In brief
Artificial intelligence is arriving on the shop floor, and two camps are squaring off: those who believe it will do everything, and those who believe it will break everything. After fourteen years in industrial quality and two years building tools for real machine shops, my answer fits on a single line: framed properly, AI does roughly 80 % of the repetitive work of quality — compiling, tracing, preparing — and leaves untouched the 20 % that matters: deciding, ruling, owning the call.
The boundary is not technical, it is moral. A machine can prepare a flawless nonconformity file; it cannot sign the decision to close it. Because to sign is to answer for your choices before a customer, an auditor, a prime. And responsibility cannot be automated. This article draws that line precisely — and explains why a chatbot delivered bare and an AI framed by trade-specific software are not playing in the same league.
The division of labour, said plainly
Let us start with the concrete, because that is where the confusion sets in. When people say "AI does the work," which work are we talking about?
In a quality system there are two kinds of work. There is mechanical work: copying out dimensions, compiling scattered nonconformities, tracking down the latest calibration certificate for an instrument, reconstructing a history the night before an audit, linking one piece of evidence to another. This work is time-consuming, thankless, and it calls for no judgment. It calls for rigour and hours — two things a manufacturing SME never has in sufficient supply.
And there is judgment work: deciding which nonconformity to correct first, judging whether a corrective action has truly fixed the root of the problem, accepting or refusing a residual risk, signing off a management review. That work is rare, dense, and it commits someone.
The right way to think about AI in quality is precisely this split: it takes the mechanical, you keep the judgment. Not because the machine is too dim for the rest — it would often produce a plausible answer — but because the rest is not a matter of calculation. It is a matter of responsibility.
AI prepares the file. The human signs it. That sentence should be posted on the wall of every shop that adopts these tools.
Why responsibility cannot be automated
This is the point I defend most firmly, so let us take the time to lay it out.
A quality system, stripped of its jargon, is a mechanism by which an organization answers for what it produces. When a prime receives a part, it is not receiving only a machined piece of metal: it is receiving the commitment of a shop that affirms the part conforms, that the process is under control, that deviations have been dealt with. That commitment has a name, in law and in practice: responsibility.
And you cannot delegate a responsibility to a thing that bears none. If an AI "decides" to close a nonconformity and the defect resurfaces at the customer's plant three months later, who answers? Not the algorithm. You do. The quality lead who validated it, or who should have. The machine has no licence, no signature, no professional conscience to put on the line. It loses nothing if it is wrong.
That is why the 20 % boundary is not negotiable. This is not timid technological caution. It is the recognition that a quality decision is an act, taken by someone who owns its consequences. ISO 9001 has always understood this: through its corrective action requirement (clause 10.2), it asks not that a problem vanish by magic, but that a person analyze its cause, decide on a response, and verify that it worked. You correct a nonconformity (clause 8.7) attached to a process; you do not tick a box.
A machine can prepare each of those steps. It can gather the facts, propose an analysis, draft a first version. It cannot be the person who answers. The day we confuse the two, we no longer have a quality system — we have a stage set.
Bare chatbot versus framed AI: the distinction that changes everything
When most people hear "AI," they picture a conversational bot. You type a question, it spits out an answer. That image feeds both the naive promise ("it will write my quality manual") and the legitimate fear ("it will make things up"). And both are right — about the bare chatbot. Delivered on its own, with nothing to anchor it, a generative AI produces plausible, groundless text. It does not know your machines, your history, your rules. It improvises with a confidence that, in quality, is a hazard.
But there is an entirely different way to use this technology: AI framed by trade-specific software. Here the AI is no longer an oracle taken at its word. It is a tool constrained by an application that forces it to rely on your real data, that files each piece of evidence in its place, that checks its outputs, and that refuses shortcuts. The difference between the two is as wide as the difference between a rumour and a documented file.
A concrete example says more than a long speech. Take the extraction of a dimensioned drawing. A general-purpose chatbot shown a technical drawing will "read" numbers and spit them back — sometimes right, sometimes invented, with no way to tell which. An AI framed by inspection software extracts each dimension within a structure: every value is tied to its balloon, presented for human verification, and nothing is validated until an eye has confirmed it. The AI proposes; the frame disciplines; the human disposes. The same engine, used bare, would be a dangerous gadget; used framed, it becomes a reliable accelerator.
This is where the whole thesis is decided. AI without a frame dilutes and improvises. AI inside a well-built application does the opposite: it applies a rigour that the human hand, tired on a Friday afternoon, does not always apply. The technology is the same. What changes is the harness.
What a good frame actually does
What is a "frame" concretely? It is not a slogan. It is a set of precise behaviours that well-built software enforces, and that a bare chatbot knows nothing about.
It anchors the AI to your real data. It does not reason over generalities: it works on your nonconformities, your instruments, your processes. What it produces is verifiable line by line, because every element comes from your own shop.
