Seven-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 7 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 Technology Company Built Unit Inspection That Learns from Technician Corrections
Every unit that crosses the line, from the frame that catches it to the verdict that clears it. The agents photograph every unit instead of one in ten, read every frame, sort each unit against its spec, and hand a person only the calls that are theirs: the held unit on the bench, the open alert on the line, the change to what counts as a defect. Seven stations, run as a loop: what the bench overturns today changes what the cameras call tomorrow.
What Was Broken
What We Built
Seven stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Line Capture
Nobody is asked anything hereAgents photograph every unit crossing the line at line speed, stamp each frame to its unit and lane, retake the bad shots, and send a unit no camera saw clearly to the hold lane whole. Nothing waits on a person here. A unit is never judged off a blurred frame, and never skipped.
Defect Inspection
Nobody is asked anything hereAgents read all four angles of every unit, find each scratch, chip and missing part, place it on the unit face by face, and pass a mark they cannot call to the hold lane rather than guess. Nothing waits on a person here. Every unit is read, not one in ten, and a mark rounded neither up to a defect nor down to clean.
Line Sorting
Nobody is asked anything hereAgents call each placed defect against the spec for that product, pass what the spec passes, fire the divert arm on what it fails, and hold anything with no spec behind it rather than guess either way. Nothing waits on a person here. A unit with no spec loaded is held whole, not passed on a guess and not failed on one.
Hold Bench
A person answers hereAgents build a docket for every held unit, the frames, the spec line and the reason it stopped, recommend a verdict with the evidence above it, and record what the technician decides, including every time the technician overturns them. This is where a person works. Everything upstream exists so that a tray of dockets, not a shift of units, reaches this bench.
Line Watch
A person answers hereAgents track every defect across units, lane by lane and hour by hour, group repeats onto the machine station that touches that face, raise an alert when a repeat crosses the rule's line, and keep paging until a person takes it. A person answers the alert and closes it. An agent can name the polisher head; only a person can swap it and say the line is right again.
Shift Record
Nobody is asked anything hereAgents count the shift, units through, passed, held, scrapped and reworked, place every defect on the line drawing by face, lane and hour, keep one trail per unit, and hold the board back the moment two counts disagree. Nothing waits on a person here. The board is on the wall screen by shift end, and every number on it opens to the units behind it.
Recalibration
A person answers hereAgents turn the bench's overturned holds into proposed threshold changes, draft a spec for the variant that arrived without one, pair every uncalled mark with its bench verdict as a training example, and hold all of it until a person signs. A person decides every change. What counts as a defect on this line is not something an agent adjusts 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 unit gets no clear picture and the line judges it off the blurry one anyway. Bad units pass on evidence nobody could have read.
The reader cannot tell a mark from a shadow and calls it anyway. Half the guesses scrap good units and the other half ship bad ones.
A held unit gets scrapped or released on the recommendation alone. Nobody made the call, so nobody can defend it when the customer asks.
The shift counts disagree with the unit trails and the board posts anyway. The number everyone plans against is the one nobody can trace.
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-intakeline/frames/<unit-id>/no-usable-frameline/frames/**, line/hold/**, the camera feeds read-onlydefect-readline/read/<unit-id>.jsoncannot-tellline/read/**, line/frames/** read-onlyspec-sortline/sorted/<unit-id>.jsondivert-missedline/sorted/**, line/read/** read-only, the divert armdocket-queuehold/dockets/<unit-id>.mdawait-your-verdicthold/**, everything else read-onlyrun-watchalerts/<alert-id>.jsonalert-raisedalerts/**, line/sorted/** read-onlyshift-closereports/shift-<date>.jsoncounts-disagreereports/**, the day's records read-onlyrecalibraterules/proposed/<change-id>.jsonaccept-changerules/proposed/**, hold/** read-only, reports/** read-onlyStack
Every system this agent reads or writes.