Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3 and 6 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 Technology Company Built a Request Loop That Keeps a Record Anyone Can Ask
Every request an employee raises, from the moment it is raised to the moment it shows up in the Friday numbers. The agents name each request, gather what it needs from the systems that hold it, carry it out once it is cleared, and keep a record anyone can ask. A person signs wherever money or access moves, and decides how the process itself changes. Six stations, run as a loop: what Friday's report finds about a snag becomes the rule intake runs on Monday.
What Was Broken
What We Built
Six stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Request Intake
Nobody is asked anything hereAgents collect every request off every route it can arrive by, attach it to the person who raised it, match it to a request type the company already runs, and hold anything that names no ask. Nothing waits on a person here. A held request is never dropped and never guessed at.
Fact Gathering
Nobody is asked anything hereAgents pull the facts each request type needs from the systems that hold them, name where every fact came from, show the ones that disagree, and say plainly when a fact exists nowhere. Nothing waits on a person here. A gap in a package is written down, never papered over.
Sign-Off
A person answers hereAgents clear the requests policy already answers, and put the rest in front of a person with the ask in one line, the cost beside it, and every conflicting or missing fact marked above the button. This is where a person works. Everything else exists so that three requests, not thirty, reach this screen.
Carrying Out
Nobody is asked anything hereAgents run each cleared request's steps across the systems it touches, read every change back before calling it done, tell the raiser where their request stands, and stop any run a system did not take. Nothing waits on a person here. A change that does not read back did not happen, whatever the connector said.
The Record
Nobody is asked anything hereAgents close each finished request onto the company record with its whole trail, answer questions out of that record with the line each claim came from, and say not found where the record does not answer. Nothing waits on a person here. An answer the record cannot support is refused, not softened.
Reporting
A person answers hereAgents build the week's report off the record alone, mark the number that broke its own pattern, and propose the process changes that would have stopped the week's repeated snags, with what each change would slow written beside what it saves. A person accepts or rejects each change. How the company runs a request is not something an agent changes quietly.
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 package looks complete. One field was filled with a guess, and the approver signs a number nobody actually pulled.
Two systems disagree on the amount. The payment goes out anyway, and the cleanup starts after the money is gone.
The agent sends the change and moves on without reading it back. The target system dropped it, and the request closes on work that did not happen.
Someone asks what happened on a request the record does not hold. The search hands back a similar one, and the wrong answer carries a real citation.
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.
request-intakerequests/open/<req-id>.jsonhold-unmatched-requestrequests/open/**, requests/held/**, the channels read-onlyfact-gatheringrequests/assembled/<req-id>.jsonfield-no-system-holdsrequests/assembled/**, the HR system, the finance system, the CRM and the drive read-onlysign-off-deskrequests/cleared/<req-id>.jsonawait-your-signoffrequests/cleared/**, requests/assembled/** read-onlycarry-outrequests/done/<req-id>.jsonwrite-did-not-read-backrequests/done/**, the finance system, the directory, the ordering tool and the CRM through their connectorsrecord-keepingrecord/closed/<req-id>.json, answers/<q-id>.jsonanswer-not-in-the-recordrecord/**, answers/**, everything else read-onlyreport-and-tunereports/week-<n>.json, process/proposed/<change-id>.jsonaccept-process-changereports/**, process/proposed/**, the record read-onlyStack
Every system this agent reads or writes.