Five-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3, 4 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 an Enterprise Technology Company Built Defect Grading with a Human Verdict on Every Held Unit
Every unit that comes down the line, from the moment it passes the cameras to the moment the shift is counted. The agents photograph every unit, find and grade every defect, clear the clean ones and pull the bad ones, and hand a person only the units the catalogue cannot call, the alerts only a person can answer, and the rule changes only a person may accept. Five stations, run as a loop: the calls a person makes at station 3 and the thresholds accepted at station 5 are what station 2 scans against tomorrow.
What Was Broken
What We Built
Five stations and 10 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Frame Capture
Nobody is asked anything hereAgents photograph every unit off the four line cameras, match each frame to the unit's serial, retake anything blurred or glared before the unit leaves the camera, and hold a unit whose serial nobody could read. Nothing waits on a person here. A held unit sits at the end of the belt with its frames kept, never passed on as somebody else's.
Defect Detection
Nobody is asked anything hereAgents scan every frame of every unit, locate each finding down to the component it sits on, grade it cosmetic or functional, and flag a finding the catalogue has no example of rather than forcing the nearest label. Nothing waits on a person here. A finding that matches nothing keeps its frames attached and travels to the verdict station as an open question, not a guess.
Unit Verdict
A person answers hereAgents clear the units whose frames came back clean, pull the ones carrying a functional finding, and queue the borderline ones for a person, each beside the catalogue examples that decide it. This is where a person works. Everything else exists so that five units, not nine hundred, reach this screen.
Line Watch
A person answers hereAgents trace every pulled unit back through the machines it passed, watch the pull rate against the last thirty shifts, and raise an alert when the same defect arrives in a run. The alert stands until a person answers it. This is where a person decides. The agents can find the run and page the floor; stopping a press is not theirs to call.
Shift Close
A person answers hereAgents count the shift's units, pulls, holds and alerts, trace what the day's run cost, and propose the threshold changes that would have caught it earlier, with what each change would wrongly pull stated beside what it catches. A person accepts or rejects each one. This is where a person decides. How the scan grades a finding and when an alert pages the floor are rules, and an agent does not change its own rules 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:
When no camera reads the serial, the frames enter the pipeline anyway. A defect gets recorded against a unit nobody can find, and the bad unit ships under a clean record.
The scanner meets a mark it has never seen and files it under the nearest known defect. The new failure hides inside an old category, and nobody goes looking for it.
A held unit sits past its window and the line lets it go. The units that most needed a person are exactly the ones that ship without one.
The alert fires, nobody answers, and after a while the system marks it resolved. The floor learns that alerts go away on their own, so nobody answers the next one either.
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.
capture-and-checkvision/frames/<serial>/<cam>.jsonhold-unreadable-serialvision/frames/**, vision/held/**, the cameras read-onlyscan-and-gradevision/findings/<serial>.jsonno-catalogue-matchvision/findings/**, vision/frames/** read-onlycall-the-unitvision/verdicts/<serial>.jsonawait-your-callvision/verdicts/**, vision/findings/** read-onlywatch-the-linevision/alerts/<alert-id>.jsonalert-awaiting-answervision/alerts/**, vision/verdicts/** read-onlyclose-the-shiftvision/reports/shift-<date>.json, vision/thresholds/proposed/<id>.jsonaccept-threshold-changevision/reports/**, vision/thresholds/proposed/**, the history read-onlyStack
Every system this agent reads or writes.