Incoming quality control with AI vision: a practical guide

What incoming quality control is

Incoming quality control (IQC), also called incoming or receiving inspection, sits between the supplier's shipping dock and your production line. A delivery arrives, an inspector takes parts from it, checks them against what was ordered and specified, and the lot is accepted, held or rejected. What gets through here is built into your product; what is rejected here becomes a claim against the supplier instead of a defect at your customer.

What an incoming inspector checks

• Conformity to the drawing or the golden sample. The part has the features the drawing calls for — holes, threads, chamfers, inserts, markings — and looks like the approved reference sample the supplier was qualified on.
• Visible defects. Scratches, dents, burrs and sharp edges, cracks, flash on molded parts, missing or incomplete features, contamination.
• Color and finish. The right color, gloss and surface finish — paint, powder coat, anodizing, plating — without runs, bare spots or discoloration.
• Labels and traceability. Part number, revision, lot or batch number and date code on the label match the order and the delivery note, and the date code is within what you accept.
• Packaging and pallets on arrival. Crushed cartons, torn wrap, broken pallets, wet or tilted loads — damage that has to be recorded before the delivery is signed for.
• Quantity. The count in the box, tray or reel matches the label and the delivery note.
• Documentation for rejects. A rejected lot needs evidence the supplier will accept: what was wrong, on which parts, from which lot, with pictures.

Why traditional methods miss defects at incoming inspection

Incoming inspection relies on people, on samples, against criteria that are partly written and partly carried in the inspector's head. Each of those has a blind spot.

Sampling plans and their blind spots

Inspecting every part of every delivery by hand is slow and costly, so incoming inspection relies on sampling. Sampling plans by attributes can be built on ISO 2859-1 (Sampling procedures for inspection by attributes — Part 1), which indexes sampling schemes by the acceptance quality limit (AQL) for lot-by-lot inspection. A sampling plan is a sound way to decide on a lot, but it has limits worth keeping in view:

• A sample decides for the lot. An accepted lot is not a lot without defects; it is a lot whose sample did not show enough of them. The parts outside the sample are not looked at.
• Defects can cluster. A worn tool, one cavity of a mold, one box packed at the end of a run: when defects sit in one part of a lot, the sample may simply not reach them.
• The plan assumes the inspector sees what is there. The numbers behind a sampling plan count the defects found in the sample. A scratch missed on a sampled part counts as a good part.
• Cosmetic criteria are read differently. "No visible scratches on the A-surface" means one thing to one inspector and another to the next, and the line moves with the person on shift.

Inspector fatigue and paperwork

• Fatigue. Turning parts under a lamp, box after box, is repetitive work. Attention drops over a shift, and small defects get harder to see.
• Delivery pressure. When the line is waiting for the parts, the inspection is under pressure to finish, not to look harder.
• Weak evidence for claims. A reject is only as strong as its record. A handwritten note and a phone photo without a lot number are hard to turn into a supplier claim weeks later.
• No memory across deliveries. Each lot is judged on its own. Whether the same supplier sent the same burr three deliveries ago lives in someone's memory or a spreadsheet, if anywhere.

Automated optical inspection with fixed rules helps where parts are uniform and presented the same way every time, as on an electronics line — see our guide on AOI false calls and PCB defects. At receiving, part types change with every delivery, parts arrive in bags, trays and boxes, and the criteria may be written in words rather than measurements. That is where fixed rules struggle.

How vision-language models make it solvable

Primarch builds on and enhances vision-language models (VLMs): models that read images and language together. For incoming inspection, that means the criteria can be given to the model the way they are given to an inspector — in words, with reference pictures and the knowledge of what the part is for — instead of being turned into a fixed measurement window for each part.

