From RPA to agentic AI: rebuilding the revenue cycle

If you lead a healthcare operations or revenue cycle team, you almost certainly already run automation across your claims processes. The question is no longer whether to automate the revenue cycle, but why an automated revenue cycle still leaks revenue. The initial claim denial rate reached 11.81% in 2024, a 2.4% year-on-year increase, according to Kodiak Solutions. Most of those claims are eventually paid, but only after healthcare organizations spend heavily to overturn them.

This is the gap that most discussions of RPA vs agentic AI healthcare automation tend to miss. The problem is not a lack of automation. It is that the automation in place was built for a revenue cycle that does not stand still.

How the revenue cycle runs today: RPA bots on rails #

The current approach in most mid-market provider organisations is robotic process automation (RPA). A bot is configured for each payer and each workflow: it logs into a portal, runs a real time eligibility check, files a prior authorization request, posts a payment, or submits a claim status inquiry. Revenue cycle management contains a high volume of repetitive tasks that are rules-based, which is exactly why RPA became the default among automation tools for the job, as TechTarget documents.

Within those boundaries, RPA works well. It is fast, deterministic, and consistent. It does the same task the same way for every patient, every time, and it does not stop at five o'clock. For stable, high-volume transactions, this is the correct tool, and nothing about agentic AI changes that.

Where deterministic automation breaks #

The limitation is structural rather than a matter of quality. An RPA bot follows a fixed path. When anything on that path changes, the bot fails. As one revenue cycle technology analysis puts it, a bot can click a field on a payer portal, but if the spacing of the website components shifts, the bot can break and require costly maintenance to correct, according to AKASA. The same brittleness appears when a payer changes a portal layout or a PDF template shifts, as Nanonets Health notes.

Payers change these things constantly. Every changed rule, every reformatted document, and every new medical-necessity requirement is a potential break point, and each break either halts the bot or pushes the case into a manual exception queue. That queue is where rework cost accumulates.

The administrative cost to fight a single denied claim rose from $43.84 in 2022 to $57.23 in 2023, and across the industry providers spent an estimated $25.7 billion in 2023 contesting denials, of which roughly $18 billion was potentially wasted arguing over claims that should have been paid at submission, according to Premier, Inc. RPA does not address this cost, because the work that generates it is precisely the work RPA cannot do.

What agentic AI adds: reasoning over the exception, not just the rule #

The distinction that matters for healthcare claim denial automation is between executing a rule and reasoning about an exception. RPA executes a fixed sequence. An agentic system plans, retrieves information from multiple sources, adapts when an input changes, and works through the cases that rules-based automation rejects. Rather than breaking when a portal changes, an agent can read the changed page and proceed; rather than dropping a denied claim into a queue, it can investigate the denial reason, gather the supporting clinical documents, and draft the appeal letters.

An RPA bot does what you told it to do last quarter. An agent works out what the payer is asking for today. In the revenue cycle, that difference is the entire exception queue, which is where the lost revenue actually sits. Antoni Kozelski, CEO and Founder, Vstorm

The two approaches are best understood side by side.

For a deeper explanation of how retrieval and multi-step reasoning combine in practice, see our write-up on agentic RAG.

The transition is layered, not a rip-and-replace #

The transition from RPA to agentic AI is not a matter of switching one off and the other on. The two are complementary, and treating the move as a wholesale replacement is how organisations introduce unnecessary risk. RPA should keep running the stable, high-volume, rules-based transactions it already handles well: eligibility checks against unchanged payers, clean-claim submission, routine payment posting.

Agentic AI layers on top, taking the work that sits beyond the reach of a fixed rule: denial root-cause analysis, appeals drafting, medical-necessity documentation, and any workflow where payer behaviour varies enough to break a bot. The result is a revenue cycle where deterministic automation handles what is predictable, and agentic reasoning handles what is not. Each layer does the work it is suited to, reducing manual intervention in the exception queue and freeing the team managing denials to focus on the cases that genuinely require judgement.

A production deployment inside a US healthcare provider #

The relevant question for any provider is whether agentic AI can run inside a real, compliance-bound healthcare environment, not just in a demonstration. We at Vstorm built a multi-channel AI Agent for a US healthcare provider serving more than 100,000 members across multiple states. The system met strict healthcare compliance requirements, saved each doctor more than five hours per week, and lifted patient engagement by over 20%.

That deployment automated pre-appointment scheduling rather than the revenue cycle, and we are precise about that distinction. What it demonstrates is the harder part of the problem: an agentic system operating reliably and compliantly inside a live US provider, integrated with existing infrastructure. The same delivery discipline applies whether the workflow is scheduling or claim denial management. For more on our work in this sector, see our agentic AI in healthcare page.

What still needs a human, and what governance the transition requires #

Agentic AI in the revenue cycle is not autonomous in the sense of unsupervised. Clinical-judgment decisions, ambiguous medical-necessity calls, and final appeal sign-off belong with people. The agent compresses the work that precedes those decisions; it does not remove the decision-maker.

Governance is the precondition, not an afterthought. Protected health information must stay within its minimum-necessary scope, business associate agreements must cover every system the agent touches, and every action taken by the artificial intelligence (AI) must be observable and auditable. These controls are also where many healthcare AI efforts quietly fail. The opportunity is real: the industry could save more than $20 billion annually by fully automating the administrative transactions still handled manually or semi-electronically, according to the 2025 CAQH Index. Capturing it depends on building the governance layer correctly, not on deploying agents quickly.

Where to start: one workflow, measured #

The practical path is narrow and measured. Choose a single workflow with a high rework cost, denial management being the obvious candidate, and establish the baseline: how many cases enter the exception queue, what each one costs to resolve, and how many are abandoned. Layer an agent onto that one workflow, measure the same figures again, and prove the return before extending to the next. This sequence keeps the risk bounded and the business case grounded in your own numbers rather than industry averages.

The revenue cycle was automated once already. Rebuilding it with agentic AI for revenue cycle management is the second pass: keeping the deterministic automation that works, and adding reasoning where deterministic automation never could.