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 HR Services Company Built a Support Agent That Answers Employees from the Employer's Handbook
Answers every employee question the moment it arrives, on whichever channel it arrived, and never sends an answer it cannot point to a page for. Written for an HR services company where the question "how do I add my newborn?" comes in at 11pm on a Sunday and the person asking cannot wait for office hours. Every answer cites the employer's own handbook, every question the bot cannot settle reaches a specialist with the whole story already written down, and the questions the bot could not answer this week become the answers it gives next week. Six stations, run as a loop: what station 6 learns is what station 3 searches.
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 Arrival
Nobody is asked anything hereAgents watch every channel questions arrive on, merge the same person asking twice, and tie each question to an employee record before anything tries to answer it. Nothing waits on a person here. A question only reaches somebody when a later station cannot settle it.
Question Triage
Nobody is asked anything hereAgents sort each question into the kind of work it is, pull the record it will be answered against, and ask one short clarifying question when the message does not say enough. A person never sorts here. A question the agents cannot make sense of goes to station 4 with everything they tried.
Answer Delivery
Nobody is asked anything hereAgents search the employer's handbooks, policies and past settled tickets, write the answer with the page it came from, and stay with the question until it is settled or stuck. A person types no answer here. The questions the agents cannot settle move on with everything already tried attached.
Specialist Handoff
A person answers hereAgents turn each stuck question into a short brief that carries the transcript, what was already answered, and the two or three ways this kind of question usually settles. This is the station where a person works. Everything before it exists to keep this queue short.
Ticket Closing
Nobody is asked anything hereAgents write every conversation into the ticketing system with who answered what and from where, ask the one-question survey, and reopen any close the survey proves wrong. A person signs nothing here. The answer already happened at station 3 or station 4, and the record writes itself as it happens.
Service Review
A person answers hereAgents group the week's handoffs by what caused them, draft a new answer for every gap the specialists had to fill twice, and put the week's numbers in front of the person who signs it. A person signs the week. New answers do not go live without that signature, because next week the bot gives them to every employee who asks.
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 agent matches an employee on name alone and pulls up the wrong record. It then answers a stranger's benefits question with someone else's plan details.
The agent finds no passage from this employer and answers from what it knows about benefits in general. The answer sounds right and describes a plan the employee does not have.
The agent hands a stuck question to a shared queue with no owner. Nobody's name is on it, so nobody picks it up, and the employee asks again in a week, angrier.
The agent closes a ticket whose last message is still a question. The employee thinks they were ignored, and the metrics count it as resolved.
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/open/<conv-id>.jsonhold-unmatched-senderqueue/**, rosters/** read-onlytriage-questionqueue/triaged/<conv-id>.jsonhold-unclear-questionqueue/**, tickets/** read-onlyanswer-questionqueue/answered/<conv-id>.jsonno-passage-foundqueue/**, knowledge/** read-onlyhandoff-briefqueue/handoff/<conv-id>.jsonawait-specialist-answerqueue/handoff/**, everything else read-onlyclose-tickettickets/closed/<conv-id>.jsonreopen-on-bad-surveytickets/**, queue/answered/** read-onlyreview-and-learnknowledge/proposed/<answer-id>.json, reports/week-<n>.jsonsign-weekly-reviewknowledge/**, reports/**, tickets/** read-onlyStack
Every system this agent reads or writes.