The state of
AI in RCM.

U.S. healthcare spending reached $5.3 trillion in 2024. This report examines administrative costs and the use of AI in revenue cycle management.

Green dots represent the report’s administrative-spending scenario, with an illustrative distribution across service categories. The report covers 2024 spending data, 2026 projections, RCM workflows, workforce distribution, and AI adoption. It also examines automation costs, performance, governance, and customer results.

CMS projects U.S. healthcare spending will reach $6.0T in 2026.

CMS projects$6.02T in US healthcare spending in 2026, up14.0% from 2024.

The June 2026 release puts spending at $6,017.4 billion, compared with $5,278.6 billion in 2024. These are nominal dollars, including both price and utilization changes.

From this point forward, the report uses the 2026 spending outlook.

Service spending comes directly from CMS. Administrative and workflow amounts are estimates derived from that outlook.

Healthcare spending spans hospitals, physicians, medicines, and other services.

In 2026, CMS projects$1870.5B for hospital care,$1244.5B for physician and clinical services, and$561.5B for retail prescriptions.

These represent 31.1%, 20.7%, and 9.3% of projected spending. The remaining 38.9% covers other services and products, public health, investment, and insurance-related costs.

Procedures and medications administered in a hospital sit inside the relevant service category. They are not separate, additive buckets.

Healthcare administrative spending reaches $1.1T in 2026.

Every visit also creates scheduling, eligibility, documentation, authorization, claims, and payment work.

Our 2026 scenario is approximately$1140B in administrative spending. It holds the administrative share constant while national spending grows14.0% from 2024 to 2026.

For this scenario, we normalize the roughly $1T estimate discussed by Penn LDI in 2025 to a 2024 reference year. This is a modeling assumption, not a CMS administrative forecast or a claim that Penn dated that estimate to 2024.

Administration is embedded in healthcare spending. Adding it on top would count the same dollars twice.

Providers account for 73% of administrative spending in 2026.

Payers verify coverage, review requests, and adjudicate claims. Providers prepare evidence, deliver care, submit claims, and pursue payment.

Applying McKinsey’s historical payer/provider mix to our 2026 administrative scenario gives approximately$312B for payers and$828B for providers—27% and 73%.

The ratio comes from McKinsey’s 2019 baseline of $260B payer and $690B provider-side administration. Only the ratio carries forward; these displayed dollar amounts are 2026 scenario estimates, not newly measured totals.

Provider administration supports inpatient and outpatient care.

This illustrative split allocates 43% of the provider administrative total to inpatient care and 57% to outpatient care. The two amounts add to the same provider total on the preceding slide; payer spending is excluded.

The allocation uses the 2024 hospital revenue mix reported by AHA, citing Moody’s, as a proxy. It assumes administrative spending follows that revenue mix and extends the hospital ratio to the broader provider total. That extension is unverified: physician practices and other providers have different care-setting mixes, and administrative cost per dollar of revenue can differ by setting.

Inpatient work includes admission review, utilization management, documentation and discharge billing. Outpatient work includes scheduling, eligibility, authorization, coding and collections for visits and procedures.

Administrative work extends from patient access through payment.

Each workflow combines software, outsourced labor, and employee costs.

Each bar now separates in-house labor, outsourced labor, and software / IT. The total length stays the same. Lighter shades represent employees, middle shades represent outsourced labor, and darker shades represent software and IT.

These allocations use GenHealth’s Series A workflow model. It assigns a different mix to each workflow and applies that mix to both payer and provider spending. They are planning assumptions, not independently measured national spending shares.

The model assigns more outsourcing to coding and patient collections, more software to payment posting and eligibility checks, and more in-house labor to clinical documentation and coordination. Select a row to inspect its mix and dollar amounts.

Rounded source shares are normalized to 100% so segments preserve every bar’s total. The gray remainder is outside the detailed workflow model and has no assigned mix.

Claims and collections account for $266B in 2026.

When people talk about RCM, they often mean the work from claim submission onward: adjudication, payment posting, denial follow-up, appeals, patient billing, collections, and payment integrity. Those rows are highlighted first.

This part of the cycle pursues and reconciles payment for care already delivered. In the report’s model, the highlighted rows total$107B for payers and$158B for providers.

But the claim depends on work completed much earlier. Coverage errors, missing authorization, and incomplete documentation can become payment problems downstream.

In reality, RCM runs from patient access through final payment and accounts for $644B in 2026.

RCM starts before a claim exists. Scheduling and registration establish the account; eligibility and authorization establish coverage requirements; documentation and coding support the claim. The highlight now extends to this earlier work as well as payment and recovery. RCM requires the full process from start to finish.

The highlighted rows cover patient access, eligibility, authorization, documentation improvement, coding, insurer communications, claims, payment posting, denials, patient collections, and payment integrity. The totals include software, outsourcing, and in-house labor on each side.

Within this report’s 2026 allocation model, these rows total $170B for payers and $474B for providers, or $644B combined. They are a subset of the $1.1T administrative scenario, not an additional expense.

