Manufacturing AI agent · Quality
Quality-Inspection Copilot AI Agent
Present candidate defects with the evidence, compare a part against the control plan and the current drawing revision, assemble the inspection record and route suspect parts — a qualified person dispositions.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Disposition
- Never the agent
- After launch
- Agent Care
What this agent does
Presents candidates and assembles the record
In
- 01
Take station images, gauge and CMM readings and the part identity from the cell and the work order.
- 02
Load the control plan, the drawing characteristics and the revision in force for that part and operation.
Reason
- 03
Compare each reading against the characteristic, its tolerance and the sampling frequency the plan names.
- 04
Present candidate defects beside the image, the station, the fixture and the reference examples matched.
- 05
Track measured characteristics against the control limits and mark runs, trends and shifts as they build.
Decide
- 06
Route a suspect part to quarantine and to a named inspector rather than judging it conforming.
- 07
Hold anything read against a superseded revision or a gauge with no current capability study.
Out
- 08
Assemble the inspection record — part, lot or serial, station, gauge, revision, reading and image.
- 09
Present it for signature. Release, scrap, rework and concession stay with the person the procedure names.
Product statement
The agent presents candidates with the evidence and assembles the record. Disposition — release, scrap, rework or concession — stays with a qualified person named in the procedure.
Example workflow
One part, end to end
- 1Part reaches the stationA vision cell, a gauge, a CMM or a manual inspection step named on the work order
- 2Part and plan identifiedLot or serial, work order, operation, station and fixture, then the control plan and the revision in force
- 3Characteristics read and matchedReadings against the named characteristics, candidate defects against the labelled reference examples
- 4Checks appliedRevision currency, gauge calibration and capability, sampling frequency and the control limits
No human action required
Stages 1 to 4 run without a person in the loop — reading, matching and the plan checks all finish before anyone is asked to look at a part. Anything outside tolerance ends that stretch on the spot.
Filed as an observation; a person signs it off.
Quarantined and put in front of an inspector.
Inspector or quality engineer
Looks at the part, the image and the reading against the control plan, then dispositions it under the procedure.
Disposition · Correct · Send to MRB
Dispositioned — handed back
- 6Record assembled and filedWritten to the quality system only where access and procedure allow; the disposition field stays empty
- 7Outcome evaluatedEscapes and false rejects against a re-inspected sample, per-defect-class recall and inspector agreement
Corrections
What the inspector overturns is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Dispositioning a non-conforming part.
- Releasing product or certifying conformance.
- Scrapping, reworking or accepting under concession.
- Closing a non-conformance or corrective action.
Automation boundary
Agent acts unaided
- ✓Present candidate defects beside the image and the reference examples matched.
- ✓Compare readings against the control plan and the revision in force.
- ✓Flag drift against control limits, and gauges with no current capability study.
- ✓Quarantine-route suspect parts and assemble the record for a person to sign.
The agent writes the inspection record and the routing flag. The disposition and the release are not its to write.
- Signing the inspection or batch record.
- Changing the control plan or an inspection method.
- Widening a tolerance or an accept threshold.
- Setting the containment boundary after an escape.
Example output
One part at one station, annotated
Everything the agent presents is attached to the exact part, station and drawing revision it was read from.
Inspection output · single part at one station
Illustrative example- Part and operation
- Machined housing, op 40
- Station
- Vision cell 2, fixture B
- Drawing revision
- Rev C, in force
- Presented as
- Candidate porosity, near edge
- Confidence
- 87%
- Disposition
- Not made by the agent
As received
The part, the station and fixture it ran on, and the revision in force at the time — taken from the work order, not inferred.
Evidence used
- Rev-C characteristic 14
- Three reference defects
- Prior runs, same fixture
Why it was presented
The feature is in tolerance, but the last runs sit on one side of the mean.
What the score decides
Confidence decides how hard the inspector looks, not whether a part ships.
Action
DispositionCorrectSend to MRBValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every part inspected
Vision cell, gauge, CMM or a manual step
Work to your own control plan
Use the plant's control plan, drawing characteristics, gauge records and the reference defects your inspectors labelled.
