Healthcare’s Future: AI Agents in Action Throughout a Patient’s Journey

We've outlined some quick wins AI agents could bring to healthcare’s record-keeping burdens—now let’s get imaginative. What if teams of specialized AI agents streamlined every step of a medical visit? In big steps (there are many between), resolving a health issue goes something like this:

1. You feel something's off.
2. You book an appointment.
3. Nurses and doctors diagnose and treat you.
4. Your clinic bills you and/or your insurance.
5. You (and/or your insurance) pay your doc.
6. You follow some protocol or receive a treatment (and hopefully improve).

Let’s now sketch out a hypothetical yet feasible story of how AI agents might optimize each of these steps.

Atlas felt off for weeks: his joints ached, especially in the mornings, and an odd blotch on his face worsened after sun exposure.

A nagging fatigue crept in most afternoons too.

Frustrated, he swiped open a new healthcare app. Its voice interface asked him a few questions, and Atlas described his symptoms to what felt like a single, attentive assistant.

After an Identity Verification Agent—akin to Hippocratic AI's identity verification system, which precisely confirms dates of birth and other numerical data—ensured it was actually speaking with Atlas, the difficult detective work began.

With emotional machine intelligence, the Orchestrator Agent actively listened, analyzing Atlas' tone, pace, and speech patterns to gauge his emotional state—adjusting its responses to ease his anxiety. Behind the scenes, the Orchestrator Agent decomposed the overall problem (identifying Atlas' malady) into smaller tasks and assigned each of these amongst a team of specialist AI agents that were powered by fine-tuned LLMs, voice AI, RAG systems, medical knowledge graphs and ontologies, and more.

AI-driven speech-to-text transcribed his every word, and a similar text-to-speech model responded intelligently.

The Orchestrator Agent didn’t just hand out tasks. It used a Hierarchical Task Network (HTN) planner, breaking down Atlas' medical case into increasingly specific subtasks. Each specialist agent (Rheumatology, Dermatology, and Immunology) maintained its own Belief-Desire-Intention (BDI) model, dynamically updating beliefs (e.g., "Atlas' symptoms are consistent with an autoimmune disorder"), generating desires (e.g., “Confirm if lupus is correct”), and forming intentions (e.g., “Order inflammatory marker tests”).

Whenever possible, the Orchestrator Agent routed partial initial outputs from Atlas and agents to other agents in near parallel time. This sped up the preliminary diagnosis enough to keep the conversation with Atlas natural and fluid. But sometimes the agents slowed down, switching to sequential, turn-taking interactions within a multiagent debate. 

When the Dermatology Agent flagged the butterfly-like rash in Atlas' photo, it formalized an argument using premises (the rash’s distinct shape and distribution), rules (classic dermatological criteria), and a tentative conclusion: this looks like lupus. The Immunology Agent quickly interjected, pointing out that photosensitive rashes also appear in dermatomyositis and mixed connective tissue disorders—so they ought to seek more data (blood tests, antibody screens, etc.) before locking in on “lupus.” Meanwhile, the Rheumatology Agent factored in Atlas' morning joint pain and potential inflammation. The trio sparred—much like human doctors do—until converging on a preliminary lupus diagnosis with 87% confidence.

Alas was sheltered from all these agents' diagnostic chatter to avoid overwhelming him with uncertain preliminary diagnoses. But the AI agents’ reasoning and dialogue were stored for audit by doctors, lawyers, engineers, and regulators.

With a preliminary diagnosis in hand, Provider Matching Agents got to work. They scrutinized everything from insurance networks and medical credentials to practice locations and communication styles to find Atlas a doctor. The Provider Matching Agents worked much like Notable Health's patient-matching technology, which connects patients with providers and streamlines pre-visit paperwork. At North Kansas City Hospital, Notable's system boosted patient satisfaction to 95% within 13 weeks, increased digital registration completion to 70%, and cut no-shows by 23%—a glimpse into how AI-driven matchmaking can improve early patient care.