Dimmed rows cover broader clinical documentation, order entry, medication administration, workflow coordination, records and quality reporting, care management, member services, and credentialing / plan design. Those functions can support RCM, but their full costs are outside this report’s selected RCM scope.

This is an editorial scope applied to the workflow allocation, not a measured national RCM total. Combined rows such as coding / risk adjustment and referrals / order management are counted whole because the source does not separate their component costs. Payer totals describe corresponding financial-administration work.

Labor accounts for most spending across RCM workflows.

The same highlighted RCM workflows total $644B across payers and providers. This spending falls into two groups: labor spend, which includes both in-house employees and outsourced teams, and software spend, which includes software and IT.

• Labor spend: $518B (80% of selected RCM spending).
• Software spend: $126B (20% of selected RCM spending).

Within labor spend, in-house employees account for $385B and outsourced teams account for$133B.

AI companies can target either spending group. Some add agents to existing software, automating tasks within an application. Others build agents to perform work currently assigned to people, including work that crosses documents, phone calls, portals, and multiple systems.

Both approaches are delivered through software. The difference is the work and budget they target: improving a software product versus taking responsibility for a labor workflow. A company can pursue both. These bars show current modeled spending, not AI vendor revenue or the share already automated.

A practical vendor test: who spends time in whose application? If employees spend substantial time operating the vendor’s application, the product is primarily adding AI to software that people still use. If the vendor’s agents complete work in your existing applications while employees spend little time in the vendor’s interface, the product is closer to automating labor.

Ask the vendor to demonstrate a complete workflow and measure employee time in its interface, agent-completed work in your systems, and human review or exception time. Compare those measures with the manual baseline. Time saved and verified completion matter more than time an agent simply spends running. A monitoring dashboard alone does not mean the underlying work still requires a person.

These are the same dollars shown in the preceding workflow bars, grouped by spending type. Each row’s modeled allocation is weighted by its payer and provider spending; broader administrative functions that were dimmed remain excluded.

The split uses the Series A workflow allocation model, scaled to the report’s 2026 scenario. It is not an independently measured national breakdown. Software includes IT; outsourced labor represents external delivery teams, and in-house labor represents employees.

We’re going to focus on labor spend for a minute.

Outsourced labor accounts for $133B and in-house labor for $385B across the selected RCM workflows. Next, we look at where these teams work and how staffing affects cost and quality.

U.S. RCM operations rely on domestic and overseas teams.

The globe starts over the United States, where the healthcare spending modeled in this report originates. The CMS 2026 projection is $6.02T; the report’s selected RCM workflow scenario is $644B. Those totals are national, not city-level estimates.

The all-spend view allocates $644B across the cities using the supplied relative weights. Scroll further to isolate $133B of outsourced labor at the same dollar-to-height scale. Documented provider and vendor locations are available in a separate layer. U.S. examples include Mayo Clinic in Rochester, Cleveland Clinic, UPMC in Pittsburgh, UCLA Health in Los Angeles, NYC Health + Hospitals, and HCA / Parallon in Nashville. Job postings establish team locations; some roles are hybrid or remote.

RCM can be performed by a provider’s own staff, a domestic service partner, or an offshore team. Outsourced does not necessarily mean abroad; a U.S. vendor can employ people in several countries.

A disclosed example: roughly 41% U.S. / 59% international. R1’s 2023 filing reports 12,300 U.S. employees (including 500 part-time) and 17,600 international employees. This is company headcount, not FTEs or a census of U.S. RCM.

A newer, different business mix: Sagility’s FY2024–25 workforce. Its sustainability report lists 38,754 employees across five countries. The country figures imply 6.3% in the U.S. and 93.7% abroad; India and the Philippines together account for 85.0%. Sagility serves both health plans and providers, so this is broader healthcare operations employment, not provider RCM alone.

These figures use the sustainability report’s reporting perimeter. Sagility’s FY25 earnings release separately reports 39,409 year-end employees; we do not combine that denominator with this country breakdown. The contrast with R1 demonstrates why one vendor’s staffing mix cannot stand in for the whole market.

We did not identify a public, comprehensive count covering in-house provider teams and every RCM service firm. The industry-wide domestic/offshore share remains unquantified here. Historical workforce evidence is dated explicitly, alongside the report’s 2026 spending scenario.

Outsourced healthcare administration spans delivery teams around the world.

The globe allocates the same $133B of outsourced RCM labor shown in the preceding spending split. Each city retains its share of the supplied dataset, scaled to this total. The map starts over the United States.

Select a spike to see the city’s allocated amount and percentage. The headline, city values, regional totals, and CSV all use the same spending basis.

These allocations assume that the relative geography in the supplied outsourced healthcare-administration dataset also applies to outsourced RCM. Its original $68.6B total is used only to calculate city weights. It is not an additional spending total. The allocation does not measure local payroll or municipal accounts.

The linked market reports describe broader healthcare BPO markets; their public pages do not substantiate this city allocation. The city figures remain a supplied illustrative dataset with no documented city-level estimation methodology or reference year.