Give the inspector evidence, not a verdict
The image, the reading, the characteristic and the reference examples arrive together, so the decision starts from what was seen.
Send the doubtful parts, not all of them
Suspect parts and weak candidates reach an inspector while the rest carry a complete record, so attention goes where the doubt is.
Build an evidence trail
Retain the part, the station and fixture, the revision, the gauge, the image, the reading, the confidence and the disposition — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Inspection stations
- Cognex
- Keyence
- Basler
- Halcon
- CMM and gauge output
Quality & MES
- Siemens Opcenter
- SAP QM
- InfinityQS
- Minitab
- Plant MES
Engineering & drawings
- Teamcenter
- Windchill
- PLM revisions
- GD&T characteristics
Agent
Quality-inspection copilot
- Compares to the plan
- Presents candidates
- Routes suspect parts
Gauge & calibration records
- Gauge registers
- calibration dates
- MSA and gauge R&R
- Reference defects
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the disposition
Each control wraps the one inside it. A candidate clears every layer before an inspector sees it, and the disposition sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Keep the comparison current and meaningful
- L3
- Decides how hard the inspector looks
- L4 – L5
- Keep disposition with a person, trail intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the whole inspection — not only the candidates it presents.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the candidate the inspector opens
- E1Final-output evaluationDid what the agent presented match what a re-inspection found?
- E2Step-level evaluationDid it read the right characteristic, tolerance and drawing revision?
- E3Tool evaluationDoes the record tie back to the exact part, station, fixture and gauge?
- E4Per-defect-class recallHow does recall hold for the classes with only a handful of examples?
- E5Slice evaluationHow does performance change across SKUs, shifts and changeovers?
- E6Business outcomeWhat escaped, what was falsely rejected, and how much was re-inspected?
Floor — the disposition a qualified person makes
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. None of them dispositions a part — the inspector's look, the control plan and the plant's non-conformance process are the controls that stop them. If one gets through and product has shipped, the part, lot, station and revision held in the record are what the quarantine, the customer notification and the containment sort are built from.
Agent lifecycle
Direction of processing →
01 · Retrieval
2 modes
- QI-01
Superseded drawing revision
Compared against a revision the drawing has moved past.
- QI-02
Angle the station cannot see
A defect visible only from a view the cell never captures.
Stage gathers
The image, the reading, the plan and the revision
02 · Reasoning
1 mode
- QI-03
Rare defect class missed
A class with few labelled examples is not presented at all.
Stage compares
The candidate, the tolerance and the control limits
03 · Record / file
1 mode
- QI-04
Record ties to no part
The record cannot be pinned to the exact part and station.
Stage files
The inspection record, written to the quality system
04 · Output
2 modes
- QI-05
Reading without gauge capability
A measurement reported as if the gauge had been studied.
- QI-06
False rejects train waving through
So much is flagged that inspectors stop looking properly.
Stage presents
The candidate and the evidence the inspector reads
05 · Change / Version
1 mode
- QI-07
Drift after a rebuild
New lighting, fixture or material moves the station quietly.
Stage tracks
Model, threshold, control-plan and station changes
- Sev-1 · a defect can reach the customer
- Sev-2 · the record cannot be relied on
- Sev-3 · false rejects rise, the flag stops being read
Affected slices
Recall is not the same on every defect class
Two parts on the same line are not the same problem. A mature SKU with thousands of labelled examples reads cleanly; a rare class on a first lot after changeover does not. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Rare defect classes | 6.6% | 3.6× | Review | |
| First-off after changeover | 5.0% | 2.7× | Review | |
| New SKU, first lot | 3.5% | 1.9× | Watch | |
| Mature SKU, settled line | 1.1% | 0.6× | Normal |
Bar: disagreement-rate lift vs. mature-SKU baseline · scale 0–4.0× · tick marks 2.0×
2 of 4 slices over threshold
Evidence-linked improvement
An escape becomes a part the station is measured on
What the inspector overturned, and what a re-inspected sample found, come back as labelled parts the next release has to get right.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Escape and false-reject rates move on a defect class, a SKU or a shift.