How’d the Provider Matching Agent surface the best doc for Atlas? Instead of a simple matching algorithm, it used a contract net protocol—a negotiation mechanism where a “manager” agent defined its clinician requirements and sent requests for proposals to multiple “contractor” agents, each representing different clinical providers. These contractor agents then “bid” for Atlas’ case, presenting their clinicians’ qualifications, specialties, availability, and relationships with relevant specialists—or declining if they weren’t a fit.

The “manager” agent then dynamically scored each proposal using a multi-criteria decision-making framework, balancing sometimes conflicting factors like:

• Clinical expertise: Finding specialists who excelled not just in lupus treatment but also in managing complex cases requiring immunologist coordination
• Logistics: Balancing trade-offs between providers’ experience levels and their accessibility (in location and wait times)
• Network dynamics: Ensuring strong relationships with other specialists that Atlas might require while maintaining scheduling flexibility
• Patient experience: Matching communication styles and appointment flexibility with clinical qualifications

The system went beyond basic rankings, modeling how each doctor would function within a broader care network. Could they seamlessly coordinate with an immunologist if needed? Would they fit into a long-term treatment strategy? When providers showed promise but had scheduling conflicts, they could counter-propose alternatives—enabling dynamic optimization of these trade-offs in seconds rather than days.

Dr. Rodriguez emerged as an ideal choice: a rheumatologist with lupus expertise, positive patient outcomes, and only a single malpractice case (which she successfully defended). Her office was only a fifteen-minute detour from Atlas' commute, and she had an established working relationship with a top immunologist, should Atlas' require one.

Atlas approved. With a tap, a Scheduling Agent—akin to Notable’s no-show reduction system, which cut missed appointments by 32% at Good Shepherd Rehabilitation—found an optimal morning slot, even factoring in Atlas' energy levels and local traffic patterns.

Once Atlas scheduled his appointment, a team of AI agents streamlined the usual administrative headaches—much like Notable's digital intake system, which doubled pre-visit digital completion rates to 50%, freed up 13 full-time employees' worth of capacity at Montage Health, saving $2 million annually.

Documentation Agents pulled historical records from Atlas' healthcare providers—an impressive accomplishment given his 7 moves in the last 4 years. It stitched together a coherent health history from several different providers’ records, something very difficult to pull off before AI agents. While reviewing the data, it flagged an old mention of joint pain from two years ago and asked Atlas to fill in missing details before passing the information to Dr. Rodriguez for review. This mirrors Notable’s chart review system at Castell, where AI scaled chart processing from 100 to 3,362 charts daily. Pre-AI, each chart required around three minutes of manual clicking through EHR systems, plus additional time to review scanned reports, faxes, and attachment uploads, and to call patients.

Next, Lab Analysis Agents—similar to Hippocratic AI's labs and vitals reference range agents—reviewed Atlas’ past tests within the context of his age, gender, and medical history. When Atlas mentioned taking over-the-counter (OTC) supplements for joint pain, another agent, equipped with capabilities similar to Hippocratic AI's OTC contra-indications-aware Polaris 2.0, quickly checked for potential drug interactions. It flagged one that could exacerbate his lupus symptoms—a level of domain-specific reasoning that general LLMs lack.

This wasn't a lone agent crunching numbers—it was a distributed problem-solving network. A group of Lab Analysis Agents collaborated through a blackboard architecture—a shared workspace where different specialist agents posted their findings, cross-checked their insights, and collaboratively refined their conclusions. For example, to analyze Atlas' Erythrocyte Sedimentation Rate (ESR), an inflammation marker relevant to lupus diagnosis, the following specialist agents pitched in:

• Normalization Agent standardized values.
• Reference Range Agent adjusted ranges based on Atlas' demographics.
• Temporal Analysis Agent evaluated Atlas’ ESR trends over time
• Context Agent correlated results with Atlas’ symptoms.
• Meta-Analysis Agent found broader patterns across all of Atlas’ labs.

Each agent subscribed to updates pertinent to its domain, forming an interconnected analytical network. When they confirmed that Atlas’ ESR test indicated elevated inflammation, a Lupus Specialist Agent dug deeper for condition-specific indicators. This collaborative framework ensured robust error checking—if one agent erred, others might detect and correct the mistake.

