Generative AI in Healthcare: Applications, Use Cases, and the Revenue Cycle Impact

Healthcare has used artificial intelligence for years — predictive models that flag high-risk patients, rules engines that scrub claims before submission, natural language processing that extracts diagnosis codes from clinical notes. But generative AI represents a fundamentally different capability. Instead of classifying, scoring, or flagging, generative AI creates. It drafts clinical documentation from a physician-patient conversation. It writes denial appeal letters tailored to the specific payer's adjudication logic. It generates patient-facing explanations of billing statements in plain language. It produces synthetic training data for rare disease coding models.

The distinction matters because the revenue cycle is, at its core, a documentation and communication process. Every claim is a generated document. Every appeal is a generated argument. Every patient billing interaction is a generated communication. The revenue cycle has always depended on human beings generating these artifacts — and doing so accurately, persuasively, and at scale has always been the bottleneck.

Generative AI doesn't just automate the revenue cycle. It changes what automation means.

The market agrees. Healthcare generative AI spending is projected to reach $22.1 billion by 2032, growing at a compound annual rate of 37%, according to Grand View Research. But investment without understanding is just cost. This article explains what generative AI actually does in healthcare, how it applies to revenue cycle operations specifically, where the real risks lie, and how to evaluate whether a generative AI solution will produce measurable ROI or just impressive demos.

What Generative AI Actually Means in Healthcare (And What It Does Not Mean)

Generative AI refers to models that produce new content — text, images, code, structured data — based on patterns learned from training data. In healthcare, the most relevant generative AI models are large language models (LLMs) that generate text: clinical notes, appeal letters, patient communications, coding suggestions, and structured data extractions.

This is distinct from the other AI categories that healthcare organizations already use.

Predictive AI answers "what will happen." A predictive model scores a claim's denial probability at 73% based on historical patterns. It doesn't write the appeal or fix the claim — it tells you the claim is likely to be denied.

Analytical AI answers "what happened and why." An analytical model identifies that Payer A's denial rate for CPT 99214 increased 340% in Q3 because of a new clinical editing rule. It explains the pattern but doesn't act on it.

Generative AI answers "what should be created." A generative model takes the denied claim, the payer's stated reason, the patient's clinical documentation, the payer's historical adjudication patterns, and relevant regulatory requirements — and drafts a complete appeal letter optimized for overturn probability. It creates the artifact that a human would otherwise spend 30-60 minutes writing.

The practical implication: predictive and analytical AI tell your staff what to do. Generative AI does the work — or produces a draft that reduces human effort by 70-90%.

What Generative AI Is Not

Generative AI is not a search engine, though it can be combined with one. It is not a database query tool, though it can generate structured outputs. Most importantly, it is not inherently accurate. Generative models produce statistically likely outputs, not verified facts. In clinical and financial contexts, this distinction is critical — and it shapes every responsible implementation strategy.

Key Applications of Generative AI in Healthcare

Clinical Documentation

The most mature generative AI application in healthcare is ambient clinical documentation — AI that listens to physician-patient conversations and generates structured clinical notes in real time. Companies like Nuance (Microsoft), Abridge, and DeepScribe have deployed these systems across thousands of provider sites.

The revenue cycle impact is direct. Clinical documentation drives coding, coding drives claims, and claims drive revenue. When documentation is incomplete, codes are missed. When codes are missed, revenue is left on the table. Studies from the American Medical Association show that physicians spend an average of 16 minutes per patient encounter on documentation — time that could be spent with patients. Generative documentation AI reduces that to 2-3 minutes of review time.

But the revenue implications go beyond time savings. Generative documentation systems capture clinical detail that physicians routinely omit when writing notes manually. A physician may document "patient has diabetes" in a hand-written note. The AI, capturing the full conversation, documents "patient has type 2 diabetes mellitus with diabetic chronic kidney disease, stage 3a, currently managed with metformin 1000mg twice daily, last A1C 7.8%." The difference in specificity cascades through the entire revenue cycle:

• HCC coding: The AI-captured note supports HCC 18 (Diabetes with Chronic Complications) rather than HCC 19 (Diabetes without Complication), a risk adjustment factor difference of 0.302 vs 0.104 — nearly 3x the risk score.
• E/M level support: The additional clinical detail supports a higher evaluation and management level, potentially moving a 99213 ($92 Medicare reimbursement) to a 99214 ($133 reimbursement) — a 45% increase per encounter.
• Denial prevention: Complete documentation reduces the likelihood of medical necessity denials by 35-50% because the clinical rationale for services is captured in the note rather than existing only in the physician's memory.

