Agentic AI for Revenue Cycle Management: AI Agents, Vendors and Rollout
Founder & CEO
15+ years building healthcare technology. Founder & CEO of Nirmitee.io, whose teams deliver EHR integrations, FHIR and HL7 platforms, ABDM and NHCX connectivity, and clinical AI for health systems and healthtech companies.

Updated October 2026
Agentic AI revenue cycle management uses AI agents that plan and carry out several billing steps on their own, then hand exceptions to people. An agent can read a denial, gather the evidence, draft the appeal and file it, where an older tool would only flag the claim for staff.
For how a US claim moves end to end, see our RCM integration guide.
The strongest public case is still an estimate of potential, and most vendor results are self-reported. McKinsey's January 2026 article estimates AI could cut cost to collect by 30 to 60 percent and recommends starting with back-end work such as denials, underpayments, follow-up and cash posting. This guide covers what AI agents for healthcare revenue cycle work can do, who sells them, and how to test one. If you plan to build rather than buy, our revenue cycle management software development team builds the integrations these agents depend on.
Correction, October 2026. An earlier version of this page credited McKinsey with "4 Levels of RCM Automation". We could not find that framework in the text of McKinsey's 2026 or 2023 revenue cycle articles, so the levels below are now labeled as our own. We also removed production results, budgets, a payback period and three state bill numbers that we could not trace to a source.
What is agentic AI in healthcare revenue cycle?
Agentic AI in the healthcare revenue cycle is software that takes a billing goal, decides the next steps and acts in connected systems within set limits. McKinsey describes it as able to "autonomously make decisions and execute complex end-to-end processes", while generative AI mainly gives advisory support.
The practical test is who does the work. A denial prediction model produces a list for staff. A denial agent works the list and asks for help when a case falls outside its limits.
| Tool | What it does | Where it breaks | Who decides |
|---|---|---|---|
| RPA bot | Follows a fixed script through screens | Portal layout changes, cases outside the script | The script's author |
| Predictive model | Scores risk, such as denial likelihood | Produces a score and takes no action | Staff |
| Generative AI assistant | Drafts a letter or summary on request | Waits to be asked, cannot finish the task | Staff |
| AI agent | Plans steps, uses tools, checks the result | Wrong actions at speed if limits and logs are weak | The agent inside limits, people outside them |
Why health systems are looking at autonomous AI agents for RCM
Revenue cycle work is expensive, repetitive and short of staff time. McKinsey puts the cost at 3 to 4 percent of an at-scale health system's revenue, or more than $140 billion a year, and says nearly 20 percent of claims are denied, with as many as 60 percent never appealed.
Read those figures with their limits in mind. The 20 percent comes from a KFF study of ACA Marketplace plans, so your payer mix may differ. Experian Health's State of Claims 2025 found that about 41 percent of surveyed providers have denial rates of 10 percent or higher. That is a share of providers, and it does not mean 41 percent of claims are denied, as this page wrongly said before. On prior authorization, an AMA survey from late 2024 found practices complete 39 requests per physician per week, with physicians and staff spending 13 hours a week on them.
Four levels of autonomy, and what McKinsey actually said
A simple way to place any product is to ask how much it may do before a person steps in. These four levels are our working model, and McKinsey did not publish them. We read the text of its two articles but not the exhibit images.
- Level 1: Advisory AI. The system suggests a code or flags a risky claim, and a person decides.
- Level 2: Assisted automation. Routine cases run on their own and exceptions go to staff.
- Level 3: Autonomous agents (agentic AI). The agent handles a multi-step job and escalates when it is unsure or out of bounds.
- Level 4: Connected agents. Agents pass work to each other across the cycle, with people still overseeing them.
McKinsey's own view is that the touchless model "will be unlocked in parts", moves toward an interconnected network of agents, and that "a human will always remain in the loop". It also says few health systems have deployed agents on their own and that most deployments so far came through vendors. The $60 million to $120 million saving it gives is an "if" scenario for a system with $6 billion in patient revenue. For the full measurement method, read our guide to the touchless revenue cycle.
Six use cases for AI agents in revenue cycle management
Back-end tasks with written rules are the safest first use, and patient-facing or clinical tasks need closer human review. The order and the review points below are our judgment, except for the back-end starting point, which follows McKinsey. No row is a measured result.
| Agent | Starts when | What it does | Main data | A person steps in when |
|---|---|---|---|---|
| Denial management | A denial arrives | Maps the reason code, gathers evidence, drafts and files the routine appeal | X12 835, CARC and RARC codes, chart | The appeal needs clinical judgment |
| A/R follow-up | A claim ages past your limit | Checks status and works the next action, including payer portals and phone trees (IVRs) | X12 276/277, portals | The payer's answer does not fit a known path |
| Payment posting and underpayments | A remittance or deposit arrives | Matches and posts, compares payment to contract | 835, bank file, contract terms | A payment does not match or a contract is disputed |
| Eligibility verification | A visit is scheduled | Checks coverage and flags gaps before the visit | X12 270/271 | Coverage is complex or there is more than one plan |
| Claims scrubbing and coding | A note is signed or a claim is built | Suggests codes, runs payer and NCCI edits, fixes clear errors | FHIR DocumentReference, 837 | Audit risk, low confidence, non-standard procedures |
| Prior authorization and patient collections | An order is placed or a balance posts | Assembles the request, or sends estimates and payment plan options | FHIR, X12 278, patient messages | Medical necessity, financial hardship or a complaint |
See also our prior authorization automation work and our guide to payment posting software.