Before the visit, AI agents also scheduled a few routine labs, finding a convenient location not far from Atlas' apartment. Every pre-visit task—from insurance authorizations to medical history forms—was streamlined and verified, similar to Notable's prior authorization AI automation at Care New England, which cut turnaround times by 80% and saved 2,841 staff hours, allowing the organization to shuffle staff to more valuable tasks.

Atlas was spared from the usual gnawing paperwork. AI agents took over most of his admin load and kept him informed via his preferred channels—email and SMS—at intervals designed to prevent notification fatigue.

Despite a crowded clinic with three different patients named Atlas, a Biometric Verification Agent kept the correct Atlas’ file open on Dr. Rodriguez’s monitor at all times. Just before the appointment, when she asked for “Atlas’ lupus notes,” the agent confirmed the correct chart—preventing a potentially dangerous mix-up, a mistake that sadly still happens even with modern EHRs.

During Atlas' appointment, a Documentation Agent transcribed and structured the conversation in real time—a capability powered even today by tools like Nova-3. Beyond simply transcribing the doctor-patient conversation, it built and stored a dynamic knowledge graph, mapping Atlas’ symptoms, treatments, and concerns—while preserving a clear "who-said-what" timeline. For example, it:

1. Added a "fatigue" node in the knowledge graph
2. Linked "fatigue" to "sun exposure" with a "trigger" relationship
3. Tagged the temporal pattern ("after")
4. Assigned a confidence score for this relationship
5. Cross-referenced lupus literature for similar patterns
6. Found possible drug interactions (some meds might have worsened Atlas’ sun sensitivity)

Notable implemented a similar approach at Austin Regional Clinic, where AI agents fetched patient data via pre-visit questionnaires, converted it into narrative clinical summaries, and then handled all the manual clicking, dropdown selecting, and data entry in Epic's EHR that doctors normally manually slog through. This cut documentation time by half, letting Austin Regional’s clinicians better focus on their patients.

Next, Symptom Assessment Agents used hierarchical planning to refine hypotheses. Each symptom triggered multiple possibilities in a belief network. When Atlas mentioned sun sensitivity, for example, they:

1. Spawned multiple hypothesis child agents, each investigating a different potential cause
2. These child agents gathered evidence (some checking autoimmune conditions, others allergic reactions)
3. They “debated” their findings
4. Higher-confidence hypotheses spawned specialized grandchild agents
5. An "exploration budget" occasionally probed unlikely but severe conditions

This spawning, multi-agent strategy mimics how a specialist team brainstorms diagnoses but explores further and faster.

When Atlas mentioned recent blood work he’d had run at a different urgent care clinic, Lab Analysis Agents instantly fetched and integrated those results into Dr. Rodriguez’ EHR system, noting Atlas’ historically elevated ESR tests, a key lupus indicator. This saved weeks of phone tag.

Ordering new tests was equally simple: Rather than clicking and scrolling across endless EHR menus, Dr. Rodriguez simply said, “Let’s get the comprehensive lupus workup,” and Lab Ordering Agents parsed the conversation context, pulled from condition-specific knowledge bases, and auto-prepped orders for ANA, anti-dsDNA, CBC, and more—even checking Atlas’ insurance coverage for each. All this freed Dr. Rodriguez from much typing and clicking, letting her offer more empathy during Atlas’ already stressful visit.

Other agents absorbed the structured data from the Documentation Agent. When Dr. Rodriguez prescribed hydroxychloroquine, a Clinical Decision Support Agent factored in Atlas’ sun sensitivity and possible OTC supplement interactions, personalizing doses and precautions.

After Dr. Rodriguez placed the order, Monitoring Agents tracked every step from blood draw to analysis. If a test was delayed by equipment downtime, they automatically rerouted samples to another lab, notifying the relevant staff. To do this, the Monitoring Agents formed a hierarchical multi-agent system mirroring the lab's physical workflow. Each stage had a specialist agent. They used partial Global Planning (PGP) to resolve real-world constraints and ensure priority cases stayed on schedule. For example, when Atlas' ANA test was delayed:

1. An Equipment Agent detected the issue
2. It shared its local plan with other lab agents
3. All agents merged their plans into a PGP
4. They used constraint satisfaction to find alternate routes for urgent samples
5. A negotiation protocol balanced competing patient priorities

These agents also maintained "mental models" of human lab technicians' schedules, factoring in shift changes, lunch breaks, and working styles, making reroutes efficient.