Staffing location changes the cost of an RCM team.

Geography sources include employer career pages and role descriptions, such as Access Healthcare’s Chennai A/R role and Sagility’s Philippines careers board. These document advertised locations and tasks, not workforce totals or measured city spending. Posted salaries are not the basis of the staffing-cost assumptions below.

Switch to the globe’s Spending scenario layer to explore a 1,000-person team. Its illustrative starting case puts 500 people in the U.S., 375 in India, 100 in the Philippines, and 25 in Colombia.

At assumed annual fully loaded costs of $80,000 / $20,000 / $24,000 / $30,000, respectively, spending is$40M / $7.5M / $2.4M / $0.75M. Half the people are abroad, but just 21% of the $50.65M staffing budget is abroad. Change the inputs to see the difference.

These are planning assumptions, not researched wage estimates or a forecast of national RCM spending. City weights are illustrative too; the tallest spikes show the largest allocations in this scenario, not a measured ranking.

Back in the broader$958B core workflow model, the assumed 65% employee / 18% outsourced / 17% technology mix corresponds to$622B /$172B /$163B. That model includes work beyond narrow RCM. Neither labor bucket has a verified geographic split, so we do not multiply these national totals by the example team’s shares.

Use the current team’s costs and performance as the baseline for an AI evaluation.

Coding accuracy changes the total cost of outsourced RCM.

Outsourcing and geography are separate choices: a U.S. team can be in-house or outsourced. We did not find a representative, current comparison of offshore vendors with in-house U.S. teams across RCM. Coding studies provide a narrower view.

In a 2018 report, KIWI-TEK’s COO described a comparison across six hospitals and health systems. Offshore outsourced coders averaged 6.5 percentage points lower accuracy than domestic outsourced coders, with six additional audit hours per coder per month. Participants used both services for at least a year with the same onboarding, auditing, and training procedures. This vendor-authored comparison is historical and does not establish an in-house advantage or a national accuracy rate.

KIWI-TEK: domestic versus offshore coding, March 2018 ↗

Managed Resources reports reviewing 14,674 cases at a U.S. health system using an offshore coding vendor, from Q4 2022 through Q1/Q2 2024. Accuracy improved across E/M, ICD-10, and CPT/HCPCS after audits, issue tracking, and education. This single-client vendor case study supports investing in quality controls; it does not isolate geography as the cause or compare performance with an in-house team.

Managed Resources: offshore coding audit and improvement ↗

Compare your teams on the same specialty, payer mix, and case complexity. Use blinded audits to measure first-pass accuracy, financially material errors, rework time, turnaround, and cost per correctly completed case. Include supervision and escalation time in labor costs. Access to clinical context, payer rules, training, and feedback should be evaluated for both teams.

AI changes the cost and supervision required for RCM work.

GenHealth’s operating assumptions put AI-only work at $2–4 per equivalent labor hour, depending on the workflow, with quality slightly above offshore labor. No quantified accuracy difference has been supplied, so the chart does not assign one.

For AI paired with U.S.-based staff, the proposed cost is $6 per hour, with collections up to 34% higher. These are supplied operating claims, separate from the published coding study on the preceding slide. They have not been independently validated here; a cohort, time period, and collections denominator were not supplied. The 34% figure is an upper-end claim, not an established average or an accuracy score.

The offshore all-in planning price is $10 per hour, including the management cost of supervising the outsourced team. AI paired with U.S. staff is shown at $6 per hour. These are supplied planning estimates, not measured market prices.

Evaluate cost per completed workflow after human review, retries, and escalation. Compare collections on matched accounts over the same collection window, accounting for payer mix, balances, and fees.

How much of your admin work is AI doing today?

Select the closest percentage of administrative work that AI completes today.

Consider completed work across your teams, including outsourced operations. Having an AI feature available is different from using it to complete work.

Then scroll to compare agents embedded in software with agents performing work in your existing applications and systems of record.

Here is how much administrative work other respondents report AI doing.

The histogram shows saved responses to the preceding question, as percentages without counts. Each bar corresponds to one of the six answer choices.

These are voluntary reader responses, not a representative industry survey. One response per poll is accepted for each browser identity. Your saved answer is included in the distribution.

If you use a vendor, are you working in their application, or are their AIs working in yours?

Which best describes where the work happens today? Consider your employees’ time in the vendor’s application and the vendor’s AI completing work in your existing applications and systems of record.

Choose both if the two approaches are about equally common, or indicate that you are not using a vendor. This answer is saved separately from your AI work-share answer.

Here is where other respondents say their vendor work happens.

The chart shows saved answers to the vendor workflow question as percentages, without counts. All four answer choices are shown, including both equally and not using a vendor.

These are voluntary reader responses, not a representative industry survey. Your saved answer is included. Each browser can contribute one response to this poll, separately from the AI work-share poll.

Answer the vendor question to see what your workflow suggests.

Employees primarily using the vendor’s application: the vendor is primarily agentifying software as a service. Your team still operates the workflow, with AI features helping inside the product.

