Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3, 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 Manufacturer Built AI Vision Inspection with the Inspector Still Making the Call
Photographs every part, calls the ones that look wrong, and hands an inspector a decision instead of a shift of scanning. Written for a precision manufacturer making machined components for aerospace and medical device customers, where the constraint was never how hard anyone was trying. A Sandia study of 82 inspectors on precision parts measured an 85% hit rate on defective items and a 35% false-alarm rate on good ones, and detection starts falling inside the first 30 minutes on task. Six stations, run as a loop: what station 6 learns from the overrides and the escapes is what station 2 calls on.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Image Capturing
Nobody is asked anything hereSubagents photograph each part from every fixed angle, check the exposure against the enclosure's own reference, and stop a station whose lighting or fixture has drifted. Nothing waits on a person here. A part reaches an inspector only when station 3 asks them to look.
Defect Scoring
Nobody is asked anything hereSubagents score every part against the defect classes this family has actually shown, and lean toward flagging, because a missed defect costs more than a second look. A person sees a flag with the drawing beside it, not a verdict.
Flag Reviewing
A person answers hereSubagents turn each flag into a decision an inspector can make in one look, with the frame, the defect class and the drawing tolerance side by side. An inspector decides. This is the station the whole build exists to make small enough to do well.
Lot Deciding
A person answers hereSubagents group every called part by lot, shift and tool, count what it is worth, and draft the disposition the evidence supports. A person decides scrap, rework or use-as-is. Nothing ships or scraps on an arithmetic.
Sample Checking
A person answers hereSubagents draw the audit sample from the parts the model passed, check each one against the classes the model calls, and count what got past. An inspector signs the audit. The point of this station is to find out where the model is wrong, not to confirm it is right.
Model Retraining
A person answers hereSubagents read every override and every escape back against the call that produced it, find the defect class the model has never seen, and propose the retraining. A person accepts or rejects each proposed change. Nothing about what the model calls changes on its own.
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 the station keeps scoring while its lighting reference drifts, every score after the drift is a score of the light, not the part. The numbers stay green while the calls go quietly wrong.
A model allowed to move its own threshold moves it toward whatever makes today's queue quiet. Your acceptance standard changes and nobody signed the change.
If the system releases a lot while a decision on it is still open, the parts leave the building before the answer arrives. A held pallet is cheap. A recall is not.
An override with no reason written on it is a data point with no meaning. A model that trains on it learns the inspector's habit instead of the standard, and the habit spreads to every future call.
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-grablines/<line-id>/frames/<part-id>.jsonlighting-out-of-referencelines/**, drawings/** read-onlydefect-scorelines/<line-id>/scored/<part-id>.jsonfamily-not-trainedlines/**, drawings/** read-onlyflag-presentlines/<line-id>/reviewed/<part-id>.jsonawait-inspector-calllines/<line-id>/reviewed/**, everything else read-onlylot-brieflots/<lot-id>/disposition.jsonawait-human-dispositionlots/**, everything else read-onlyaudit-drawlots/<lot-id>/audit.jsonescape-foundlots/**, lines/** read-onlymodel-readanalytics/model/<period>.jsonpropose-retraininganalytics/**, lines/** read-onlyStack
Every system this agent reads or writes.