Seven-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3, 5 and 7 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 Billing That Codes Only What the Signed Note Supports
Everything around one patient visit, from the reminder that fills the slot to the payer's answer on the claim. The agents prep the chart, draft the note off the provider's own dictation, code only what the signed note supports, send the clean claims and hold the doubtful ones, and stop for a person exactly where the law or the money says a person signs. Seven stations, run as a loop: what the week's denials teach at station 7 changes how station 4 codes and how station 1 preps next week.
What Was Broken
What We Built
Seven stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Visit Confirmation
Nobody is asked anything hereAgents send the reminders, record the confirmations, verify each patient's coverage, and put a prepped chart behind every slot on tomorrow's book. Nothing waits on a person here. A coverage that does not verify is flagged on the schedule, never guessed at.
Visit Documentation
Nobody is asked anything hereAgents transcribe the provider's dictation, draft each note into its SOAP sections, check the draft against the chart, and leave blank what the audio does not settle. Nothing waits on a person here. A word the audio does not settle stays blank, because a guessed drug name is worse than a gap.
Note Signing
A person answers hereAgents queue each drafted note to the provider who saw the patient, put the flags and the blanks on top, and apply the spoken corrections line by line. This is where a person works. The signature is the fact the whole record hangs on, and no agent carries it.
Visit Coding
Nobody is asked anything hereAgents draft the diagnosis and procedure codes off the signed note's own sentences, check every pair against the payer's rules, and hold any code they cannot trace to a sentence. Nothing waits on a person here. A held code stops the claim it was bound for, and the claim carries the reason to station 5.
Claim Submission
A person answers hereAgents assemble each claim, scrub it against the payer's format, submit the clean ones, and hold the rest with the reason written on each. Denials come back traced to the line the payer names. This is where a person works. A held claim leaves fixed or not at all, and the fix is a decision.
Recall Booking
Nobody is asked anything hereAgents send each after-visit summary through the portal, confirm the prescriptions reached the pharmacy, book the follow-up the signed plan names, and route every clinical question to the provider who signed. Nothing waits on a person here. A clinical question is never answered in a provider's name, only routed to one.
Practice Review
A person answers hereAgents check every visit against where it ended, count what settled without a person, propose the rules that would have stopped the repeated denial, and draft the compliance report off the visit trail. A person accepts or rejects each rule. How a note is coded and what a claim carries are practice rules, and an agent never changes them 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 scribe hears an unclear word and writes its best guess instead of a blank. The provider signs the note without catching it, and a wrong drug or dose is now the record of care.
A code the note does not support goes onto the claim anyway. An auditor traces it back months later, and one untraceable code puts every similar claim under review.
The builder sends a claim that carries a held code or coverage nobody verified. The denial arrives weeks later, and staff spend that week's margin unwinding it.
The sender puts a diagnosis or a result into an ordinary text message. That is a reportable privacy breach, and the practice owns it, not the agent.
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.
visit-prepvisits/<visit-id>/prep.jsonhold-clinical-textvisits/**, the schedule read-only, the portal send-onlynote-draftvisits/<visit-id>/note-draft.mdunsettled-termvisits/<visit-id>/**, the chart read-onlynote-signvisits/<visit-id>/note-signed.mdawait-signaturevisits/<visit-id>/**, everything else read-onlyvisit-codevisits/<visit-id>/codes.jsonhold-unsupported-codevisits/<visit-id>/**, the signed note read-only, the payer rules read-onlyclaim-outclaims/<claim-id>.jsonawait-held-claim-decisionclaims/**, visits/** read-onlyfollow-throughvisits/<visit-id>/follow.jsonroute-clinical-replyvisits/**, the portal send-only, the scheduleweek-closerules/proposed/<rule-id>.json, reports/week-<date>.jsonaccept-rule-changerules/proposed/**, reports/**, the history read-onlyStack
Every system this agent reads or writes.