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 Members from Their Own Records
Answers a member's question the moment it arrives, on whichever channel it arrives, and never sends a sentence it cannot trace to an article or to the member's own record. Written for an org where the queue was the worst part of the product: the repetitive questions get answered at once and around the clock, the hard ones reach a person carrying the whole story, and the one thing a person signs each week is what the bot is allowed to say. Six stations, run as a loop: the article station 6 approves is the article station 3 answers from.
What Was Broken
What We Built
Six stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Inquiry Intake
Nobody is asked anything hereAgents watch every channel members write on, match each sender to a member record, and stitch a member's email and chat about the same matter into one conversation. Nothing waits on a person here. An unverified sender reaches somebody only through station 4.
Intent Reading
Nobody is asked anything hereAgents read each message for what the member wants, pull out the claim numbers and dates the answer will need, and screen every turn for anything a bot must never answer. Nothing waits on a person here. What cannot be read goes to station 4 with its transcript, and anything clinical goes straight to the nurse line.
Answer Writing
Nobody is asked anything hereAgents search the article library, past settled tickets and the member's own claim and appointment records, then write a reply with a source behind every sentence and send it on the channel the member asked on. Nothing waits on a person here. A question with no answer on file is handed over, never improvised.
Human Handoff
A person answers hereAgents pull a conversation out of the bot's hands the moment it is stuck or the member is upset, and hand it to the right person as one line they can act on, with the transcript and the history already attached. 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 member that the answer settled it, write the full transcript, cause and outcome onto the ticket, and reopen the ticket when the member comes back on any channel. Nothing waits on a person here. A reopened ticket goes back through the loop on its own.
Knowledge Review
A person answers hereAgents group the week's handoffs by what the bot was missing, draft the article that would have answered the commonest gap, and trace complaints back to the answer that caused them. A person signs the week. The bot says only what somebody approved, and an answer proven wrong comes down before anything else goes up.
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 bot matches a sender to a member record on a name that looks close. It reads one person's health history to answer another person's question, and nobody knows until the member does.
A member describes chest pain, and an article about chest pain matches it. The bot sends the article, and a machine is now the first responder to a medical complaint.
The bot writes what sounds right instead of what a passage says. A member acts on a coverage claim that no policy backs, and the company owns that sentence.
The bot keeps answering a member who has already stopped trusting it. Every correct reply after that point makes the conversation worse, because the member wanted a person three messages ago.
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/inbound/<conv-id>.jsonidentity-unverifiedqueue/**, members/** read-onlyread-intentqueue/understood/<conv-id>.jsonintent-unclearqueue/**, history/** read-onlyanswer-from-kbqueue/answered/<conv-id>.jsonno-source-foundqueue/**, kb/** read-onlyhandoff-briefqueue/handoffs/<conv-id>.jsonawait-human-answerqueue/handoffs/**, everything else read-onlyclose-tickettickets/closed/<conv-id>.jsonmember-still-waitingtickets/**, queue/** read-onlyweekly-reviewkb/proposed/<article-id>.json, reports/week-<n>.jsonsign-kb-changekb/**, reports/**, tickets/** read-onlyStack
Every system this agent reads or writes.