When Dr. Rodriguez prescribed hydroxychloroquine for Atlas', Alert Management Agents avoided a flood of irrelevant warnings. Instead, they surfaced only two critical interactions: one with Atlas' antacid and another tied to his sun sensitivity. Dr. Rodriguez easily adjusted the medication schedule and added UV precautions to his care plan.

Unlike static-threshold systems, these Alert Management Agents used a Dempster-Shafer-based belief-revision model—more nuanced than simple probabilities—to maintain dynamic belief networks about what Dr. Rodriguez needed to know. These AI agents decided not just what to alert about, but when and how to alert. To do this, they used a meta-reasoning layer comprised of the following:

1. Urgency Assessment: Using temporal logic to time alerts properly
2. Context Awareness: Estimating Dr. Rodriguez's current cognitive load
3. Information Bundling: Grouping related alerts
4. Delivery Optimization: Tailoring alerts to the best modality (visual, audio, haptic)

And for medication interactions, these agents:

1. Calculated "belief masses" of potential severities
2. Combined evidence from multiple knowledge bases using Dempster's rule
3. Applied a utility function to weigh the risk of missing an interaction against interrupting Dr. Rodriguez
4. Used reinforcement learning to refine their interruption strategy based on her past responses

By bundling related alerts and tracking her cognitive load, the Alert Management Agents kept Dr. Rodriguez informed without overwhelming her. These agents’ reasoning was transparent—if Dr. Rodriguez questioned a particular agent’s recommendation, she could trace its logic.

A Hospital Policy Agent ensured all recommendations followed hospital and department guidelines. And as Atlas’ appointment wrapped up, an AI Summarization Agent distilled Dr. Rodriguez’s notes into a concise, actionable summary—balancing medical precision with patient-friendly language.

After diagnosing Atlas, a Prescription Management Agent helped Atlas navigate his new medication: identifying it, pronouncing it, and tracking doses. Meanwhile, a Care Plan Management Agent built a treatment schedule tailored to Atlas’ daily routine. It also closed care gaps (i.e., missed preventive and chronic management opportunities) by sending automated text reminders for follow-ups (after learning that Atlas recently started responding more to texts than emails). ThedaCare used similar AI agents to close 963 care gaps in 3 months with 89% patient satisfaction. Rather than optimizing for a single metric, the Care Plan Management Agent used multi-objective reinforcement learning, drawing on data from thousands of lupus patients. Its reward function weighed factors like:

• Maximizing medication adherence
• Reducing symptom severity
• Improving quality of life
• Lowering costs
• Boosting patient engagement
• Mitigating complication risks

When Atlas missed a dose, the Care Plan Management Agent didn't just blast out a generic reminder. It:

1. Updated its belief state on Atlas' adherence patterns
2. Simulated the impact of a missed dose using Monte Carlo methods
3. Tweaked its communication strategy via a contextual bandit algorithm (balancing exploration and exploitation)
4. Sent a personalized message tied to Atlas' symptoms and treatment goals

Prior authorizations steal physicians’ time. In a 2021 poll, ~81% of responding medical groups reported increased prior authorizations, with physicians devoting ~16 hours per week to these tasks. When Atlas’ insurer insisted on a cheaper alternative to the hydroxychloroquine that Dr. Rodriguez prescribed, a Claims Resolution Agent—similar to Notable Health's prior authorization agents at Fort HealthCare, which hit a 91% success rate (for prior authorization requests)—quickly retrieved proof that Atlas had already tried the cheaper option with poor results. It fired off the evidence straight to the insurer—sparring on Dr. Rodriguez’s and Atlas’ behalf to spare them the nasty, drawn-out appeals gauntlet that many practices increasingly endure.