Organizations deploying ambient documentation report average revenue increases of $15,000-$25,000 per provider per year from improved documentation specificity alone.

Coding Assistance

Generative AI coding tools don't just suggest codes — they generate the reasoning chain that connects clinical documentation to code selection. Traditional computer-assisted coding (CAC) systems use NLP to extract keywords and map them to codes. Generative coding AI reads the full clinical context and produces a code set with explanations for each selection.

For example, a traditional CAC system might extract "chest pain" and suggest R07.9 (Chest pain, unspecified). A generative coding AI reads the full note, identifies that the chest pain is described as substernal, exertional, relieved by rest, occurring in a patient with known coronary artery disease — and suggests I25.119 (Atherosclerotic heart disease of native coronary artery with unspecified angina pectoris) with a rationale citing the specific documentation elements that support the code.

The accuracy difference is significant. Traditional CAC systems achieve 60-75% accuracy on initial code suggestion. Generative coding AI achieves 85-92% accuracy because it understands clinical context rather than just extracting keywords. For a coding team processing 500 encounters per day, that accuracy improvement reduces human review time by approximately 40% and catches an estimated $380,000-$720,000 in annual undercoding.

Denial Appeal Letter Generation

This is where generative AI delivers perhaps its most immediate and measurable revenue cycle ROI. Writing denial appeals is one of the most labor-intensive, specialized, and consequential tasks in the revenue cycle. A well-crafted appeal that addresses the specific denial reason, cites relevant clinical documentation, references applicable regulations and payer policies, and presents a persuasive clinical argument can overturn 50-70% of denials. A generic, template-based appeal overturns 20-30%.

Generative AI appeal systems ingest the denied claim, the remittance advice with the specific denial reason code, the patient's complete clinical record, the payer's known adjudication patterns, and applicable CMS guidelines or state regulations. The output is a complete, payer-specific appeal letter that would take a skilled denial analyst 30-60 minutes to write manually.

The economics are compelling:

Metric	Manual Appeals	Generative AI Appeals
Time per appeal	30-60 minutes	2-5 minutes (including human review)
Appeals processed per FTE per day	8-12	40-60
Overturn rate	35-45%	55-68%
Cost per appeal	$25-$50	$4-$8
Annual capacity per FTE	2,200-3,000	10,000-15,000

For a health system managing 25,000 denials per year, switching from manual to generative AI-assisted appeals recovers an estimated $2.8-$4.2 million in additional overturned revenue annually while reducing appeal staff requirements by 60-70%.

Patient Communication

Patient billing communication is a generative problem. Every patient's financial situation is different — different insurance plans, different deductibles, different out-of-pocket maximums, different services rendered. Generating personalized, plain-language explanations of what patients owe and why has historically required either expensive human interaction or generic, confusing form letters.

Generative AI produces patient-specific billing explanations that translate complex EOBs into language patients actually understand. "Your insurance covered $4,200 of the $5,800 total charge. Your remaining responsibility of $1,600 reflects your $500 annual deductible (which had $320 remaining) and your 20% coinsurance on the remaining $6,400 in covered charges. Based on your payment history, we've set up a suggested payment plan of $267/month for 6 months."

Organizations using generative patient communications report 23-35% improvement in patient payment rates and 40-55% reduction in billing-related call center volume. For a health system with $30 million in annual patient responsibility, a 25% improvement in collection rate represents $7.5 million in additional revenue.

Care Gap Identification and Outreach

Generative AI analyzes patient records to identify gaps in care — missed screenings, overdue wellness visits, lapsed chronic disease management — and generates personalized outreach communications. The revenue impact comes from increased visit volume and improved quality measure performance under value-based contracts.

A generative system might identify that a diabetic patient is overdue for an eye exam and A1C check, then generate a personalized message referencing the patient's specific conditions and explaining why these services matter for their health. The message is clinically accurate, personally relevant, and actionable — attributes that drive 3-5x higher response rates compared to generic recall notices.

Generative AI in the Revenue Cycle: A Complete Use Case Map

The revenue cycle is a sequence of documentation-intensive, communication-heavy processes. Generative AI touches nearly every stage.