How a denial agent should work
A good denial agent files only what its playbook covers and sends the rest to people. This flow is an illustrative design.
Which vendors provide AI agents for healthcare revenue cycle operations and claims management?
Waystar, R1, Ensemble, UiPath and FinThrive have all publicly described agentic products or programs for the revenue cycle. Every claim in this table is the company's own and has not been independently tested, and we did not rank them because we found no independent head-to-head test.
That also answers the question of the best agentic AI platform for revenue cycle management in healthcare: no AI-powered revenue cycle automation platform has been shown to be best for all healthcare organizations. The platforms available today differ by payer connections and by how much they act without staff. Some vendors call theirs an agentic revenue operating system, so look past labels and "leading innovation" claims, and ask which platforms health systems are using to reduce denials and improve collections.
| Company | What it has said publicly | Ask them |
|---|---|---|
| Waystar | Announced agentic AI in January 2026 and expanded its Google Cloud work in March 2026 using Gemini models. It says AltitudeAI has prevented over $15 billion in denials. | Which workflows act on their own today, and how was "prevented" measured? |
| R1 | Launched the R37 lab with Palantir in March 2025 and expected to deploy "agentic RCM worker" solutions to enterprise customers in the second half of 2025. | Which of those are live, and at how many customers? |
| Ensemble Health Partners | Announced an agentic platform with Cohere, starting with insurance reimbursement. It expected agents to manage at least 40 percent of A/R follow-up tasks at first. | Is that expectation now a measured result? Can you buy it without outsourcing? |
| UiPath | Announced agentic healthcare solutions at ViVE 2026. Its site lists prior authorization and claim denial resolution. | How does it hold up when a payer portal changes? |
| FinThrive | Its newsroom lists an agentic AI RCM platform shown at HIMSS in March 2026, with 50 or more AI and automation use cases. | How many of those use cases are agents acting without staff? |
| Experian Health | Sells AI Advantage for denial prediction and Patient Access Curator. We did not find an agent launch that fits this definition. | What acts automatically, and what only scores? |
| Adonis | Raised a $31 million Series B in June 2024 for revenue cycle intelligence and automation. We did not confirm its current agent features. | Which actions run without staff approval? |
| Ember | Raised $4.3 million in seed funding in November 2025 for denial prevention and reports over 57 percent fewer denials. | To evaluate Ember on AI RCM, ask for the baseline, the customer count and the payer mix behind that figure. |
| Nirmitee | An engineering partner that builds custom integrations and agents. We have no packaged agent product and no published production RCM agent result, so choose us for a build, not for a ready-made platform. | Ask to see our open-source repositories and how we would log and limit an agent. |
Search results often connect certain companies to deals we could not confirm. We found no public announcement of an Experian Health and UiPath partnership or case study, no Experian Health automation investments in RPA, and no Waystar investment in RPA automation for the revenue cycle, though absence from our search is not proof. For digital transformation claims of this kind, ask for the press release or a customer you can call. We did not review Availity's RPA and prior authorization automation, or Operator Labs or its alternatives for revenue cycle phone work. For wider lists, see our reviews of top RCM companies and RCM software.
How to evaluate AI-driven RCM vendors for complex billing environments
Test the vendor on your own denied claims, with your payers, and compare against a baseline you measured first. Published rankings for "most secure" or "most scalable" AI revenue cycle management software are mostly marketing, so check each point yourself. These are the capabilities and criteria we would use when evaluating agentic AI, whichever brands or products reach your shortlist.
| What you are judging | What to ask for |
|---|---|
| Security and data security | A signed BAA, including any model provider. Current third-party audit reports. Limits on which patient records each agent can open. A log of every read and write. |
| Scalability | Performance at your claim volume across multi-location sites and specialties. What happens during a payer outage. How rework grows with volume. |
| Functionality | Which actions finish without staff. How many of your payers and EHR versions, including Epic, are live today. Whether it integrates by API or by screen scraping. |
| After-sales support | Who fixes a broken payer path, how fast, and whether you can reach them when volumes spike. |
| Proof | Customer references with dates, baselines and payer mix. Cost per claim, days in A/R, denial write-offs and overturn rate, measured before and after. |
Which RCM agent reduces days in A/R fastest? We found no independent comparison that can answer this, so treat any single winner with caution. A short test on your own aged claims, measured against a holdout group, will tell you more. The same goes for AI voice agent services for businesses that promise a 30 percent cut in cost per claim: ask for case studies showing the reductions, with the ROI math.
