Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 2, 4, 5 and 6 need a person. Step 4 is waiting for approval right now. Click it to open the full walkthrough.
Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.
How a Healthcare Provider Built a Claims Pipeline That Codes, Checks and Sends with a Coder in the Loop
Reads what the clinician documented, codes it, checks it against the payer's own rules before it is sent, and stops at the claims a person has to decide. Written for a regional healthcare system where the revenue cycle was losing to rework rather than to volume: 200,000 claims a month across 12 facilities, 72 out of every 100 paid first time, and the other 28 coming back to be done again. Six stations, run as a loop: what station 6 learns from a quarter of denials becomes the rule station 3 checks against.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Charge Reading
Nobody is asked anything hereSubagents read every closed encounter, pull out the services, diagnoses and modifiers the note supports, and hold anything the note does not carry. Nothing waits on a person here. A charge reaches a coder only when station 2 cannot settle it.
Code Assigning
A person answers hereSubagents suggest the codes the note carries and show the line each one came from, so a certified coder reviews a suggestion instead of starting from an empty field. A coder signs every code. Nothing is submitted on a suggestion alone.
Claim Checking
Nobody is asked anything hereSubagents test every coded claim against the payer's own edits, check that nothing required is missing, and score how likely this payer is to deny one like it. A person sees only the claims the pipeline expects to lose. Everything it expects to be paid goes out.
Claim Releasing
A person answers hereSubagents read every held claim back against the rule that held it, draft the correction the rule supports, and stop. This is the station where a human works. Everything else exists to make this decision small and well-lit.
Denial Working
A person answers hereSubagents sort every returned claim by its reason code, separate what can be corrected from what cannot, and draft the appeal that has worked on this reason before. A person decides what to appeal. The draft is ready either way.
Payer Learning
A person answers hereSubagents count every denial by payer and reason, watch what each payer has published since, and propose the validation rule that would have caught it before it was sent. A person accepts or rejects each proposed rule. Nothing about what gets sent changes on its own.
Where a Person Decides
Operating Model
This changes how work flows through the team.
What Transfers, What Must Be True
Failure Modes
What breaks this pattern:
Nothing holds a charge when the note falls short. The claim pays, an audit later finds the gap, and the practice refunds the money with penalties.
A suggestion becomes a billed code without a certified coder's name on it. When a payer questions the code, no person can say why it was chosen.
A claim goes out for a service that needed authorization first. The payer denies it, and the money is gone, because authorization cannot be granted after the fact.
The system writes a blocking rule from a single denial. One payer mistake becomes a standing rule, and clean claims get held for a reason that was never real.
Directional Outcomes
What the agent counts, and the station that counts it.
These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.
What Runs Where
Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.
charge-readclaims/captured/<enc-id>.jsonnote-does-not-supportclaims/**, ehr/** read-onlycode-suggestclaims/coded/<enc-id>.jsonawait-coder-signatureclaims/coded/**, ehr/** read-onlypre-submitclaims/checked/<claim-id>.jsonmissing-prior-authorizationclaims/**, payers/** read-onlyhold-briefclaims/decisions/<claim-id>.jsonawait-human-approvalclaims/decisions/**, everything else read-onlydenial-workclaims/denials/<claim-id>.jsonpropose-appealclaims/denials/**, ehr/** read-onlypayer-readanalytics/payers/<period>.jsonpropose-rule-changeanalytics/**, payers/** read-onlyStack
Every system this agent reads or writes.