Vendor AI primarily working in your applications: the vendor is closer to agentifying labor. Its agents take on execution inside your existing applications and systems of record.

Both equally: the vendor may combine the two approaches. If you are not using a vendor, use this question during product demonstrations.

The distinction depends on who completes the work. A dashboard for monitoring agents does not make the product SaaS automation by itself. Check verified task completion, employee time saved, and review or exception effort before drawing a conclusion.

Most HFMA respondents are piloting AI rather than deploying it at scale.

HFMA’s February 2026 survey of 95 healthcare finance professionals found 27% reporting AI deployment at scale across multiple functions and 53% conducting pilots in select areas. That makes the implementation gap a central RCM opportunity: buying or testing AI is much more common than operating it broadly.

These are respondent-reported organizational stages, not the percentage of claims handled autonomously or a nationally representative estimate of all providers. The survey does not establish achieved savings or collections uplift.

HFMA: The Revenue Cycle of the Future, April 2026, p. 4 ↗

84% of surveyed health insurers use AI or machine learning.

NAIC’s May 2025 report found that 78 of 93 responding health insurers used AI/ML in some capacity. Sixteen states surveyed companies from November 2024 through January 2025, selecting larger insurers or those with significant state market share.

This is broad enterprise adoption, including statistical models, rather than a measure of generative AI agents or autonomous RCM. The surveyed business covers comprehensive major medical and student health products; it is not a census of every plan.

Do not subtract HFMA’s 27% from NAIC’s 84% to measure a payer lead. The dates, populations, technologies, and deployment thresholds differ. Our inference: an early-mover advantage must come from better execution in a particular workflow; providers should not assume the payer is starting without AI.

NAIC: Health AI/ML Survey Report, methodology and Table 1 ↗

Surveyed insurers already use AI in claims and prior authorization.

NAIC’s Table 5 reports 31 companies with AI/ML already in production for claims adjudication, 18 for prior authorization, and 42 for utilization, severity, or quality management. These are current-production responses, separate from planned deployments.

For the provider, these touch the front-end authorization and back-end payment stages in our RCM timeline. The practical response is to improve evidence completeness, policy alignment, submission quality, and exception resolution, then measure acceptance and collections on matched case mixes.

Production use does not establish autonomous approval or denial, the share of claims touched, or accuracy. The report describes Table 5’s respondent base inconsistently; we retain its published company counts without calculating penetration percentages. These surveys also do not quantify how much of administrative spending estimated here is already automated.

NAIC: Table 5, printed pp. 14–15 ↗

Eligibility checks lead provider transaction spending in CAQH’s analysis.

CAQH’s 2023 Index specialty analysis identifies eligibility, claim submission, and claim-status inquiries as the three largest medical transaction spending categories. Applying its reported provider shares gives approximately $41.2B, $18.4B, and $12.1B, respectively.

This is a narrower, historical transaction-cost lens—not an allocation of the report’s 2026 core workflow scenario. The numbers must not be added together across those two models.

Total cost depends on transaction volume as well as the cost of each transaction.

Specialists face higher costs for several manual billing tasks.

For manual eligibility, authorization, and claim-status work, CAQH reports higher unit costs for specialists and behavioral-health providers than generalists. Use the selector to compare tasks.

Operationally, a provider-focused product needs more than one generic queue. Useful segments include primary care, procedural specialties, behavioral health, and hospital-based services. Specific workflows should be validated with each practice rather than inferred from its specialty name.

• Primary care: repeated coverage checks, routine billing, and patient balances.
• Procedural specialties: authorization evidence, procedure documentation, and claim edits.
• Behavioral health: coverage details, authorization requirements, and recurring-visit billing.
• Hospital services: documentation, coding, and coordination of facility and professional billing.

These are workflow-design examples, not quantified specialty market shares. Named fields such as oncology, cardiology, orthopedics, radiology, and DME need their own volume and cost data before assigning national RCM dollars.

Complex encounters cost more to bill.

A 2018 JAMA study at one academic health system estimated billing and insurance costs from $20.49 per primary-care visit to $215.10 per inpatient surgical encounter. The analysis included professional and hospital billing. The graphic compares all five encounter types.

This supports a complexity gradient, not a current national ranking. High case cost, high claim volume, and a high percentage of revenue consumed by billing are three different measures.

CMS publishes Medicare Part B expenditure and service counts by specialty. Those are payments for care, not RCM expense. They can supply a scoped volume lens, but multiplying them by an unrelated administrative percentage would produce an unsupported estimate.

Our takeaway: target repeatable volume in routine workflows and expensive exceptions in complex care. We have evidence for those two opportunities; we do not yet have a defensible all-payer national RCM spending leaderboard by individual specialty.

AI is automating routine steps across the revenue cycle today.

Automation today includes checking eligibility, validating claims, retrieving claim status, and posting payments. Much of this uses structured transactions and rules. It does not require a conversational agent for every step.

Waystar advertises automated claim-status checks, payment posting, and reconciliation. Its claim-management offering also connects coverage checks with claim validation. These are product capabilities, not evidence that the entire revenue cycle runs unattended.

