Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 2, 5 and 6 need a person. Step 5 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 Growth-Stage Healthcare Provider Built Visit Follow-Through Where a Billing Manager Decides Every Appeal
One patient visit, from the evening before it happens to the day its claim is paid. The agents confirm the visit, prep the chart, draft the note from the room's own recording, code what the signed words support, keep the patient informed, and work what the payers send back. A provider signs every note, a billing manager decides every appeal and write-off, and nothing clinical is ever answered by an agent. Six stations, run as a loop: the denials station 5 sorts are the rules station 6 proposes, and an accepted rule is what station 3 codes with on Monday.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Visit Prep
Nobody is asked anything hereAgents confirm every visit on tomorrow's schedule, collect the intake forms, pull each chart from the EHR, check coverage with each plan, and hold anything that does not add up. Nothing waits on a person here. A patient nobody can reach, or a chart with two birthdates, is handed to the desk with its reason attached, never guessed into shape.
The Visit Note
A person answers hereAgents draft each visit note in SOAP shape from the room's recording, carry every drug and dose exactly as spoken, check the draft against the chart, and leave blank what they could not hear. This is where a person works. The provider reads, corrects and signs every note, and nothing downstream moves without that signature.
Coding and Claims
Nobody is asked anything hereAgents read each signed note, match every diagnosis and procedure to the code its words support, check each claim against that payer's own edits, and submit the clean ones the same day. Nothing waits on a person here. A code the note does not support goes back to the provider as a question, on the same screen where notes are signed.
Patient Follow-Through
Nobody is asked anything hereAgents send each patient their summary in plain words, book the next visit the plan names, pass on results the provider has released, and hold every clinical question for the provider who owns it. Nothing waits on a person here. A clinical question is held for the provider with the chart open beside it, never answered by an agent.
Payer Response
A person answers hereAgents post every payer response against its claim, match each payment to the contracted rate, sort every denial by its stated reason, and draft the resubmission or appeal the note supports. This is where a person works. Resubmissions, appeals and write-offs leave only on the billing manager's decision.
Week Review
A person answers hereAgents close the week's record, trace every visit from prep to payment, check every chart open against the visit behind it, and propose the rules that would have stopped this week's repeated denials. A person accepts or rejects each rule. What the practice tells payers, and what patients hear from it, is not something an agent changes quietly.
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:
The recording drops a number and the note writer fills in a likely one. The chart now carries a dose nobody said, and a person could take it.
A code the note does not support goes out anyway. Each claim like it sits in the payer's file as billing the practice cannot defend, and it stacks up until an audit finds it.
A message goes to contact details that match a different patient's record. That is a privacy breach the moment it sends, and the practice reports it, not the software.
A gap in the chart gets filled with the system's best guess instead of waiting at the desk. The provider walks into the room trusting numbers nobody checked.
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.
prep-deskvisits/<visit-id>/prep.jsonhold-for-the-deskvisits/**, the schedule and the EHR read-onlynote-draftvisits/<visit-id>/note-draft.mdawait-provider-signaturevisits/<visit-id>/**, the chart read-onlycode-and-claimclaims/<claim-id>.jsoncode-unsupported-by-noteclaims/**, the signed notes read-onlypatient-messagesmessages/outbound/<msg-id>.jsonclinical-question-heldmessages/**, the portal, the chart read-onlyremit-and-denialsclaims/<claim-id>/response.json, denials/<claim-id>.mdawait-your-decisiondenials/**, claims/**, the contracts read-onlyclose-the-weekrules/proposed/<rule-id>.json, reports/week-<date>.jsonaccept-rule-changerules/proposed/**, reports/**, the week's record read-onlyStack
Every system this agent reads or writes.