• It reads the part, not just the pixels. The model sees what the part is, which surfaces matter and which features it should have, and reads a defect in that context: a scratch on a hidden face is not the same as a scratch on the visible one.
• It brings domain knowledge to the image. Your acceptance criteria, the features from the drawing and what a burr or a run in the paint looks like are put on top of the model, so its reading follows the rules your inspectors work to.
• It applies the same criteria every time. The model does not tire at the end of a shift and does not read "visible scratch" differently from one day to the next. How much of each lot is imaged — a sample or every part that passes the camera — is a decision about your receiving area, not a limit of the reading.
• It explains its finding in plain language. Each result is a sentence a quality engineer can check — what was found, where on the part, why it counts against the criteria — next to the image it was read from.
• It remembers. Findings are kept, so they can be searched later and earlier findings showing the same defect can be brought up, instead of each lot being judged on its own.

The useful question is not "did the sample pass?" It is "what did these parts actually look like, and can we show it?" Reading each imaged part against your criteria and keeping the evidence answers both.

Vision can only judge what the image shows. Dimensions within tight tolerances, material properties, hidden internal features and anything that needs a gauge or a test stay with measurement equipment and the lab. Comparing color and gloss with the reference image needs the same lighting at the bench; measured color and gloss values stay with a colorimeter and gloss meter.

What to look for in a solution

Whichever vendor you talk to, these are fair questions to ask about an AI vision system for incoming inspection:

• Uses your criteria. Can your drawings, acceptance criteria and approved samples be built into how it decides, or does it apply its own idea of a good part?
• Copes with variety. Receiving sees many part types. How much work is it to add a new part or a new supplier — and does that need retraining from scratch?
• Works with how parts arrive. Bench camera, photos, trays, bags, pallets: what image setup does it need at your receiving area?
• Explains every finding. Each reject should come with a reason and an image a person can check, not just a pass or a fail.
• Produces evidence a supplier will accept. Can a reject be documented with its images, the part and the lot, in a form you can send with a claim?
• Shows patterns across deliveries. Can findings be searched later, so you see what keeps coming back from which supplier?
• Is proven on your parts. A pilot on your own parts and deliveries, judged against measurable criteria agreed before it starts — not a demo set.
• Keeps your data with you. Drawings and part images can be confidential, yours and your customers'. Is there an on-premise option where the data never leaves your facility?

Example use cases

The walk-throughs below are illustrative examples of how an incoming inspection finding is built. They are not customer cases and contain no performance figures.

Example 1: machined housings with a missing thread and a burr

1. 01

   Moment

   A delivery of machined aluminum housings arrives. The inspector places sampled parts at the inspection bench camera.

2. 02

   What the inspection checks

   The features the drawing calls for, written into the expert's knowledge base in discovery, and visible defects on the machined edges.

3. 03

   What the expert reads

   One housing has three of its four mounting holes threaded; the fourth is drilled but not tapped. Another has a raised burr along the edge of the sealing face.

4. 04

   The finding + image

   "Suggested reject: mounting hole at the lower right is drilled but not threaded; the drawing calls for a threaded hole. Second part: burr along the sealing face edge." The images of both parts are attached.

5. 05

   Who sees it

   The incoming inspector, who puts the lot on hold, and the supplier quality engineer.

6. 06

   What the pattern shows

   Asked in plain language — "show earlier findings of an untapped hole" — the memory returns earlier findings showing the same untapped hole, which the engineer can review before raising it with the supplier.

Example 2: powder-coated brackets in the wrong finish

1. 01

   Moment

   Powder-coated steel brackets for a visible part of the product arrive in boxes.

2. 02

   What the inspection checks

   Color and gloss against the approved reference sample, and coating defects.

3. 03

   What the expert reads

   The brackets are a visibly lighter shade and glossier than the approved sample; several have bare spots at the hanging point.

4. 04

   The finding + image

   "Suggested reject: finish does not match the approved sample — lighter and glossier. Coating missing at the hanging point on several parts." The images, next to the reference sample image, are attached.

5. 05

   Who sees it

   The incoming inspector and the quality engineer, who decides whether the lot goes back.