Waystar also offers denial prioritization, generated appeal letters, payer forms, submission, and delivery tracking. These automate parts of recovery without guaranteeing payment.

Select an example in the graphic to see the actions, the technology involved, and the point where a person takes over. Each maps to the front-to-back RCM timeline.

AI can turn clinical documentation into inputs for billing.

Language models can extract relevant evidence, summarize documentation, and identify possible coding gaps. That extends automation beyond copying fields and checking transaction formats.

AKASA describes re-analysis of clinical records to surface supported coding opportunities, followed by human expert review before billing. Its authorization-assistance offering is another example of helping staff prepare the work rather than independently deciding whether care should be approved.

The practical distinction is between generating a suggestion, executing a permitted action, and closing an account correctly. A generated summary is useful, but it is not the same outcome as an accepted claim or reconciled payment.

Outsourced work is easier to automate; in-house labor offers larger savings per hour.

Standardized outsourced queues are likely to be automated first. Work that has already been packaged for an external team often has documented steps, repeatable inputs, measurable outputs, and clear handoffs. Eligibility checks, status follow-up, and routine posting are examples of bounded actions that can be easier to automate.

This is our sequencing thesis, not a measured industry-wide order. Outsourcing does not itself make a task easy: vendors also handle complex coding and appeals, while many in-house teams run highly standardized processes.

Higher-cost in-house labor creates a larger incentive to reduce labor spend per hour saved. When a domestic employee’s fully loaded cost exceeds that of an offshore delivery role, the same successful automation has more potential dollar value in-house. Actual savings depend on the role, service contract, review burden, and whether capacity can be redeployed or expense reduced.

In the illustrative workforce scenario, an $80,000 annual role costs four times a $20,000 role. Saving the same fraction of time therefore has four times the gross labor value—before implementation, oversight, and exceptions. Those costs are editable assumptions, not national wage estimates.

Prioritize repeatable tasks. Evaluate more complex workflows against their labor costs.

The existing model also allocates 65% to employee labor versus 18% to outsourced labor. Those are broad administrative scenario shares, not measured RCM staffing shares or immediately realizable savings.

AI can keep routine RCM queues moving outside office hours.

AI does not need sick time, vacation coverage, or a shift change. For bounded work, it can reduce scheduling constraints and keep a backlog moving outside office hours. Parallel execution can add capacity without recruiting and training a new person for every increment of volume. But compute, rate limits, maintenance, integration support, and human oversight still cost money. Around-the-clock availability is a system design objective, not an inherent uptime guarantee.

Recheck pending claims overnight; escalate unresolved accounts when the team returns.

Compare the full cost per completed task →

AI can review long patient records without fatigue.

An AI workflow can be designed to process an entire 200-page authorization or appeal packet, rather than relying on a quick skim. That makes comprehensive review more economically feasible. It should account for all pages, check extraction quality, and connect each conclusion to its source. A large context window alone does not guarantee exhaustive comprehension: missing pages, poor scans, retrieval omissions, and degraded recall can all change the answer.

Track every page, extract relevant evidence with citations, and flag unreadable scans or conflicting dates.

Compare the full cost per completed task →

Automated workflows need validation and ongoing monitoring.

Organizations want consistent checklist execution and visibility into what happened on each account. Software can make versioned rules and audit trails easier to apply across a large queue. These controls require explicit implementation. Scaling requires validation, outcome monitoring, and a way to stop or roll back a faulty workflow. Track completion quality alongside throughput.

Verify the patient, policy version, and submission outcome; prevent duplicate claims before retrying.

Compare the full cost per completed task →

Missing context? Pause and ask.

Experienced staff notice when the literal instruction is no longer the right action. A payer representative may mention an exception on the phone; a clinician may change the treatment plan; a patient may reveal a hardship absent from the record. An AI can reason about unfamiliar situations, but it cannot reliably reconstruct facts it never received. Out-of-band events need a path into shared context, and material ambiguity needs a human handoff rather than a confident guess.

A patient says a financial counselor paused collections. Stop outreach and verify the arrangement instead of following the next reminder step.

Explore repeatability and exception handling →

RCM AI needs current payer rules and provider-specific context.

General model training may not include a provider’s contracts, local workflows, recent payer changes, or uncommon encounter types. Retrieving current sources helps supply facts; it does not substitute for representative evaluation of the actual task. Useful labeled data must show correct actions and verified outcomes, including difficult cases—not only whether a claim happened to be paid. Teams need authorized data, expert review, and ongoing checks as policies and systems change. Evaluate unfamiliar cases separately to identify gaps in task performance.

A new denial code or conflicting policy date should trigger review—not an invented explanation or an unsupported appeal.

Explore repeatability and exception handling →

Document AI can extract evidence from medical records.

Healthcare work arrives as PDFs, scanned orders, referral packets, and faxes. Document AI turns those inputs into structured information.

For a bounded task, this is usually the easiest starting point: the input is fixed and extracted fields can be checked against the original page.

