"Can AI do quality inspection?" It can do the looking that tires people out. A camera and software can check every part against what a good part looks like, flag the ones that differ, take measurements the same way every time, notice when a process is slowly moving away from its target and keep a record of everything it saw. What it cannot do is decide how much difference is too much, decide what happens to a batch that might be bad, sign the certificate that says the order meets the customer's requirements, or pick up the phone when a shipment is rejected. This article is for the owner of a small manufacturer whose inspection is done by a few experienced people with gauges, a bench and a clipboard, and who is being offered a system that says it will "automate quality".
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them, so I have seen what a customer does with a quality escape: it becomes a conversation about whether they can trust you. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs.
What cameras and software do well
Flagging defects. Show a system good parts and parts with the faults you care about, and it can learn to flag a scratch, a missing feature, a colour that is off or a gap that should not be there. It looks at every part with the same attention on the last one of the shift as on the first.
Measuring parts. Software can take dimensions from an image or a sensor the same way every time and compare them with the drawing. That removes the difference between one person's reading and another's.
Spotting drift. A single part can pass while the process is creeping toward the limit. A tool can chart the measurements as they come in and warn you that a dimension is trending before it goes out of tolerance, which is often the cheapest moment to fix a tool or a setting.
Keeping the inspection record. Which part, which lot, when, what was measured, what the result was, an image of what the system saw. Many small shops hold this on paper or in one person's memory, and finding it after a complaint takes days. A tool that keeps it as it goes makes the record easy to search.
Sorting what needs a human. A system can pass the clear cases and send the borderline ones to an inspector with the image and the reading, so the person spends time where judgment is needed.
Reading the paperwork. Drawings, specifications and customer requirements can be summarised into a checklist for a new job. Can AI write standard operating procedures covers the same step for the written method behind an inspection.
What a person must keep
The accept and reject line. Where good ends and bad begins is a decision. The drawing gives limits, but the real line is also what this customer will accept, what the part is used for and what a cosmetic mark means on this product. A system can apply a line. It cannot set one.
What happens to a suspect batch. Hold it, sort it, rework it, scrap it, ask the customer for a concession or ship it with a note. Each choice has a cost and a risk, and it is the owner's or the quality manager's to make.
The certificate of conformance. When your name goes on a document that says the order meets the requirements, a person is stating that it is true. A report from a system is evidence for that statement. It is not the statement.
Talking to the customer. When something escapes, the customer wants to hear from someone who understands the part, the process and the fix. A tool can pull up the record in seconds, and a person has to explain it.
Finding the cause. A tool can show that defects rose on Tuesday afternoon. It does not know that the second shift changed the coolant or that a supplier sent a different lot of material. The people on the floor do.
Calling it when the system is wrong. Every system misses some faults and flags some good parts. An experienced inspector can tell when a result does not look right, and that judgment is worth protecting.
The risks that are specific to inspection
Missed defects. A system finds the kinds of fault it was shown. A new kind, in a new place, may pass without notice. Do not assume that no flag means no fault.
False alarms. If the system flags too many good parts, people stop believing it and start overriding it, and then a real flag gets ignored with the rest.
Changed conditions. A change in lighting, a new material finish, a different supplier, a camera that has been knocked out of place or a worn fixture can change what the system sees. Something that worked in March may not work in June.
An unclear line. If two inspectors disagree about what counts as a defect, the system will inherit the disagreement. Agree the standard first, with samples, and write it down.
Records that nobody checks. A tidy record only helps if someone reads it. Decide who looks at the trends, how often and what they do about them.
Customer rules. Many customers set requirements for how inspection is done, what records you keep and who may sign. Check your agreements before you change the method, and tell the customer when the contract says you must.
Drawings and part images in the wrong place. Drawings and photos of parts often carry confidentiality terms. What happens to your data explains what to ask a provider, and is ChatGPT safe for business use lists what to check. Use a company account, not a personal one.
False confidence. A dashboard full of green looks like quality. It is an answer to the checks you set up, and a customer will find the ones you did not.
Starting safely
Begin with one defect on one part. Pick a fault that costs you real money or complaints, on a part you make often, where today's check is a person looking.
Agree the standard first. Gather good parts, bad parts and the borderline ones, and have your inspectors say which are which. Write the rule down.
Run it beside your inspectors. Let the system check the same parts that a person checks, and compare the results for a few weeks. Count what it caught, what it missed and what it flagged by mistake. Nothing it says should stop a shipment yet.
Keep it advisory. The system flags and the quality person decides. A suspect batch is held by a person, and a certificate is signed by a person.
Use an approved tool on a company account. Check what the provider keeps and whether it learns from what you give it.
Keep a log. What the system flagged, what the inspector decided, and what the customer later said. After a couple of months, keep it, change it or stop.
If the same machines are also the bottleneck, can AI do production scheduling draws the same line between a proposal and a promise. For the shop side, the AI for a metal fabrication shop page walks through where estimating, programming, scheduling and quality sit for a shop of that kind.
Picture it going wrong
A machine shop with thirty people trials a camera system on one machined part. For two weeks it matches the inspector, and the owner lets it replace the second check on that part to free a person for the floor. Then a new lot of material arrives with a different surface finish. The system keeps passing parts, and a small burr that the inspector would have felt with a thumb goes through. A customer's incoming check finds it on the third delivery, and the certificate of conformance carries the owner's signature. Nobody lied and the system did what it was shown. A tool that had been checked on one condition was trusted on another, and a signature stood behind a result nobody had looked at.
Mistakes that repeat
Removing the human check before running side by side. Letting the system set the reject line. Trusting it after the material, lighting or fixture changed. Starting before the inspectors agree what a defect is. Signing certificates from a report nobody read. Ignoring false alarms until the real one is ignored too. Skipping the customer's inspection requirements. Putting drawings and part photos in a personal account.
Frequently asked questions
Can AI do quality inspection?
It can check every part for visible faults, take measurements the same way each time, spot drift and keep the record. A person sets the accept and reject line, decides what happens to a suspect batch and signs the certificate of conformance.
Can AI replace my quality inspector?
It is better thought of as taking the repetitive looking, so your inspector has time for the borderline parts, the causes and the customer. Run it beside the inspector first and keep a person responsible for every release.
Is there AI quality inspection software for a small manufacturer?
There are systems sold for this, from camera-based inspection to software that charts your measurements. Whichever you look at, test it on your own good and bad parts, under your own lighting and materials, before you trust it.
What does an AI inspection system need to work?
Good examples of acceptable and unacceptable parts, a consistent setup for the camera and lighting, the drawing limits and a clear written standard for what counts as a defect. If the examples or the standard are unclear, the results will be too.
Can AI sign the certificate of conformance?
No, a person should. The certificate is a statement that the order meets the requirements, and someone has to stand behind it. The system's record supports that person.
Can I use drawings and part photos in an AI tool?
Only if your agreement with the customer allows it and the tool is approved, on a company account. If you are unsure, ask your lawyer before you upload.
Where should I start?
Pick one defect on one part, agree the standard, run the system beside your inspectors, keep it advisory and log what it flags and what the inspector decides.
Not sure where you stand? Take the 3-minute AI readiness scorecard.