Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4 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 an Enterprise Healthcare Provider Built a Support Agent That Answers Patients from the Plan Document
Answers a patient's question in seconds at any hour, and never lets a question it could not evidence pass as one it could. Written for an org where the record is the point: every answer carries the plan document it came from, every handoff carries the whole conversation, and the only thing left for a person is the conversation the knowledge base genuinely cannot settle. Six stations, run as a loop: what station 6 learns changes what station 3 answers.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Message Arrival
Nobody is asked anything hereAgents watch every channel a patient can write on, verify who is writing before anything is disclosed, and join a new message to the conversation it belongs to. Nothing waits on a person here. A conversation only reaches somebody when a later station cannot settle it.
Question Reading
Nobody is asked anything hereAgents read each verified message, name what it asks for, pull out the claim numbers and dates it carries, and read how the patient is doing across the whole conversation. Nothing waits on a person here. Emergency wording leaves the bot for the nurse line the moment it appears.
Answer Building
Nobody is asked anything hereAgents search the knowledge base and the settled tickets, write a reply grounded in this patient's own plan, check it for anything that reads as medical advice, and send it. Nothing waits on a person here. A question the record cannot answer goes to station 4 with everything already attached.
Human Handoff
A person answers hereAgents assemble every conversation the bot could not settle into a card that already carries the transcript, what was tried, and how the patient is doing, then put it in front of the one team that can settle it. This is the station where a person works. Everything else exists to keep this queue short.
Ticket Closing
Nobody is asked anything hereAgents confirm with the patient that the answer settled it, write the full record of who said what and what answered it, and close the ticket. A ticket the patient disputes reopens on its own and goes back to the team. A person signs nothing here. The answer already happened at station 3 or station 4, and closing is the record of it.
Weekly Review
A person answers hereAgents group the week's handoffs by what actually caused them, draft the knowledge base article that would have answered the commonest one, and put the week's numbers and the drafts in front of a person to sign. A person signs every change to what the bot says. That is the signature that has to stay human, because the bot will repeat a signed article thousands of times.
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:
A patient writes that they have chest pain, and the bot files it as a billing question because a category matched. The answer arrives in an hour, and an hour is the one thing an emergency does not have.
Anyone who types a patient's name gets that patient's record read back to them. One unconfirmed sender is a privacy breach, and the bot handles thousands of senders a week.
The bot finds no passage that covers the question, so it writes the answer that sounds most likely. A made-up dosage or a made-up policy reads exactly like a real one, and the patient acts on it.
The patient asked two things, the bot answered one, and the ticket closed as resolved. Nobody ever sees the second question, and the patient learns that asking twice gets you half an answer.
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.
channel-watchqueue/arrivals/<ticket-id>.jsonhold-unverifiedqueue/**, patients/** read-onlyread-intentqueue/read/<ticket-id>.jsonemergency-wordingqueue/**, conversations/** read-onlybuild-answerqueue/answers/<ticket-id>.jsonsafety-holdqueue/**, kb/** read-onlytakeover-briefqueue/handoffs/<ticket-id>.jsonawait-human-takeoverqueue/handoffs/**, everything else read-onlyclose-tickettickets/closed/<ticket-id>.jsonreopen-disputedtickets/**, queue/answers/** read-onlyweekly-learnkb/drafts/<article>.json, reports/week-<n>.jsonsign-kb-changekb/**, reports/**, tickets/** read-onlyStack
Every system this agent reads or writes.