Validation still matters. Missing pages, handwriting, ambiguous identifiers, and contradictory evidence need explicit handling. Extraction accuracy is not the same as a correctly completed authorization.

Voice agents can handle routine RCM calls—best kept under 5 minutes today.

Start with short, routine calls: status checks, scheduling, reminders, and information collection. Under five minutes is a practical starting guideline, not a universal technical limit.

The conversation can follow a defined objective, confirm what was heard, and escalate when the other party cannot provide the answer. Identity checks, consent requirements, background noise, IVR menus, and write-back still shape the implementation.

Our practical ordering is PDF → voice → web → computer. This is a deployment heuristic, not a universal benchmark.

Web agents can complete RCM tasks across payer and provider portals.

Portals and EHRs require more than understanding language. The agent must authenticate, find the correct record, navigate permissions, enter data, and verify that the system accepted the change.

Sessions expire. Forms change. Different payers expose different workflows. The difficulty comes from reliably executing the whole sequence.

Structured interfaces and authorized network requests can reduce reliance on repeated visual interpretation. Where those interfaces are unavailable, browser and human fallbacks still matter.

Computer-use automation still faces cost and reliability limits.

A visual computer-use agent reads screenshots and moves a cursor through applications. It can reach legacy systems that lack a convenient integration.

But screen interpretation, coordinate selection, changing layouts, and long action sequences introduce failure points. Repeated screenshots, reasoning, and retries also consume tokens and time.

Computer-use models are capable, but a successful demonstration does not establish repeatability or favorable economics on a production workflow. Both need to be measured.

Check payer requirements. Send unclear cases to a person.

AI gathers facts from the patient record. Validated rules check them against current payer requirements. Missing or conflicting evidence goes to human review.

GenHealth uses network interactions to automate browser workflows.

GenHealth’s approach is to translate a workflow into repeatable execution through a purpose-built browser and the authorized network interactions behind the application.

Instead of asking a vision model to rediscover each button on every run, the workflow uses structured actions, known state, and explicit checks.

This architecture is designed to reduce repeated model calls and improve consistency. Its advantage must still be demonstrated on the customer’s systems, including authentication, changing endpoints, and error recovery.

RCM automation needs repeatable results and explicit error handling.

Test repeated execution using the same approved inputs again under the same policy, workflow version, and system state.

Look for validated preconditions, safe retries, duplicate prevention, outcome checks, version control, and an audit record. When the environment changes, the workflow should stop or escalate deliberately.

Define completion criteria and the conditions that require a workflow to stop or escalate.

Deterministic execution is a design property. Accuracy, coverage, and availability remain separate measurements.

Errors compound across the steps of an automated workflow.

At 98% success per step, a 20-step workflow has only about 67% end-to-end success under a simple independent-error model. Adjust the sliders to explore the difference.

Published benchmark scores also need context. Anthropic’s February 2026 system card reports OSWorld-Verified scores of 61.4% for Sonnet 4.5 and 72.5% for Sonnet 4.6. These are general computer-task results under specified evaluation conditions, not RCM accuracy rates or a current leaderboard.

Document extraction, coding accuracy, task completion, and claim acceptance have different denominators. This report does not assign unsupported, comparable accuracy percentages to vendors or modalities.

Compare the full cost per completed task.

For tasks people complete quickly, repeated model calls, visual processing, retries, and human rework can make AI more expensive than manual execution.

That is a scenario to test, not an inevitable property of computer use. Model choice, caching, workflow length, supervision, and labor rates change the result.

Use the calculator to compare cost per completed task. The starting values are illustrative assumptions, not current model prices or measured GenHealth results.

Include review, rework, and exception handling in cost-per-task calculations.

Chatbots answer questions when people ask.

The first familiar interface was a conversation: summarize a policy, explain a denial, draft a letter. A person still carried the answer into the operational system.

This remains useful work. But generating an answer and owning an outcome are different capabilities.

AI agents can complete assigned tasks.

An agent adds tools and execution: retrieve a record, check eligibility, fill a form, or submit an approved request. It can carry out an ad hoc task across several steps.

The important boundary is the handoff. What can the agent actually complete, and what does it return to a person?

Adoption is uneven: HFMA’s February 2026 survey of 95 finance professionals found 27% deploying AI at scale across multiple functions and 53% piloting selected areas.

AI coworkers can manage defined work queues.

The next step is an AI coworker: persistent context, assigned responsibilities, a work queue, and clear escalation rules.

It does not wait for a new prompt for each account. It monitors pending work, takes the next authorized action, and hands exceptions to the right person.

“Coworker” describes an operating model, not independent legal or clinical accountability. People set the permissions, policies, and quality standards.

AI operators would manage entire processes under human supervision.

Our longer-term thesis is an operating layer that coordinates whole administrative processes. One person—or a small team—could supervise work previously distributed across many queues and applications.

The shift is from executing tasks on demand to automating their initiation, completion, verification, and follow-up.

The operator model would fail to scale if exception volume, supervision, or maintenance costs rise as fast as the work it takes on. Persistent dependence on human judgment could favor bounded assistants instead of a small team supervising an entire operation.