- 02
Diagnose
Traced to the view captured, the reference set, the revision read or a fixture that moved.
- 03
Improve
The reference set, threshold or station setup changes under your change control, with a named approver.
- 04
Verify
Re-run against held-out parts from that class, including the ones that escaped, before it is released.
- 05
Learn
The escaped part is kept as a labelled case and the reason enters the control-plan review.
Learn → Detect
The return edge. A change to an inspection method is a control-plan change — it is approved and recorded before the station runs on it, not after.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, station and plan access, candidates and records, evaluation, integration, then supervised running and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Inspection-workflow discovery and boundaries.Week 1
- 02Control plan and drawing-characteristic mapping.Week 1, Week 2
- 03Station, gauge and MES access assessment.Week 1, Week 2
- 04Part, lot and station traceability keys.Week 2
- 05Image and measurement ingestion per station.Week 2, Week 3
- 06Reference defect library, labelled with your QA.Week 2, Week 3
- 07Candidate presentation and evidence assembly.Week 3, Week 4
- 08Revision-currency and gauge-capability gates.Week 4
- 09Control-limit and drift monitoring.Week 4, Week 5
- 10Evaluation suite and held-out defect classes.Week 4, Week 5
- 11Inspection-record write-back and routing.Week 5, Week 6
- 12Observability, deployment and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne station, one part family | ProductionProduction quality-system integration | AdvancedMulti-line / multi-site |
|---|---|---|---|
| Introduced at Pilot | |||
| Candidate defects presented with evidence | ✓ | ✓ | ✓ |
| Comparison against your control plan | ✓ | ✓ | ✓ |
| Drawing-revision currency checks | ✓ | ✓ | ✓ |
| Dimensional checks and gauge-capability gates | ✓ | ✓ | ✓ |
| Disposition by a qualified person | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Control-limit and drift monitoring | — | ✓ | ✓ |
| Inspection-record assembly and write-back | — | ✓ | ✓ |
| Observability and evaluation | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Per-defect-class and changeover slices | — | — | ✓ |
| Multi-line and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on stations and part families in scope, control-plan complexity, quality-system and gauge integrations, labelling effort and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your control plan and the drawing characteristics it names
→Control plan and drawing-characteristic mapping
Week 102Access to station images, gauges, CMM output and the MES
→Station, gauge and MES data-access assessment
Week 203How a part, lot and station are identified on the floor today
→Part, lot and station traceability keys
Week 204Labelled examples of the defects you actually see, rare ones included
→Reference defect library, labelled with your QA
Week 305Your gauge register, calibration status and MSA studies
→Revision-currency and gauge-capability gates
Week 406Escapes and false rejects from the last year, with the parts if you kept them
→Evaluation suite, held-out defect classes and failure-mode testing
Week 407Named inspectors and a quality engineer who dispositions
→Routing and inspection-record workflow, then supervised running
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 carries both the held-out defect classes and the first parts your inspectors disposition from.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Inspection-workflow discovery, control plan and drawing characteristics
W2Station, gauge and MES access, then part and lot traceability keys
W3Reference defect library, candidate presentation and evidence assembly
W4Evaluation suite, per-defect-class recall and gauge-capability gates
W5Held-out defect classes, first supervised parts and targeted corrections
W6Your inspectors disposition from the station, then Agent Care starts
- Reading the band
- Per-defect-class recall, rare classes included, is tested in week 4 — before any candidate is put in front of an inspector in week 5.
- At the end of W6
- The station has run alongside your existing inspection and been read by the inspectors who disposition from it, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Manufacturing AI agent
Build an inspection copilot around your control plan.
Show us one part family, its control plan and drawing characteristics, and the escapes and false rejects you have on file. We'll run your own parts past one station, report recall per defect class including the rare ones, and show you the record an inspector would sign.
- Typical delivery
- 4–6 weeks
- Deployment
- Your quality stack
- Disposition
- A qualified person, always
- After launch
- Agent Care available
Nestack Agents · Quality-inspection copilot
AGT-MFG-02 · Agent Care available after launch