6. 06

   What the pattern shows

   It stays a single event in the record, with the evidence kept for the claim.

Example 3: a label that doesn't match the delivery

1. 01

   Moment

   Reels of electronic components arrive with the supplier's labels and a delivery note.

2. 02

   What the inspection checks

   The part number, lot number and date code on each label, against the order and the delivery note.

3. 03

   What the expert reads

   One reel carries a different lot number from the delivery note, and its date code is older than the age the order allows.

4. 04

   The finding + image

   "Suggested hold: lot number on this reel's label does not match the delivery note; date code is older than the order allows." The image of the label is attached.

5. 05

   Who sees it

   The receiving clerk and the incoming inspector, before the reel goes into stock.

6. 06

   What the pattern shows

   The quality engineer asks for earlier label findings on deliveries from the same distributor to see whether this has happened before.

Example 4: a damaged pallet recorded on arrival

1. 01

   Moment

   A pallet of molded plastic covers is unloaded at the receiving dock, where a camera sees the unloading area.

2. 02

   What the inspection checks

   The condition of the load on arrival: pallet, wrap and cartons.

3. 03

   What the expert reads

   One block of the pallet is broken on the left side, and the load is tilting toward it.

4. 04

   The finding + image

   "Damage on arrival: pallet block broken on the left side, load tilted toward that side." The time-stamped clip of the unloading is attached.

5. 05

   Who sees it

   The receiving supervisor, before the delivery is signed for, so the damage can be noted on the paperwork.

6. 06

   What the pattern shows

   The covers from the tilted load go to incoming inspection; any parts rejected there get their own findings, and both records are kept for the claim.

How Primarch's Custom Experts help

This section is about our product. Custom Experts put your domain knowledge on top of the Primarch platform core, for problems that can be seen but are not in a catalog. We don't train models from scratch; what makes the expert custom is the knowledge. For incoming quality control:

• Configured for your parts and your criteria. The expert profile is built from a domain knowledge base — your drawings, acceptance criteria and defect definitions — with rule templates and thresholds, and targeted fine-tuning if needed.
• Visual quality control on the parts you receive. Which defect types the expert reads — scratches, dents, burrs, missing features, finish — is defined with you in discovery and tested in the pilot.
• Reads any analyzable visual input. A camera at the inspection bench, photos taken by the inspector, periodic captures or microscope output can all feed the expert.
• Writes each finding in plain language, next to the image or clip it was read from, so the inspector, the engineer and the supplier see the same thing.
• Keeps a queryable memory. Findings are written into a memory you can ask in plain language.
• Can be configured to compare parts with your approved reference sample — color, finish and appearance — using images of the golden sample in the knowledge base.
• Reads label text — part number, lot number, date code — and checks it against the order and the delivery note.
• Can be configured to record packaging and pallet damage on arrival where a camera sees the receiving dock.
• Can be configured to count parts in a box, tray or bag at the inspection bench — stock counting from drone footage is one counting example.
• Records each finding with the supplier, part number and lot it belongs to, so findings can be searched by supplier and used in a claim.

Putting it to work

1. 01

   Discovery

   We define the problem and the success criteria together: which parts and suppliers, which defects, which decision the expert supports — accept, hold or reject.

2. 02

   Configuration

   An expert profile is built from a domain knowledge base, rule templates and thresholds; targeted fine-tuning if needed.

3. 03

   Pilot

   Validation on real images from your own receiving area, tested together against measurable criteria. Depending on scope, discovery to pilot typically takes a few weeks.

4. 04

   Deployment

   On-premise or cloud rollout, team training and a continuous improvement loop.

• Alerts, digests and dashboards. Instant notifications, periodic digests and role-based dashboards, so each finding reaches the inspector, engineer or buyer who acts on it.
• Your data stays with you. With on-premise deployment, development and validation can happen inside your facility; your drawings, part images and data never leave.