This is a forecast, not today’s industry baseline. It depends on dependable integrations, repeatable execution, monitoring, and humans who can resolve the cases the system cannot.

Cross-system agents can connect work across multiple applications.

Application creators can add native agents with deep access to their own product. That can make a single system much more effective.

Third parties can build an orchestration layer across the EHR, payer portal, billing platform, documents, and phone. This reaches the handoffs where healthcare work often stalls.

GenHealth’s thesis: the cross-system approach is more likely to deliver cross-system automation because RCM workflows involve several vendors’ systems. Native agents can still be valuable components; the outcome is not settled.

Native agents could win in a single-vendor environment with deep EHR access, supported write operations, and fewer external handoffs. Cross-system automation can lose its advantage when integration maintenance, fragmented permissions, or exception work outweigh the benefits of broader coverage.

We would revise our thesis if matched real-world evaluations showed native agents delivering better verified completion, lower all-in cost, and fewer safety incidents across the same multi-system case mix. A hybrid approach may outperform either alone.

CMS’s prior authorization rule also pushes interoperability forward, with API requirements generally beginning in 2027 for affected payers.

AI can automate administrative labor as well as software tasks.

Adding a copilot to SaaS improves the application. Automating labor means completing the work people perform across applications and organizations.

The reference model illustrates the distinction: approximately 17% SaaS/IT, 18% outsourced labor, and 65% employee labor. These are directional estimates, not audited national accounts.

Toggle the graphic to compare the two scopes. GenHealth’s ambition is to address the cross-system administrative workload, including the work done by internal teams and service providers.

Spend is not equivalent to automatable savings, and automating a task does not mean eliminating an entire job.

Optum and R1 lead the estimated outsourced RCM revenue ranking.

The five largest estimates among the outsourcing providers reviewed are Optum RCM ($4.2B), R1 ($2.6B), Ensemble ($1.7B), Conifer ($1.3B), and Smarter Technologies / Access ($0.8B). These are annual service-revenue estimates assembled from public information available in September 2026, not uniformly reported 2026 results or market shares.

Optum’s disclosed $140B of annual billings managed is multiplied by an assumed 3% fee yield. Ensemble’s June 2026 $55B managed net patient revenue uses the same 3% proxy. Conifer’s $25B managed volume uses a 5% proxy. These yields are assumptions, not disclosed contract rates; billings and net patient revenue also differ. The positions can change with the fee assumptions. None of these managed volumes is itself vendor revenue.

R1’s $656.8M Q3 2024 revenue is annualized and rounded. Smarter’s estimate uses the $800M floor in its May 2025 combination announcement, including Access, SmarterDx, and Thoughtful.ai. The estimates use different source dates and have not been grown to 2026.

This ranking covers the reviewed outsourcing operators and their RCM platforms. It excludes captive HCA shared services and pure software vendors. Public disclosures do not establish a definitive industry-wide top five: Guidehouse, Omega, and GeBBS remain difficult to size consistently and could change the ranking.

Optum managed billings ↗

R1 Q3 2024 results ↗

Ensemble June 2026 managed revenue ↗

Ensemble historical fee disclosures ↗

Conifer managed revenue ↗

Smarter Technologies combination announcement ↗

Build AI capability now to capture more value later.

In retail banking, digital leaders delivered 8.1% annual shareholder returns versus 4.9% for laggards from 2018–2022. McKinsey compared 20 leaders with 20 laggards.

This measures digital maturity, not adoption timing or AI alone. It supports investing in capability; it does not prove a guaranteed early-mover return in healthcare.

Move now, before AI raises the competitive baseline.

Our outlook is that AI will become embedded across the revenue cycle. Starting now gives teams time to improve collections, lower costs, and build operating experience.

Plans are adopting AI too. More automated review could make collections harder for providers that cannot respond with complete evidence and timely follow-up.

Providers with lower administrative costs may have more room to offer lower rates. As contracts and prices adjust, competitors and customers may capture more of the savings.

This is a competitive scenario, not a measured healthcare margin forecast. Lower administrative costs do not automatically produce lower negotiated rates.

GenHealth customers report gains in collections, workload, and billing speed.

Higher revenue collections — Piedmont. GenHealth’s case study reports a +34.2% paid-to-date collections regression effect relative to the January 2025 baseline, based on a short observation window. The analysis adjusts for trend and seasonality; it is an association, not proof that AI alone caused the increase. Only three treated paid months were available.

Prior-authorization adjudication support — GuideHealth. The deployment reports up to 60% lower utilization-management workload and approximately $1.2M in projected annual cost savings. The savings are projected, and the workflow supports clinical review and decision recommendations.

Less manual administration — MedExpress. Automated intake, eligibility checks, and order creation shift staff toward exceptions. The case study reports 62% lower billed days and more than 35,000 orders built. These describe operational activity; the study does not quantify a realized administrative-cost reduction in dollars.

