust measures before scaling, recognizing that technical performance alone doesn’t guarantee real-world success.

Beyond pilot programs and operational improvements, Pair Team argues that CMS should signal a bold ambition: to establish Medicaid as a leading example of an AI-first public service. This isn’t just about making the current system more efficient, it’s about demonstrating what’s possible when public programs embrace intelligent technologies to better meet individual needs. The implications extend far beyond Medicaid to Medicare, other safety net programs, and public services globally.

The Path Forward

What Pair Team is building represents more than incremental improvement to an existing system. It’s a fundamental reimagining of how safety net programs can function when designed around AI infrastructure from the ground up.

The Precision Benefits framework addresses a core limitation of traditional care management: the gap between identifying risk and actually intervening before crisis occurs. Predictive models have been around for years but without automated intervention capabilities, those predictions sit unused. By combining prediction with real-time, personalized, automated outreach and support, Pair Team’s approach closes that gap. The AI can act on signals immediately, at scale, with zero marginal cost per interaction.

The embedded staff model solves another fundamental problem, which is that Medicaid beneficiaries interact with dozens of different touchpoints (clinics, pharmacies, community organizations, benefits agencies, schools, housing programs) and none of those systems talk to each other effectively. By making AI assistants available to everyone who touches Medicaid, not just as a member-facing portal, Pair Team creates connective tissue across a fragmented ecosystem. The food bank worker, the school nurse, the 211 operator all get augmented with tools that know how to navigate complexity on behalf of the people they serve.

The focus on 24/7 multilingual support addresses access barriers that workforce expansion alone can’t solve. You can’t hire enough bilingual care coordinators to cover every language, every time zone, every moment when someone needs help. But you can build AI that operates around the clock in any language with cultural competence built in. This isn’t about replacing human connection, it’s about making sure everyone has access to basic support regardless of when they need it or what language they speak.

The emphasis on behavioral health is particularly important given the massive access gaps in this area. The Therabot study that Pair Team cites showed remarkable results: 51% reduction in depression scores, 31% reduction in anxiety, 19% reduction in eating disorder risk from a generative AI mental health chatbot. Participants reported feeling meaningful connections with the AI. This doesn’t replace therapy or psychiatric care, but it provides support that many Medicaid beneficiaries simply can’t access today because of provider shortages and wait times.

What makes Pair Team’s work distinctive is the combination of technical sophistication, practical implementation focus, and genuine commitment to responsible development. The brief doesn’t handwave away concerns about bias, fairness, privacy, and safety. It addresses them directly with specific mitigation strategies. It acknowledges workforce concerns while reframing AI as augmentation rather than replacement. It provides regulatory clarity that makes the path forward navigable rather than daunting.

The economic case is compelling with potential savings of $37-67 billion annually from administrative efficiency alone. But the human case is what really matters. Medicaid serves people who face multiple overlapping challenges: chronic health conditions, social determinants of health, language barriers, transportation issues, housing instability, food insecurity. The current system is designed for these challenges to be addressed separately, if at all, by disconnected programs and providers who don’t communicate effectively. Pair Team’s vision is a system where AI coordinates across all these domains to catch people before they fall, providing the right support at the right time to keep everyone on an upward trajectory.

The work that Pair Team is doing extends beyond building a product. They’re establishing a model for how public programs can leverage AI to better serve vulnerable populations. That model has implications for Medicare’s 65 million beneficiaries, for safety net programs serving hundreds of millions globally, and for public services far beyond healthcare. This is infrastructure work in the truest sense. Not the flashy stuff that makes headlines but the foundational systems that determine whether society functions equitably or leaves people behind.

Pair Team is building the rails for an AI-first safety net, and if they succeed, the implications will reshape how we think about the relationship between technology and social support. The question isn’t whether AI will transform public programs. It’s whether that transformation will be done thoughtfully, responsibly, and in ways that actually serve the people who need help most. Pair Team’s work in Medicaid shows what’s possible when you get that approach right.