AI agents still need human oversight

Artificial intelligence (AI) and automation are already changing healthcare revenue operations. But the useful question is not whether technology can perform more tasks. It is which work should be automated, what information the technology can reliably evaluate, and who owns the decision when an account falls outside the routine path.

The opportunity is substantial. The Council for Affordable Quality Healthcare (CAQH) 2024 Index estimated that electronic administrative transactions helped the medical and dental industries avoid $222 billion in annual administrative costs, with another $20 billion in potential savings from moving remaining manual and partially electronic transactions to fully electronic workflows. The report also estimated that medical providers could save an average of 18 minutes per claim-status inquiry through electronic transactions. CAQH

These are industry estimates for electronic administrative transactions, not savings attributable specifically to AI agents. They establish why reducing repetitive administrative work matters.

Routine tasks offer a useful starting point, but clear inputs do not make either systems or people infallible.

A U.S. Department of Health and Human Services (HHS) Office of Inspector General (OIG) review provides a useful example. In its review of a sample of Medicare Advantage denials issued in 2019, OIG found that 13% of denied prior-authorization requests met Medicare coverage rules. It also found that 18% of denied payment requests met both Medicare coverage rules and the Medicare Advantage organization’s billing rules. Among the causes identified were both manual-review errors and system-processing errors. OIG HHS

This was a historical review of denials issued during June 1–7, 2019, not a study of today’s AI agents. It illustrates how both manual review and system processing can produce errors.

Repetition is where automation earns its place

Eligibility checks, claim-status inquiries, routing work to a queue, retrieving documents, identifying missing information, and organizing account history are good candidates for automation because the inputs and expected actions are relatively structured.

Artificial intelligence can extend that further. A 2025 article in The American Journal of Managed Care (AJMC), written by CAQH researchers, describes revenue-cycle applications that include identifying claim errors before submission, recognizing services that may require prior authorization, retrieving supporting documentation from the electronic health record (EHR), and identifying conditions that could lead to a denial before the claim reaches the payer. AJMC

Surfacing a missing authorization, inconsistent eligibility response, required attachment, or enrollment exception before submission gives the team an opportunity to address it earlier. Those tasks depend on reliable inputs and a defined next action.

The account rarely tells its story in one field

A denial code identifies the payer response, but it may not capture the full history. An authorization exception can still require review of the approved dates, previous payer actions, submitted documentation, and the client’s instructions before the next step is clear.

The Artificial Intelligence Risk Management Framework from the National Institute of Standards and Technology (NIST) specifically addresses this issue. NIST notes that organizations using AI in operational settings should clearly define human roles and responsibilities for decision-making and system oversight. It also warns that data-driven systems can lose important context when complex situations are reduced to measurable fields or model inputs. NIST AI Resource Center

Technology can bring the relevant history together. The reviewer still needs to interpret how the dates, responses, and instructions relate to the service and the action now required.

Prior authorization is becoming more structured

The Centers for Medicare & Medicaid Services (CMS) Interoperability and Prior Authorization Final Rule requires affected payers to provide a specific reason for denied non-drug prior-authorization decisions, beginning in 2026. It also sets decision limits of 72 hours for expedited requests and seven calendar days for standard requests for impacted payers other than Qualified Health Plan issuers on the Federally Facilitated Exchanges.

Prior-authorization application programming interface (API) requirements generally begin in 2027, with exact compliance dates varying by payer type. The APIs must communicate documentation requirements and whether a request is approved, denied with a specific reason, or needs more information. These requirements apply to the payer groups covered by the rule, rather than every health plan. CMS

Defined documentation requirements, specific denial reasons, and clear requests for additional information give an automated workflow more reliable inputs than a generic status. The task still needs a defined review path when those inputs are incomplete or inconsistent.

The 2025 CAQH/AJMC analysis makes the same point: data consistency and standardization are essential if healthcare organizations expect AI to perform reliably across revenue-cycle workflows. AJMC

Human oversight should be designed into the workflow

Automation can gather account history, surface inconsistencies, and route routine tasks. Review responsibility should be defined before an exception arrives, with enough context for the person receiving it to understand why it needs attention.

Escalate an account when it requires:

• Interpretation of conflicting information or an unclear payer requirement;
• Client direction, payer communication, or financial judgment;
• Review of an approaching filing or appeal deadline;
• A decision about the correction or next action.

For example, a workflow may retrieve a payer status and summarize the response. The reviewer still needs to decide whether that response calls for a correction, a document, an appeal, or a later follow-up. A completed task alone does not establish progress toward payment.

See how Alico uses technology and oversight within its revenue services.

The measure is not how many tasks AI completed

A high transaction count can look impressive while unresolved revenue continues to age. More useful measures show whether automation reduces unnecessary touches, brings exceptions forward earlier, shortens the time to a meaningful next action, and gives staff better information for decisions.

The strongest use of automation in revenue cycle work is not removing people from every decision. It is giving people more time and better information for the decisions that require them.

All insights