As Atlas adjusted to his routine, a network of AI agents optimized his care. Progress Tracking Agents monitored his symptoms via data from Atlas’ smartphone, doctor visits, and labs—similar to Hippocratic AI's longitudinal lab analysis agent, which improves chronic care coaching by interpreting trends across sequential lab results. Over time, it learned Atlas’ specific flare-up triggers. When his joint pain spiked one week, it informed both him and Dr. Rodriguez that the increased pain correlated strongly with recent weather patterns and increased physical activity.

A Prevention Planning Agent—similar to Notable’s Care Gap Closure AI agents at MUSC Health, which scheduled 1,100 preventive mammogram screenings and found 122 abnormal results without staff involvement—recommended new screenings for Atlas based on emerging lupus research. The Prescription Management Agent kept Atlas on track with his complex regimen, shifting reminders from afternoons to mornings as it detected changes in his response patterns.

A Nutrition Management Agent—like Hippocratic AI's condition-specific menu system—helped Atlas make healthier dining choices. It analyzed menu photos, factored in his recent lab values and nutrient needs, and recommended ideal dishes while flagging problematic ones. When Atlas traveled for work, Care Continuity Agents kept his care consistent across different locations, sending lupus markers tested at out-of-state labs to his normal rheumatologist and broader trends to his normal primary care physician while maintaining HIPAA and security compliance. These agents also helped Atlas find Greek-speaking doctors (his native language), similar to Notable’s multilingual agents at Lowell Community Health Center, which achieved a 96% satisfaction rate and a 98% digital completion rate among 5,000+ intakes from patients with limited English proficiency.

With his secure healthcare app, Atlas could grant temporary access to any clinic or ER—eliminating the weeks-long record requests he once endured. On one trip, a specialist adjusted Atlas’ hydroxychloroquine dosage. Version Control Agents recorded who changed what and why, then notified Dr. Rodriguez and Atlas' pharmacist, staving off the type of documentation errors that had once plagued his treatment schedule.

As Atlas focused on his health, Financial Counseling Agents analyzed thousands of lupus cases to forecast treatment costs for different scenarios, uncovered assistance programs, including a manufacturer's discount for one of Atlas’ medications, and optimized his health savings account contributions.

Payment Processing Agents managed the maze of claims, balances, and assistance programs that Atlas navigated. These agents structured a payment plan aligned with Atlas’ pay schedule, monitored for financial strain, and proactively recommended extra resources when needed. A single, easy-to-read dashboard gave Atlas clear visibility into his healthcare costs—no more wading through insurance legalese. 

Provider-side systems have already reaped similar benefits. MUSC Health, for example, increased their copay collections by $1.7 million using AI agents to automate billing workflows and identify new billable opportunities.

Atlas’ journey illustrates the broad potential for healthcare-related AI agents. Coordinated networks of specialized agents—handling everything from diagnoses to follow-up care—could make medicine more efficient, personal, and accessible.

AI agents might help us with:

• Faster Diagnostics: Human-AI agent collaboration could reduce weeks-long processes to minutes.
• Reduced Administrative Burdens: Automation of routine tasks frees clinicians' time.
• Better Coordination: AI agents can streamline workflows involving insurance, labs, treatments, and care transitions.
• Increased Patient Guidance: Assisting patients in navigating healthcare systems.
• More Face Time: Enabling doctors to focus on patient care.
• Earlier Detection: Using pattern recognition for earlier disease identification.
• Greater Financial Support: Helping patients and doctors with insurance policies, pre-authorizations, and claims for better affordability

Still, many AI agent-assisted medical tasks will likely remain human-in-the-loop for some time. In Atlas' story, for example, Dr. Rodriguez held the final authority over diagnosis and treatment, with AI agents supporting her by analyzing medical data, scanning literature, and completing routine tasks. Even for relatively straightforward, low-risk tasks, like processing health insurance claims, we may still want a culpable human to bless off on AI agents' suggestions.

And now that we’ve caught a glimpse into what the future of healthcare could look like, let’s see what the journey there would entail. In the next part of this series, we’ll discuss the promise and the perils of AI—both what’s already working and the obstacles ahead.