AI helps most with bounded, checkable revenue cycle tasks: documentation improvement, coding support, eligibility and claim-status checks and drafting appeal letters. People must stay in charge of anything that decides, denies or delays care or payment on medical grounds, final coding on complex or contested cases, and every payer dispute that does not fit a rule. The working model is a named human reviewer for each AI task, an exception queue and an audit trail.
Which revenue cycle tasks are mature enough to use today?
Maturity differs by task. Documentation improvement, coding support and defined back-office tasks are proven at task level. Agents that work claims and appeals across several systems, with audit trails, are emerging. An AI that resolves every payer dispute without human exception handling is speculative. Adoption data points the same way: in a 2025 HFMA and Solventum survey of 272 provider leaders, about half said their organizations had implemented AI-driven mid-revenue-cycle solutions, and roughly one-third were testing, evaluating or implementing coding automation. Those are provider responses, not a census of billing vendors. A February 2026 HFMA report found only about one in five providers used AI for denials management.
A practical map of task, maturity and who stays accountable:
- Eligibility and claim-status checks: proven at task level. The billing manager reviews the exception queue.
- Documentation improvement and coding suggestions: proven as suggestions. Certified coders accept, change or reject each one, and auditors sample the results.
- Denial sorting and appeal drafting: emerging, with a person in charge. An appeals specialist signs off, with a clinician where medical necessity is argued.
- Working claims across payer portals: emerging. It needs audit trails, access limits and a named owner for exceptions.
- Autonomous medical-necessity denials: speculative and high risk. Do not automate.
What do prior authorization rules say about AI?
CMS-0057-F requires affected payers to meet 72-hour urgent and seven-calendar-day standard prior authorization decision timeframes and to give specific denial reasons, generally from 1 January 2026. The required FHIR APIs generally begin on 1 January 2027, and the prior authorization provisions generally exclude drugs. In California, SB 1120, in force since 1 January 2025, says an AI tool in a health plan's utilization review may not itself deny, delay or modify medically necessary services. A competent licensed professional must make the medical-necessity determination.
For providers, the useful distinction is checking versus deciding. Bain and KLAS describe an Epic and Humana pilot in which 75% of coverage checks were processed automatically. A coverage check is not every authorization decision. For health insurers and administrators, the design rule is simple: AI can sort, summarize and check completeness, and a licensed clinician makes and owns the determination.
Physicians are wary. In an AMA survey reported in May 2026, six in ten physicians were concerned AI could increase prior authorization denials, and only 24% said medical-necessity denials were consistently reviewed by appropriately qualified clinicians. That is concern, not a measured rise in denials, but it tells you what clinicians will ask about any tool that touches authorizations.
Where must people stay in charge?
- Medical-necessity decisions, approvals and denials alike, belong to licensed clinicians. The medical director owns the policy for when AI output may inform one.
- Coding on complex, high-value or contested cases belongs to certified coders, who stay accountable for the claim.
- Appeal letters that make clinical arguments need a clinician's review. A drafted letter that cites the wrong date or diagnosis is worse than a late one.
- Payer disputes that match no rule go to an experienced specialist, not a retry loop.
- Patient balance and hardship conversations stay with trained staff, and the compliance officer owns the audit sample.
How do you read ROI claims about revenue cycle AI?
Read them as reports, not proof. Treat any return multiple you are shown, from a vendor or an analyst, as self-reported unless you can see how it was measured and against what baseline. The HFMA and Solventum survey shows what survey data looks like: 37% of leaders estimated documentation and coding discrepancies cost their organizations at least $1 million, and they named cost (57%), inadequate IT resources (55%) and uncertain ROI (41%) as the leading barriers. Those are estimates and opinions from the people surveyed.
For billing and RCM companies, this is also a sales question, because practice clients will ask what AI does to their claims and who is accountable. Measure before you automate: first-pass claim acceptance, days in accounts receivable, denial rate by reason, appeal overturn rate, cost to collect and coder accuracy on audit.
How should you set up the workflow?
Fix the workflow first. If a team re-keys one eligibility result into three systems, an AI step on top copies the mess faster. Good AI automation separates fixed rules, such as a payer rule that always applies, from judgment steps, such as reading a denial letter. It routes anything below a confidence threshold to an exception queue and logs who approved what. Confirm a business associate agreement with any vendor that handles claims data.
Next steps
- List every revenue cycle task where AI is proposed and label it proven, emerging or speculative.
- Name a human reviewer and a backup for each.
- Capture baseline measures before any pilot.
- Keep AI out of medical-necessity decisions.
- Review exception and override logs monthly with the compliance officer.
Frequently asked questions
Can AI submit claims without human review?
Bounded back-office tasks are proven at task level. Agents that work claims and appeals across systems are emerging and need audit trails and people handling exceptions. Sample and audit even the tasks you automate.
Can an insurer use AI to deny care?
Not on its own in California. SB 1120 requires a competent licensed professional to make the medical-necessity determination, and CMS-0057-F requires specific denial reasons from affected payers. Rules differ by state and program.
Will AI reduce billing headcount?
It depends on your volume, baseline and how much work moves to exceptions. Plan for roles to change before you plan for roles to shrink, and treat a headcount promise without a measured baseline with suspicion.