Public research brief · July 2026
Agentic AI for Healthcare Revenue Cycle
Pressures on the cycle. Problems across front, middle, and back. How agents — implemented securely — change the process. Not a chatbot on a workqueue.
Abigail Stafford · not legal or clinical advice
The argument
Revenue cycle work is information-heavy, rules-heavy, and chronically understaffed. Agentic AI is a supervised coworker that plans, retrieves evidence, calls tools, and completes steps under policy. Done securely, it reduces burnout and leakage. Done carelessly, it creates audit, privacy, and patient-harm risk. Security and correctness are the product.
~11.8%
Initial denial rate (2024, Kodiak)
$48.4B
Net leakage, Kodiak platform (2025)
39/wk
Avg PAs per physician (AMA)
30–60%
Potential cost-to-collect reduction*
*McKinsey agentic / touchless RCM analysis — directional at full scale, not a guarantee.
Practitioner lens
Written from the floor of the revenue cycle, not only from market research. Healthcare is drowning in RCM work that agents can finally help with — and almost nobody has deep experience on both sides of that sentence: the denial code and the control loop.
Field vignette · anonymized
On a denials desk, the same medical-necessity denial would land fifty times in a week — different patients, same missing clinical attachment and the same filing clock. The work was not “write a better letter.” It was find the note, cite the policy, prioritize by dollars and deadline, and stop the root cause from regenerating. A secure agent would have prepared the packet. A human would still have signed the appeal. That line — propose vs dispose — is how revenue cycle has always been done when it is done correctly.
Where to start
Denial triage + cited appeals
Highest authenticity from ground-truth denials work; clear human-in-the-loop; measurable overturn and hours saved.
Underpayment / contract variance
Scarcer skill set. Silent leakage. Strong differentiation from generic coding bots.
Epic readiness / exception analytics
Trusted KPIs, exception queues, reversible automation — the same governance agents need.
The design rule
The agent proposes. A human disposes. Least-privilege tools. Evidence travels with the action — or the action does not fire. Audit trails on both paths. That is not a product feature. It is how this work has always been done when it is done correctly.
See it run on synthetic data in the Agent Control Room, or read the org design on GitHub.