Pre-Service

• Eligibility verification communications: Generating patient-facing notifications about coverage limitations, required authorizations, and estimated costs before service delivery.
• Prior authorization requests: Drafting authorization requests with supporting clinical documentation tailored to each payer's specific requirements and approval criteria.
• Financial counseling scripts: Producing personalized financial counseling materials based on the patient's specific insurance, service costs, and financial assistance eligibility.

Point of Service

• Ambient clinical documentation: Real-time note generation from physician-patient conversations.
• Charge capture verification: Generating alerts and documentation when captured charges don't align with documented services.
• Real-time coding suggestions: Producing code recommendations with clinical rationale during or immediately after the encounter.

Post-Service

• Claim narrative generation: Writing claim attachments and narratives that support medical necessity for complex procedures.
• Appeal letter drafting: Generating payer-specific appeal arguments for denied claims.
• Patient statement generation: Creating personalized, plain-language billing statements.
• Collection communications: Producing payment reminder sequences calibrated to the patient's communication preferences and payment history.

Generative AI vs. Rules-Based Automation: Why the Difference Matters

Healthcare organizations have used rules-based automation for decades — claim scrubbers that check for missing fields, eligibility verification systems that query payer databases, payment posting algorithms that match remittance data to claims. These systems follow predetermined logic: if condition X exists, take action Y.

Generative AI operates differently in three fundamental ways.

Rules handle known scenarios. Generative AI handles novel ones. A rules engine can catch a missing modifier on a claim because someone programmed that rule. It cannot draft a persuasive appeal for an unusual denial reason it has never encountered. Generative AI can — because it understands the underlying principles of clinical documentation, payer adjudication, and regulatory requirements, not just specific rules.

Rules produce fixed outputs. Generative AI produces contextual ones. A rules-based appeal template fills in blanks: patient name, date of service, denial reason. The argument structure is identical for every appeal. Generative AI crafts the argument based on the specific clinical circumstances, the specific payer's known adjudication patterns, and the specific regulatory framework that applies — producing a unique, optimized appeal every time.

Rules require human maintenance. Generative AI adapts. When a payer changes its clinical editing rules, a rules-based system requires a human to identify the change, write new rules, test them, and deploy them. This lag — typically 4-8 weeks — means thousands of claims are processed under outdated logic. Generative AI models can adapt to new patterns as they encounter them, though they still require human oversight to validate adaptations.

The practical implication: rules-based automation handles the predictable 70-80% of revenue cycle transactions. Generative AI addresses the complex, variable 20-30% that currently requires skilled human judgment — and that's where the majority of revenue leakage occurs.

Generative AI vs. Agentic AI in Healthcare

The healthcare AI conversation in 2026 increasingly distinguishes between generative AI and agentic AI. Understanding this distinction is important because the two capabilities are converging in revenue cycle applications.

Generative AI produces outputs. It drafts an appeal letter. It generates a clinical note. It creates a patient communication. But it doesn't take action — a human must review the output, decide whether to use it, and execute the next step.

Agentic AI takes action. It identifies a denied claim, determines that an appeal is warranted, generates the appeal letter, submits it to the payer through the appropriate channel, monitors for a response, and escalates to a human only if the appeal is denied a second time. Agentic AI combines generative capability (creating the appeal) with autonomous decision-making (deciding to appeal) and execution (submitting the appeal).

The revenue cycle is one of the first healthcare domains where agentic AI is being deployed at scale. The reason is structural: revenue cycle tasks are high-volume, rule-governed, and have clear success metrics (claim paid or not paid, appeal overturned or not). These characteristics make autonomous action safer and more measurable than in clinical domains.

QuickIntell's AI-native platform exemplifies this convergence. Rather than simply generating suggestions for human review, the platform's AI agents autonomously execute revenue cycle workflows — verifying eligibility, scrubbing claims, posting payments, generating and submitting appeals — while routing exceptions and edge cases to human staff. The generative capability (creating documents, communications, and analyses) is embedded within an agentic framework that acts on those outputs.

The trajectory is clear: standalone generative AI tools that produce outputs for human review are transitioning to agentic systems that produce outputs and execute workflows. Organizations evaluating generative AI for revenue cycle should assess not just what the AI can generate, but what it can do with what it generates.

