The short answer: The useful question is not whether AI belongs in the revenue cycle. It is which decisions an agent may make without a person. A defensible line for 2026: an agent may retrieve, prepare, draft and rank. A person approves anything that changes what a payer or a patient sees, anything that moves money and anything that closes a balance. Advantum Health runs AI inside its own platform, operated by its own staff, which is what makes that line enforceable rather than aspirational.
Why this matters: autonomy is a liability decision, not a technology decision
Most AI governance conversations in healthcare start with the model. That is the wrong starting point for the revenue cycle, because the risk here is rarely that the model is unsophisticated. The risk is that a confident output becomes an irreversible action against a payer or a patient before anyone reads it.
The pressure to move is real. Deloitte’s 2026 health care outlook survey found more than 80 percent of health systems prioritizing agentic AI for revenue cycle management, and HFMA’s research on the revenue cycle of the future found 27 percent of finance organizations deploying AI at scale with another 53 percent piloting. Deployment is outrunning the governance built around it.
A submitted claim is a representation. An appeal is an assertion about clinical facts. A write-off closes a balance the organization had a right to collect. A patient communication about money lands on someone already under financial stress. Each of these carries consequences that do not reverse cleanly, and none of them cares whether the decision was made by a person or a process.
So the governing question is not capability. It is reversibility. The more expensive a decision is to undo, the more a person belongs in front of it.
A working framework: five levels of autonomy
Borrowing the structure used elsewhere for autonomous systems makes the tradeoff explicit. Assign every workflow a level and write it down.
Level 0, observe. The system reports. Dashboards, trend detection, anomaly flags. No decision, no action. Almost no risk, and often more value than organizations expect.
Level 1, prepare. The system retrieves and assembles. Pulling payer policy, gathering documentation, reconciling a remittance to a claim. A person still decides. This is where the largest share of realistic revenue cycle time savings sits.
Level 2, recommend. The system ranks or drafts and a person chooses. Denial worklists ordered by recoverable value, appeal letters drafted for review, coding queries suggested. Errors surface at the review step.
Level 3, act with review. The system executes but a person approves before anything leaves. Appropriate for high-volume, low-variance work with a clear approval gate: status checks, standard-form submissions, routine follow-up sequences.
Level 4, act without review. The system decides and acts with no human in the loop, monitored only in aggregate. Defensible for a narrow set of reversible, low-value, high-volume tasks. Not defensible today for anything that asserts clinical facts, closes a balance or contacts a patient about money.
Where the line should sit today
Coding. Agent-assisted, human-assigned. A code is a representation to a federal payer. Automated code suggestion is well established and useful. Automated code submission without review moves a compliance decision into a system that cannot be deposed.
Appeals. Draft autonomously, submit with review. Generative drafting is genuinely good at the structure of an appeal. It is not reliable at asserting what the record actually says, and an appeal that misstates a clinical fact damages more than one claim.
Patient communication about balances. Keep a person on approval. Financial messages reach people in hardship, in bereavement and in active treatment. The cost of a tone-deaf automated message is not measured in the balance.
Write-offs and adjustments. Human approval, with a threshold. An agent may recommend which balances are uncollectible. Closing them is a decision with revenue integrity and audit implications.
Eligibility, status checks and data assembly. Reasonable candidates for higher autonomy. They are reversible, they are verifiable against the source and being wrong produces a retry rather than a representation.
What governance actually requires
A policy that says AI is used responsibly is not governance. Five things make it operational.
Traceability. For any output, you can reconstruct what the system saw, what it produced and who approved it. If you cannot answer an auditor’s question about a specific claim, the autonomy level was set too high.
Reversibility. Before raising an autonomy level, ask what undoing a wrong action costs. If the answer is a payer conversation, keep the person.
A named owner. Every automated workflow has a person accountable for its output, by name, not a department.
Measurement against a baseline. Accuracy and rework rates tracked against what the process produced before automation. Without a baseline, you cannot tell improvement from drift.
An escalation path. Defined behavior when the system is uncertain. A system with no abstention behavior is not cautious. It is confidently wrong at volume.
Common mistakes
Setting autonomy by vendor capability rather than by consequence. What the tool can do is not the question. What it should do in your process is.
Letting autonomy creep. A pilot at Level 2 quietly becomes Level 4 because review was slowing things down. Review the levels on a schedule.
Treating payer AI as a reason to abandon your own standards. Payers automating adjudication is a real pressure. It is not an argument for submitting unreviewed assertions.
Writing the policy after deployment. The autonomy level is cheap to set before go-live and expensive to lower afterward.
A checklist for revenue cycle leaders
- Inventory every AI-touched workflow and assign each an explicit autonomy level.
- Name an accountable owner for each one.
- Document what the system does when confidence is low.
- Confirm every output is logged and attributable to an approver.
- Set a reversibility test before raising any workflow’s autonomy level.
- Keep human approval on coding assignment, appeal submission, write-offs and patient financial communication.
- Baseline accuracy and rework before automation, then track both after.
- Review autonomy levels quarterly and record the rationale for any change.
- Confirm your vendor agreements and business associate agreements cover automated actions taken on your behalf.
Draw the line before the technology draws it for you
Autonomy in the revenue cycle will keep increasing, and it should. The workflows that consume the most staff time are the ones best suited to it. But the direction of travel is not an argument for skipping the decision about where to stop.
Organizations that write the line down keep control of it. Organizations that do not will find it drawn by whichever vendor ships the most aggressive default.
Advantum Health’s structure reflects that view. The AI in our platform supports our own experienced staff rather than replacing their judgment, and clients do not operate the models themselves. That is not a limitation of the technology. It is where accountability has to sit when the output is a representation to a payer.
Part three looks at what this means for how organizations should think about automation and people together. If you are evaluating an agentic tool now, start by writing down which of the five levels you are actually buying, then check it against a revenue cycle assessment of the workflow it would touch.