Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4, 5 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 Healthcare Provider Built Tray Inspection with Human Approval Before a Batch Ships
Every tray the line makes, filmed, read and judged, from the moment it passes the cameras to the moment its batch ships. The agents inspect every tray instead of the sample a tired pair of eyes used to pull, hold what they cannot call, and reach a person only where the call, the signature or the threshold is theirs. Six stations, run as a loop: the human calls made at the hold bench today are the thresholds the reading station inspects against tomorrow.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Line Capture
Nobody is asked anything hereAgents collect every frame off every camera at line speed, tie each one to the tray and batch under it, retake what blurs, and mark a tray uninspected the moment its coverage fails. Nothing waits on a person here. A blind camera pages maintenance and marks its trays uninspected, never inspected-clean.
Tray Inspection
Nobody is asked anything hereAgents find every component on every tray, name each defect with the frames behind it, and close each tray to one verdict: pass, defect, or unsure. Nothing waits on a person here. A tray the angles cannot agree on is held unsure, not guessed through.
Line Disposition
Nobody is asked anything hereAgents act on each verdict in the second it lands: pass routes on, defect diverts to rework with its record, unsure goes to the hold bench, and a defect cluster pages the line lead. The line-stop call is the one thing an agent will not make. The page goes to the line lead, and the trays hold while it waits.
Hold Review
A person answers hereAgents build the case for each held tray, recommend a call where the past holds like it point one way, and present the tray bare where the frames support no recommendation. This is where a person works the trays. Each held tray gets a human call, made with the frames on screen and recorded with them.
Batch Release
A person answers hereAgents close every tray's verdict, hold and human call into one batch record, name any gap in it, and draft the release summary a signer reads in one screen. This is where a batch waits. Nothing ships until a person signs the record, and a record with a gap blocks its own release.
Shift Close
A person answers hereAgents reconcile every tray against where it ended, plot where the defects sit and how the hold rate is moving, propose the threshold change the day's human calls point to, and queue the calls as retraining examples. A person accepts or rejects each change. What the model holds tomorrow 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:
A camera fails and the trays behind it keep moving. Nothing marks them, so the record says inspected when nobody looked.
Glare hides half the tray and the model scores what it can see. The pass looks like every other pass, so nobody knows an inspection did not happen.
A reviewer clears a hold from the queue screen without opening the frames. The log records a human decision that never happened.
The record closes while trays are still uninspected. The signature goes on a batch nobody can fully account for, and the gap surfaces in an audit, not on the line.
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.
frame-intakevision/frames/<tray-id>/camera-blindvision/frames/**, the cameras read-onlyinspect-trayvision/verdicts/<tray-id>.jsontray-unreadablevision/verdicts/**, vision/frames/** read-onlyroute-and-alertvision/routed/<tray-id>.json, alerts/<alert-id>.jsonline-stop-callvision/routed/**, alerts/**, the divertershold-benchholds/<tray-id>.jsonawait-your-callholds/**, vision/** read-onlybatch-releasebatches/<batch-id>/record.jsonawait-your-signaturebatches/**, holds/** read-only, vision/** read-onlyshift-closereports/shift-<date>.json, thresholds/proposed/<change-id>.jsonaccept-threshold-changereports/**, thresholds/proposed/**, the history read-onlyStack
Every system this agent reads or writes.