Risk and Governance: What Can Go Wrong

Generative AI in healthcare carries risks that don't exist in other industries. The consequences of errors are financial (incorrect billing), legal (regulatory violations), and potentially clinical (incorrect documentation influencing care decisions). Responsible implementation requires understanding five specific risk categories.

Hallucination

Generative AI models sometimes produce outputs that are fluent, confident, and wrong. In healthcare, hallucination can manifest as:

• Fabricated clinical details: An appeal letter that references a diagnostic test the patient never received.
• Incorrect code rationale: A coding suggestion justified by clinical criteria that don't apply to the documented diagnosis.
• Invented regulatory citations: An appeal that cites a CMS ruling that doesn't exist.

Hallucination rates in general-purpose LLMs range from 3-15% depending on the task and domain. Healthcare-specific models fine-tuned on clinical and regulatory data reduce hallucination rates to 1-4%, but they don't eliminate them entirely. Every generative AI output in healthcare requires human verification — the question is how to make that verification efficient enough to preserve the productivity gains.

Mitigation strategies: Retrieval-augmented generation (RAG) that grounds outputs in verified source documents. Confidence scoring that flags low-certainty outputs for enhanced review. Domain-specific fine-tuning that reduces out-of-domain hallucination. Structured output formats that constrain the model's generation space.

HIPAA and Data Privacy

Generative AI models process protected health information (PHI) to generate useful outputs. This creates three HIPAA compliance requirements that don't apply to traditional software:

1. Business Associate Agreements (BAAs) must cover the AI model provider, not just the application vendor. If the application uses a third-party LLM (e.g., OpenAI, Anthropic, Google), the BAA must extend to that provider.
2. Data residency and retention: PHI used for model inference must be handled according to HIPAA's minimum necessary standard. Organizations must verify that PHI isn't retained in model training data, cached in inference logs, or transmitted to jurisdictions without adequate privacy protections.
3. De-identification for training: If an organization fine-tunes a generative model on its own clinical data, that data must be de-identified per HIPAA's Safe Harbor or Expert Determination methods — or the model training must occur within a BAA-covered environment.

Organizations should require vendors to provide detailed data flow diagrams showing exactly where PHI travels during inference, whether it's ever used for model training, and what retention policies apply.

Bias and Equity

Generative AI models trained on historical healthcare data inherit the biases present in that data. If historical denial appeal success rates are lower for certain demographic groups — not because the clinical merits differ, but because of systemic disparities in documentation quality or payer behavior — the model may generate less effective appeals for those groups.

Revenue cycle bias manifests differently than clinical bias, but it's equally consequential. If an AI coding system consistently assigns lower-acuity codes for patients from underserved communities because the training data reflects historical undercoding for those populations, the result is systematic revenue loss that disproportionately affects safety-net providers.

Regulatory Uncertainty

The regulatory framework for generative AI in healthcare is evolving rapidly. The FDA has cleared over 950 AI/ML-enabled medical devices but has not yet established a clear framework for generative AI used in administrative and financial contexts. CMS has issued guidance on AI-assisted coding but has not addressed generative AI specifically. State-level AI regulations vary significantly.

Organizations should build governance frameworks that exceed current regulatory requirements, because regulations will tighten. The cost of retrofitting compliance into a deployed system far exceeds the cost of building it in from the start.

Workflow Disruption

The most overlooked risk is organizational. Generative AI changes what revenue cycle staff do — and that change, poorly managed, creates resistance, errors, and turnover. A denial analyst who spent 80% of their time writing appeals now spends 80% of their time reviewing and editing AI-generated appeals. The skills required are different: critical evaluation rather than composition, exception identification rather than routine production.

Organizations that deploy generative AI without redesigning workflows, redefining roles, and retraining staff consistently report lower adoption rates, higher error rates, and lower ROI than organizations that treat implementation as an operational transformation rather than a technology deployment.

Current Market Landscape

The healthcare generative AI market in 2026 segments into four categories.

Clinical documentation platforms (Nuance DAX, Abridge, DeepScribe, Suki) focus on ambient documentation and clinical note generation. These are the most mature products with the widest deployment.

Revenue cycle AI platforms (QuickIntell, Thoughtful AI, AKASA) apply generative AI across the billing lifecycle — coding, claims, appeals, patient communications. QuickIntell's approach is notable for integrating generative capabilities within an agentic framework, enabling autonomous workflow execution rather than just document generation.

