LAB NOTES
Revenue cycle AI creates a lot of questions that don’t fit neatly inside a case study.
Lab Notes are where I pull those questions apart: evaluation, agent behavior, QA, domain judgment, failure modes, edge cases, and the difference between a system that works in theory and one that survives actual operations.
LESS “THOUGHT LEADERSHIP.” MORE FIELD NOTES.
WHY LAB NOTES
I’m interested in the places where the workflow gets ambiguous, the evidence gets incomplete, the agent gets confident anyway, or the technically correct answer turns out to be operationally useless.
Those moments are not side effects of the work. They’re often where the most useful learning happens.
FEATURED NOTE
A completed workflow is not necessarily a correct workflow.
If an agent performs every expected action but uses the wrong evidence, misunderstands the account state, or leaves the underlying problem unresolved, did it actually succeed?
Read note →Task completion measures whether the system finished. RCM QA has to determine whether the work was right.
ALL NOTES
RCM, agents, QA, evidence, edge cases, and whatever else refuses to behave cleanly.
Completing the documented workflow only proves that the workflow ran. It does not prove the conclusion was supported, the action was appropriate, or the actual revenue-cycle problem moved toward resolution.
Read noteModels are very good at generating answers that sound reasonable. Revenue cycle work asks a harder question: what does the available evidence actually support?
Read noteThe right action at the wrong time is still wrong. So is the right conclusion with unsupported evidence, or a successfully completed task that changes absolutely nothing about the underlying account.
Read noteSometimes the correct answer is not a conclusion. It’s recognizing that the available information cannot safely support one yet.
In the labIf every difficult case quietly lands on a specialist’s desk, the workflow may technically contain human oversight without actually having designed a meaningful human role.
In the labTwo systems can fail at the same rate and create completely different operational risks. What failed, why it failed, and whether we can detect it matter just as much as the score.
In the labWHAT THIS ISN’T
I’m much more interested in what happens when the future gets assigned a payer portal and asked to work a denial.
These notes are grounded in operational questions: what works, what breaks, what evidence matters, what specialists notice, and what useful AI actually has to understand.