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 a Growth-Stage Educational Institution Built a Support Agent That Cites the Ledger Behind Every Sentence
Answers every student and parent inquiry the record can actually support, on whatever channel it arrives, at any hour, and hands a person only the questions the record genuinely cannot settle. Written for an institution where the reply is the institution speaking: every sentence sent carries the article or ledger entry behind it, every hold carries a reason, and the support team's day is the short queue of things only they can answer. Six stations, run as a loop: what station 6 signs is what station 3 answers with next week.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Inquiry Intake
Nobody is asked anything hereAgents watch every channel a question can arrive on, match each message to a student and to any conversation already open, and drop what names no student at all. Nothing waits on a person here. A sender nobody can place reaches somebody later, with everything known about it attached.
Question Triage
Nobody is asked anything hereAgents read each conversation, settle what is being asked and how soon it matters, and ask one follow-up question when the message could mean two things. A person is not asked anything here. A question still unclear after one follow-up goes to a person with the exchange attached.
Answer Drafting
Nobody is asked anything hereAgents search the knowledge base, the closed-ticket archive and the student's own record, compose a reply with a source behind every sentence, and hold instead of guessing when no source exists. A person reads none of these before they go out. What has no source behind it never goes out at all.
Escalation Desk
A person answers hereAgents assemble everything the pipeline held, the unplaced senders, the unclear questions, the answers no source supports, into one line a person can decide from, routed to the person who owns that kind of question. This is the station where the support team works. Every other station exists to keep this queue short.
Ticket Closing
Nobody is asked anything hereAgents send the reply on the channel the student is actually on, watch what comes back, and close the ticket only when nothing in it is still open. A person signs nothing here. The answer was signed at station 4, or the sources held and there was nothing to sign.
Knowledge Review
A person answers hereAgents group the month's handoffs by what actually caused them, draft the knowledge base articles that would have answered the recurring ones, and count the month off the tickets themselves. A person signs every draft before the bot answers with it. A person works here, because what enters the knowledge base is what the bot says in the college's name from then on.
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 student mentions self harm, then asks a routine question about a deadline. Without the refusal to lower a distress flag, the routine tone wins, the flag drops, and no person ever reads the message.
Nothing in the sources answers the question, but the search still returns the closest match. The student gets a confident reply about a different policy than the one they asked about.
A shared surname or a reused family email looks like the same person, and the two threads merge. One student then reads answers written for the other, including things they were never meant to see.
A conversation goes quiet and an age rule closes it. The second question in that thread never gets an answer, and once the ticket is closed, nobody goes looking for it.
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-intakequeue/conversations/<conv-id>.jsonidentity-unresolvedqueue/**, directory/** read-onlyintent-triagequeue/triaged/<conv-id>.jsonsafeguarding-detectedqueue/**, calendar/** read-onlyanswer-draftqueue/replies/<conv-id>.jsonno-source-foundqueue/**, kb/** read-only, tickets/** read-onlyhandoff-briefqueue/escalations/<conv-id>.jsonawait-human-answerqueue/escalations/**, everything else read-onlyclose-watchtickets/closed/<conv-id>.jsondelivery-failedtickets/**, queue/replies/** read-onlygap-reviewkb/drafts/<article>.md, reports/week-<n>.jsonapprove-kb-draftkb/**, reports/**, tickets/** read-onlyStack
Every system this agent reads or writes.