To compare the operational efficiency of various AI agents across US healthcare revenue cycles, use the same claims, the same period and the same definition of a staff touch for each one.
Marketing tools for multi-location healthcare campaigns are a different category. Keep them out of this comparison, even when they use the word "autonomous".
Autonomous RCM agents vs offshore staff for eligibility verification
We found no independent 2026 study comparing accuracy, HIPAA compliance and cost for insurance eligibility verification, so compare the two on your own sample. Both need a BAA and access limits. Agents scale with volume and log every step, and trained staff handle odd coverage questions better until an agent has proven otherwise. Hiring decisions should wait for that result.
The technology stack behind agentic RCM
An agent needs four layers: data connections, models and rules, agent control, and oversight. In our view most of the effort sits in the first layer, because US billing runs on a mix of X12 files, FHIR APIs and payer portals.
Analyze these technical requirements before deploying an agent. Elective care adds one more: revenue attribution, so each payment and patient balance ties back to the right service and location.
Language models read notes and policies, while traditional prediction models often suit scoring tasks. Control means a durable record of where each claim stands, a fixed list of actions the agent may take, and a clear handoff to people. Read more on how AI agents integrate with EHR systems, our X12 EDI integration services and how to build HIPAA compliant AI agents.
Rules that shape agentic RCM in 2026
Federal rules set payer prior authorization deadlines and APIs, and several states now limit how insurers can use AI in denials. These state laws bind payers and review agents, so check with counsel how they affect a provider-side agent.
| Rule | What it requires | Status |
|---|---|---|
| CMS-0057-F | Covered payers decide in 72 hours for urgent and seven calendar days for standard requests, give a specific denial reason, publish metrics and build FHIR APIs. Exchange plans are excluded from the timeframes. | Operations from 1 January 2026, first metrics by 31 March 2026, APIs generally from 1 January 2027 |
| California SB 1120 | Medical necessity decisions must be made by a licensed healthcare professional, with safeguards on AI used in utilization review. | Summarized by Morgan Lewis, bill text not opened |
| Texas SB 815 | A utilization review agent may not use an automated decision system to make an adverse determination, wholly or partly. Administrative support and fraud detection are allowed. | Effective 1 September 2025 |
| Arizona HB 2175 | A medical director must review medical necessity denials using independent judgment. | Reported in force 1 July 2026 |
| Maryland HB 820 | Sets conditions on AI and software tools used in utilization review. | Approved May 2025, Chapter 747. We read the synopsis only. |
On HIPAA, we found no AI-specific rule. The existing requirements still apply: a HIPAA compliant agent needs a BAA with each vendor that handles PHI, access limited to the minimum necessary, and audit controls under 45 CFR 164.312(b).
What is a realistic implementation timeline for an AI RCM platform?
We found no published benchmark for how quickly a health system can deploy AI agents for revenue cycle work, so plan by checkpoints and treat a fixed promise as a sales claim. McKinsey expects leaders to move from pilots to production-scale deployments over the next two to three years, and that is an industry forecast, not a project plan.
Your time depends mostly on the data connections, payer count and how fast your compliance team can review. Move forward only when each checkpoint passes. This order is our judgment:
- Baseline. Record denial rate, days in A/R, cost to collect and overturn rate for the one task you chose.
- Connect. Get read access to the data, sign the BAA, and write down what the agent may never do.
- Shadow. The agent runs alongside staff and takes no action. Compare its answers to theirs.
- Supervised. The agent acts and staff approve or spot-check every result.
- Limited autonomy. The agent acts on low-value, rule-based cases. Keep a holdout group so the ROI is measurable.
- Expand. Add a payer or a task. McKinsey warns against getting stuck in "pilot purgatory", so set the scale-up test before the pilot starts.
Do not trust ROI figures without a baseline. Ask how the savings split between hard cost, recovered revenue and staff time moved to other work.
How Nirmitee can help
We build the integrations and agent controls for teams that need something a packaged platform does not cover. We are ISO 27001:2022 certified and sign BAAs. Public work you can inspect includes our open-source Da Vinci CRD, DTR and PAS servers tested with Inferno, and an X12 clearinghouse simulator. We have also built 350+ Mirth Connect interfaces. We do not run billing operations, and we have not published a production RCM agent result.
Choosing your first revenue cycle agent? Talk to our engineers about your denial data and systems. We will tell you which task we would start with and what it would need, including when buying a platform makes more sense.
Ready to scale?
Talk to our healthcare engineering team about building, integrating, and shipping faster.
Frequently Asked Questions
What is agentic AI in revenue cycle management?
How much can agentic AI save on revenue cycle costs?
Is agentic AI for RCM HIPAA compliant?
What is the difference between RPA and agentic AI in healthcare revenue cycle?
Which RCM functions should we automate with agentic AI first?
Which vendors provide AI agents for healthcare revenue cycle operations and claims management?
How quickly can a health system deploy AI agents for revenue cycle work?
Can AI agents replace revenue cycle staff?