These are GenHealth-reported customer outcomes, not typical-return guarantees or direct evidence comparing early and late adopters. Each case study provides the scope and methodology behind its figures.

RCM teams should validate AI before delegating work.

Adoption has several meanings: a purchased license, a pilot, a tool staff use, or a workflow that reliably completes work. Report the fraction of eligible accounts actually processed and the human touches still required, not only the number of AI projects.

The 2026 Guidehouse/HFMA survey page reports 69% outsourcing some or all of RCM and 88% identifying payer issues among their top three concerns. These are respondent findings, not AI penetration or market-share estimates.

The path in the graphic is our proposed rollout sequence. Advance based on evidence, with explicit scope and a named operations owner. A mature eligibility workflow can coexist with a coding pilot inside the same organization.

RCM automation creates value through savings, collections, and faster cash.

Four benefits need separate accounting: staff capacity released, expense actually avoided, additional revenue collected, and cash received sooner. They have different implications for a CFO.

Subtract integration, licensing, model usage, supervision, rework, and ongoing maintenance. Labor time saved may reduce overtime or hiring needs; it does not automatically reduce current payroll. Evaluate recurring benefits separately from one-time implementation costs.

The cash calculator illustrates how five fewer A/R days on $1B annual net patient revenue could release approximately $13.7M. That is accelerated cash, not new revenue or recurring profit. The estimate assumes stable daily revenue and collectible receivables.

Compare cost per completed task →

RCM automation depends on usable data, clear ownership, and exception handling.

Successful rollout starts with a narrow queue, a measured baseline, reliable system access, and staff who can handle exceptions. Validate on ordinary cases and difficult ones, including payer variation, missing documents, conflicting records, and failed writes.

A common failure mode to test for is a strong demo followed by expensive review: the model looks accurate, but staff still check every field, repair integrations, and chase unresolved accounts. Another is optimizing throughput while increasing denials or patient complaints.

Track first-attempt completion, verified outcome, review minutes, backlog age, final denials, and full cost per account. Compare matched case mixes over sufficient time for claims to settle. Agree on stop conditions before the pilot, and retain a staffed fallback.

Expand deployment after verifying improvements in operational and financial results.

Automated RCM decisions need source evidence and verified outcomes.

Trust comes from inspectable evidence and correction mechanisms. Staff should see which record and policy supported an action, what changed in the destination system, and why an exception was escalated.

For an appeal, keep links to the supporting chart passages and approved argument; verify delivery separately from eventual payment. For coding, preserve the expert review where required by the workflow. Measure error severity, not only average accuracy.

Test performance across payer and specialty groups and, where appropriate and lawful, relevant patient populations. Track complaints and overrides. Give reviewers time and authority to disagree, so human oversight does not become a rubber stamp.

Federal policy aims to reduce healthcare administrative work.

Administrative burden is a recognized federal policy problem. In 2019, CMS estimated that its Omnibus Burden Reduction rule would remove 4.4 million hours of provider paperwork annually by eliminating unnecessary, obsolete, or excessively burdensome requirements. This was a forecast for a specific rule, not measured savings from AI.

CMS also publishes concrete changes to simplify Medicare claims documentation. These efforts acknowledge that government requirements themselves can create work that can be reduced while preserving patient protection and payment integrity.

Healthcare regulations also create documentation and reporting work.

Documentation supports accountable payment and oversight, but assembling evidence, maintaining records, and communicating decisions also consume staff time. Our reading: regulation is both a source of administrative requirements and a tool for simplifying them. It is not the only source; payer policies and fragmented systems also create work.

The 2024 prior authorization rule illustrates the tradeoff. It requires specific denial reasons and public reporting of authorization metrics, alongside standardized electronic exchange. Existing notice requirements remain. Greater transparency therefore requires operational work even as standardization aims to reduce repeated calls and submissions.

This work sits inside the documentation, authorization, and payment stages of our RCM model. We do not assign a national dollar share of administrative spending to regulation: the sources here do not establish that allocation or the net cost of all U.S. regulation.

New prior-authorization standards require operational changes.

Under CMS-0057-F, operational requirements generally start in 2026 and API requirements in 2027; exact dates depend on payer type. Covered payers include Medicare Advantage, Medicaid and CHIP programs and managed care, and federally facilitated exchange plans. This is not a mandate covering every commercial payer.

For non-drug prior authorization, impacted payers other than those exchange plans face 72-hour expedited and seven-calendar-day standard decision limits. The rule’s prior authorization provisions exclude drugs. Teams should confirm their applicable program, deadline, and current rule before changing workflows.

Our recommendation: capture source evidence once, reuse it for authorized submissions, retain policy versions and acknowledgments, and route missing evidence to an accountable person. Measure total staff time, rework, and reporting effort—not just faster data entry. AI can help perform required documentation; it does not waive the underlying obligation or justify inventing evidence to complete a record.

Automate the work. Measure the outcome.

Start with one repeatable workflow. Let AI work in your existing systems, with people handling exceptions. Build capability now, before AI becomes the competitive baseline.

Measure completed work, collections, and the full cost—including review and management—before expanding.