Five-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3 and 5 need a person. Step 3 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 an Inquiries Agent That Answers Every Question It Can Prove
Answers every student question it can prove, on the channel it arrived on, at any hour, and hands a person the ones it cannot. Written for an org where the institution's word matters: every answer carries the page or the staff reply it came from, a claim with no source never sends, and the only thing left for a person is the short queue the knowledge base genuinely cannot settle. Five stations, run as a loop: the articles a person approves at station 5 are the answers station 2 sends next week.
What Was Broken
What We Built
Five stations and 11 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Inquiry Intake
Nobody is asked anything hereAgents watch every channel questions arrive on, tie each message to the person who sent it, and label what it is asking before anything tries to answer it. Nothing waits on a person here. A student reaches somebody only when a later station cannot settle their question.
Question Answering
Nobody is asked anything hereAgents search the knowledge base and past staff replies, draft the answer with a source on every claim, send it back on the channel it arrived on, and hold the conversation across turns. A person sees only the conversations the bot holds back, and station 3 carries those to them.
Staff Handoff
A person answers hereAgents route each held conversation to the office that can actually answer it, with the whole thread, what the bot already said, and a drafted reply to start from. This is the station where a person works. Everything else exists to keep this queue short.
Ticket Closing
Nobody is asked anything hereAgents confirm the student got an answer, send the one-question survey, write the ticket record with its sources, and reopen a ticket when the same student comes back on the same question. Nobody signs a closing. The answer was either a staff reply from station 3, or a bot reply that cited its source and drew no comeback.
Weekly Learning
A person answers hereAgents group the week's handoffs by what the bot was missing, draft the knowledge base articles that would have settled the recurring ones from the staff's own replies, and assemble the week's counts. A person approves every article before the bot is allowed to repeat it, because the bot will say it all year.
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 asks about a deadline the knowledge base does not cover. The bot sends the closest paragraph it found, and the student acts on a date that was written for a different program.
The bot could not help, so it drops the case into a shared queue with nobody named on it. Every staff member assumes someone else owns it, and the student who most needed a person gets nobody.
A conversation stalls, the student stops replying, and the ticket closes as resolved. The dashboard shows the question was handled, and nobody ever finds out it was not.
The week report says 94 percent of inquiries were resolved, and nobody can trace that number back to a ticket. You make staffing decisions on a figure the system invented, and you find out the season it mattered.
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.
inquiry-intakequeue/inbound/<conv-id>.jsonhold-unreadablequeue/**, records/** read-onlyanswer-draftconversations/<conv-id>/reply.jsonclaim-without-sourceconversations/**, kb/** read-onlyhandoff-briefqueue/handoffs/<conv-id>.jsonawait-staff-replyqueue/handoffs/**, everything else read-onlyticket-closetickets/closed/<ticket-id>.jsonclose-without-answertickets/**, conversations/** read-onlyweekly-learningkb/drafts/<article-id>.md, reports/week-<n>.jsonapprove-articlekb/**, reports/**, tickets/** read-onlyStack
Every system this agent reads or writes.