General-purpose AI infrastructure (Microsoft Azure OpenAI, Google Cloud Healthcare AI, AWS HealthLake) provides the underlying models and infrastructure that healthcare applications build on. These are tools for building solutions, not solutions themselves.

Point solutions address specific generative tasks — appeal writing tools, patient communication platforms, coding suggestion engines. These deliver value in narrow use cases but create integration complexity when deployed alongside other revenue cycle systems.

Real-World Implementation: What the Evidence Shows

Generative AI in healthcare is past the proof-of-concept stage but still early in scaled deployment. The evidence base, while growing, comes primarily from early adopters.

Documentation: A 2024 study across 12 health systems using ambient documentation AI found a 72% reduction in physician documentation time, a 14% increase in charges captured per encounter, and a 22% reduction in documentation-related coding queries. Physician satisfaction scores increased by 31 percentage points.

Denial appeals: Organizations using generative AI for appeal drafting report 45-65% reductions in appeal preparation time and 12-18 percentage point improvements in overturn rates compared to template-based approaches.

Patient communications: Health systems deploying generative patient billing communications report 28% increases in patient payment rates within 90 days, driven primarily by improved patient understanding of their financial responsibility.

Coding: Early data from generative coding AI deployments shows 15-25% reductions in coding turnaround time and 8-12% improvements in coding accuracy, though the variance is high and depends heavily on specialty mix and documentation quality.

ROI Framework: How to Calculate the Financial Impact

Generative AI ROI in healthcare revenue cycle comes from four sources.

Labor productivity. If generative AI reduces the time required for a task by 70%, and you have 10 FTEs performing that task, you've recovered 7 FTEs of capacity. At a fully loaded cost of $65,000-$85,000 per billing FTE, that's $455,000-$595,000 in annual labor savings — or redeployment to higher-value work.

Revenue recovery. Better appeals recover more denied revenue. Better documentation captures more accurate codes. Better patient communications collect more patient responsibility. Quantify each stream separately.

Error reduction. Every billing error has a cost — rework time, delayed payment, potential compliance exposure. Generative AI reduces certain error categories by 40-60%. Calculate your current error rate, apply the reduction, and multiply by the average cost per error.

Speed. Faster documentation means faster coding. Faster coding means faster claim submission. Faster submission means faster payment. Each day of AR reduction has a calculable cash flow impact based on your organization's daily revenue run rate.

A conservative ROI model for a 200-provider multi-specialty group:

ROI Category	Annual Impact
Documentation time savings (200 providers x 45 min/day x $150/hr physician cost)	$3,375,000
Improved charge capture (14% increase x $25M annual charges)	$3,500,000
Denial appeal recovery (18 pp overturn improvement x 8,000 appeals x $850 avg claim)	$1,224,000
Patient payment improvement (25% increase x $6M patient responsibility)	$1,500,000
Coding staff productivity (40% reduction in review time x 12 coders x $72,000)	$345,600
Total annual impact	$9,944,600

Against a typical implementation cost of $800,000-$1.5 million in the first year (including licensing, integration, training, and workflow redesign), the ROI is 6-12x in year one.

Implementation Roadmap

Organizations considering generative AI for revenue cycle should follow a phased approach.

Phase 1 (Months 1-3): Foundation. Assess current workflows and identify the highest-value generative AI applications based on your specific pain points. If denials are your primary challenge, start with appeal generation. If documentation quality drives coding issues, start with ambient documentation. Evaluate vendors against your specific requirements, integration needs, and governance standards.

Phase 2 (Months 3-6): Pilot. Deploy in a controlled environment — a single department, a single payer, or a single workflow — with rigorous measurement. Track accuracy rates, human review time, throughput, and downstream financial metrics. Establish the human review protocols that will govern production deployment.

Phase 3 (Months 6-12): Scale. Expand to additional workflows, departments, and payers based on pilot results. Redesign roles and workflows to leverage generative AI effectively. Retrain staff on review and exception management rather than manual production.

Phase 4 (Months 12-18): Optimization. Integrate generative capabilities with predictive and agentic AI for end-to-end workflow automation. Move from AI-assisted (human does the work with AI help) to AI-autonomous (AI does the work with human oversight) for appropriate workflows.

What to Look for in a Generative AI Healthcare Solution

Not all generative AI solutions are created equal. Evaluate vendors on these criteria.