It refuses to break a piece of evidence. When one object is referenced by another — a corrective action (clause 10.2) attached to its nonconformity, an action to address a risk (clause 6.1) attached to a risk analysis — good software flatly refuses to delete it. You cannot make evidence disappear "by accident." That is the control of documented information (clause 7.5) applied automatically, without relying on anyone's discipline.
It shows the gaps instead of papering over them. This is perhaps the behaviour that best captures the philosophy. When a piece of data is incomplete, a good frame displays it as incomplete. It does not fill the void with a reassuring invention. A bare chatbot, ordered to produce an answer, will plug the hole at any cost. A framed AI flags it: "this is missing, I am not deciding in your place." In quality, that honesty is worth its weight in gold — it is the exact opposite of the box-ticking theatre that seasoned ISO hands, rightly, dread.
It stops at the right place. Good software brings the prepared file — the linked facts, the alerts, the state of readiness — and then it hands back control. It is built to assist judgment, never to simulate it. The point where it stops is precisely the 20 % boundary.
| What framed AI does well — 80 %, the mechanical | What stays 100 % human — 20 %, the judgment |
|---|---|
| Extract, structure, compile information (dimensions, files, history) | Decide what to correct, and in what order of priority |
| Link each piece of evidence by the rules of the trade, without breaks | Judge whether a nonconformity is truly closed, on the substance |
| Monitor deadlines and flag deviations | Accept or refuse a residual risk (clause 6.1) |
| Prepare the file and the draft of a corrective action (clause 10.2) | Validate the action, sign it, answer for it before the auditor |
| Display a complete, up-to-date picture, including its own gaps | Give it meaning, keep the quality culture alive on the floor |
Look at the left-hand column: real work, time-consuming, without judgment. This is exactly what an SME cannot afford to carry on the shoulders of a full-time engineer. Look at the right-hand column: nothing in there should ever leave the hands of a responsible person. Honest software is software that knows, on every screen, which column it belongs to.
The trap to avoid: automating the proof, forgetting the substance
The danger has to be named, because it is real. The temptation, with these tools, is to automate the proof while forgetting the substance. To end up with a system that ticks every box in the standard beautifully, without a single real improvement having occurred on the floor.
This drift is not caused by AI. It is caused by a misuse of AI — the kind that crosses the 20 % boundary and lets the machine "decide" in order to go faster. A shop that lets an algorithm close its nonconformities with no human eye will save time in the first year and lose its customers' trust in the second, when the same defects come back.
Picture the scenario, because I have seen situations that came dangerously close to it. An overstretched owner-manager connects a generic tool to their history and asks it to "sort" and "close" the pending nonconformities. The machine complies with aplomb: in a few minutes, everything is sorted, everything looks in order. Nobody reads it back. Then comes the audit — or worse, a nonconforming part that leaves for the prime. It is discovered, then, that serious cases were filed as minor, that root causes were never analyzed, that boxes were ticked over nothing. The tool answers for none of it. The owner-manager does. Responsibility did not move an inch: it stayed, entire, on the human side — except that control had been abandoned.
I have seen it from the other side, in the field. At a well-regarded machine shop in the Sorel region, we took on several years of scattered nonconformity history. The software did the heavy lifting: gathering everything, linking everything, making everything legible. But what created value was not the compilation itself — it was the moment a human looked at the reconstructed whole and asked: how many times has this same defect come back? On which machine? What have we never actually fixed at the root? The tool made the question possible. Only human judgment could answer it. Remove the human, and you are left with nothing but a fine inventory of problems that carry on.
The substance of quality — the will to improve, the decision to act, the responsibility owned — is delegated to no machine. A good frame protects that substance; a bad use empties it out. The technology does not settle the matter; the intention of the person using it does.
In closing: the right line, not the right machine
The question, then, is not "is AI good or bad for quality." That is a badly framed question. The real question is: where do we draw the line — and does the tool we adopt respect it?
My conviction, forged on real shop floors, sits in the split I have described. AI framed by trade-specific software absorbs the mechanical work that used to crush SMEs and keep them shut out of structured quality. It does it better, faster, and without ever forgetting a piece of evidence. But it stops dead at the boundary of judgment, because on the other side responsibility begins — and responsibility cannot be programmed.
Open questions remain, which I prefer to ask rather than pretend to resolve. How far does a company want to entrust the memory of its quality to a tool, even an honest one? How should an auditor verify a system where the mechanical is traced automatically — by focusing more on human judgment, since that is where the risk shifts? And we, the builders of these tools: how far is it our duty to make visible, on the screen, the line the user must not cross?
I have no final answer. I have a practitioner's conviction: a good AI tool for quality is not the one that does the most. It is the one that knows, better than anyone, what it must never do in your place.
The principles described in this article are the ones that guided the development of Asterion Solutions, a suite of trade-specific software built for manufacturing SMEs that want to structure their quality without multiplying administrative tasks.