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 Insurance Company Built an Inquiries Agent That Answers from the Policy Clause
Answers the customer questions that used to queue for a person, at any hour, and never sends a sentence it cannot point to a policy clause or a record for. Written for an org where the promise is the point: every reply carries its sources, every handoff arrives briefed, and the only thing left for your team is the customer no document can settle. Six stations, run as a loop: what station 6 learns from this month's handoffs is what station 3 answers with next month.
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 Intake
Nobody is asked anything hereAgents watch every channel a customer can write on, tie each sender to their policy record, and fold a customer writing on two channels into one conversation. Nothing waits on a person here. A sender nobody can match reaches somebody through station 4, with everything that is known attached.
Question Triage
Nobody is asked anything hereAgents name what each message wants, carry the earlier turns so a follow-up reads as one conversation, and send a cancellation or a run of falling patience straight to a person. A person is not asked anything here. What needs one is already on its way to station 4 before anybody could have spotted it.
Answer Drafting
Nobody is asked anything hereAgents pull the clause and the customer's own numbers, draft the reply with a source on every sentence, and check the draft against the record before it sends. Nobody reviews the replies that pass. A person sees only the questions no document answers, and those arrive at station 4 already declined.
Person Handoff
A person answers hereAgents assemble everything the bot declined or triage routed into a short queue, each item one line, with the whole conversation, the policy record and two or three drafted replies attached. This is where your team works. Every other station exists to keep this queue short and this brief complete.
Conversation Closing
Nobody is asked anything hereAgents confirm the answer landed, reopen the conversations where it did not, and file a ticket record that carries every channel, every turn, the answer sent and the sources behind it. Nobody signs a closing. A customer who writes back unhappy goes back through triage by themselves, with their history attached.
Service Review
A person answers hereAgents group the month's handoffs by cause, draft new knowledge base articles from the replies the reps actually sent, and sample the sent replies against the sources they cited. A person approves every new article before the bot answers with it, and signs the review. Those are the two acts that stay human, because the bot must not teach itself an answer nobody checked.
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:
When the bot guesses which account a stranger belongs to, it answers with someone else's order history, billing terms, or contract. One good guess in ten is still a data leak nine times.
When no passage covers the question, the bot can still write a fluent reply from nearby material. The customer acts on it, and nobody in the company ever said it or can defend it.
A handoff dropped in a shared queue with no name on it waits for whoever feels responsible, which is nobody. The customer's hardest question is the one that sits longest.
If the bot can close a thread whose last message is an unanswered question, your resolution numbers rise while your customers stop asking. The dashboard says solved, the customer says ignored.
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/open/<conv-id>.jsonhold-unmatched-senderqueue/**, customers/** read-onlytriage-intentqueue/triaged/<conv-id>.jsonroute-on-tonequeue/**, policies/** read-onlyanswer-draftqueue/replies/<conv-id>.jsonno-source-for-answerqueue/**, kb/** read-only, policies/** read-onlyhandoff-briefqueue/handoff/<conv-id>.jsonawait-rep-answerqueue/handoff/**, everything else read-onlyclose-recordtickets/closed/<conv-id>.jsonreopen-on-replytickets/**, queue/** read-onlyservice-reviewkb/proposed/<article-id>.md, reports/review-<period>.jsonapprove-articlekb/**, reports/**, tickets/** read-onlyStack
Every system this agent reads or writes.