Healthcare specificity. General-purpose models hallucinate more in healthcare contexts than domain-specific models. Ask whether the model is fine-tuned on healthcare data, what that data includes, and how hallucination rates compare to general-purpose alternatives.

Integration architecture. Generative AI that operates in isolation from your EHR, practice management system, and clearinghouse creates manual handoff points that erode productivity gains. Look for deep, bidirectional integrations.

Human-in-the-loop design. The best systems make human review efficient, not optional. Look for confidence scoring, source attribution, and structured review workflows that let your staff verify outputs in seconds rather than minutes.

Compliance infrastructure. BAAs with all model providers. Data residency guarantees. Audit trails for every generated output. De-identification protocols for any data used in training. These aren't nice-to-haves — they're requirements.

Measurable outcomes. Any vendor claiming generative AI capabilities should be able to demonstrate measurable improvements in accuracy, throughput, and financial outcomes — not just impressive demos.

QuickIntell's platform addresses these criteria by combining healthcare-specific generative models with agentic workflow automation, deep EHR and PM integrations, comprehensive compliance infrastructure, and transparent outcome measurement. The result is a system that doesn't just generate documents — it autonomously executes revenue cycle workflows with human oversight where it matters.

Frequently Asked Questions

What is generative AI in healthcare?

Generative AI in healthcare refers to artificial intelligence models — primarily large language models — that create new content such as clinical documentation, coding suggestions, denial appeal letters, patient communications, and billing narratives. Unlike predictive AI (which forecasts outcomes) or analytical AI (which identifies patterns), generative AI produces the documents, communications, and analyses that healthcare operations depend on. In the revenue cycle, generative AI automates the creation of claims, appeals, patient statements, and other documentation-intensive outputs.

How is generative AI different from traditional healthcare automation?

Traditional healthcare automation follows rules-based logic: if a specific condition is met, the system takes a predetermined action. Generative AI understands context and produces novel outputs tailored to specific situations. A rules-based system might apply a template to an appeal letter. Generative AI crafts a unique appeal argument based on the specific clinical circumstances, payer adjudication patterns, and applicable regulations. The practical difference is that rules-based systems handle standardized, predictable workflows, while generative AI handles variable, judgment-intensive tasks that previously required skilled human staff.

What are the biggest risks of using generative AI in healthcare?

The five primary risks are hallucination (generating plausible but incorrect content), HIPAA compliance (ensuring PHI is handled appropriately throughout the AI pipeline), bias (inheriting disparities from training data), regulatory uncertainty (evolving and inconsistent AI regulations), and workflow disruption (organizational resistance to changed roles and processes). Hallucination is the most technically challenging risk — healthcare-specific models reduce but don't eliminate it, making human review essential for all clinical and financial outputs.

How does generative AI improve revenue cycle management?

Generative AI improves RCM by automating the creation of documentation, appeals, patient communications, and coding suggestions — tasks that currently consume 60-70% of revenue cycle staff time. Specific impacts include 70-90% reduction in appeal writing time with higher overturn rates, 15-25% improvement in coding accuracy and speed, 23-35% improvement in patient payment rates through clearer communications, and 35-50% reduction in documentation-related denials through more complete clinical capture. The aggregate ROI for a mid-size organization typically ranges from $5-$15 million annually.

Is generative AI in healthcare HIPAA compliant?

Generative AI can be deployed in a HIPAA-compliant manner, but compliance is not automatic. Organizations must ensure Business Associate Agreements cover all parties in the AI pipeline (including the underlying model provider), PHI is not retained in training data or inference logs without appropriate safeguards, data residency requirements are met, and the minimum necessary standard is applied to all PHI processing. Organizations should require vendors to provide detailed data flow diagrams and compliance certifications before deploying generative AI with patient data.

What is the difference between generative AI and agentic AI in healthcare?

Generative AI creates outputs — it drafts a document, suggests a code, or writes a communication. Agentic AI takes autonomous action — it identifies a problem, determines the appropriate response, generates the necessary output, executes the action, and monitors the result. In revenue cycle operations, the distinction is between an AI that drafts an appeal letter for a human to review and submit, versus an AI that identifies the denial, determines that an appeal is warranted, drafts the letter, submits it to the payer, and only involves a human if the appeal fails. The healthcare industry is moving from generative to agentic AI, with revenue cycle operations among the first domains where agentic AI is being deployed at scale.