In brief
In a machine shop, the inspection report isn't paperwork: it's a mandatory deliverable demanded by the customer, on the same footing as the part itself. Without it, the order isn't complete. And yet we treat it as a necessary evil — copied out by hand, at the end of the shift, by our most skilled people.
The hidden cost isn't the price of software. It's the time of your best inspector, spent transcribing dimensions instead of measuring parts. The good news: automatic reading of dimensioned drawings — AI framed by trade-specific software — can absorb the repetitive part. The judgment stays with the human. That balance is what holds up at audit (clause 8.6), not magic.
The inspection report, the deliverable we pretend to ignore
I spent fourteen years in quality before I started designing software. One thing never changes from one shop to the next: everyone knows the inspection report has to be done, and almost no one owns it as real work.
Let's be clear about what we're talking about. You machine a part from a dimensioned drawing. That drawing carries dozens of dimensions — each with its tolerance, sometimes its geometric tolerancing symbols. The customer doesn't pay you only for the part: they pay you for the proof that every dimension meets the drawing. That proof is the first-article inspection (FAI), then the dimensional inspection report that ships with the delivery.
In the standard, this step has a clear name: release of products and services (clause 8.6). You can't ship until the planned verifications are done and documented, with traceability of who authorized release. In other words: no report, no release. The document isn't beside the part. It's part of it.
And yet, go and see how it actually gets made. An inspector opens the drawing on screen or prints it. He reads a dimension. He copies it into a grid — often a spreadsheet. He notes the tolerance, works out the minimum and the maximum, sets up a column for the measurement. Then he moves to the next dimension. On a complex part, that's the routine, repeated dozens of times, before he's even touched a single part with a caliper.
The report isn't the boring part of the job. It's the part that proves the job was done.
Where the time really goes
Ask yourself honestly, with your own shop's numbers in hand: out of the total time of an inspection, how much goes to measuring, and how much goes to preparing and transcribing?
Because measuring is the craft. That's where experience counts: choosing the right instrument, understanding an ambiguous dimension, sensing that a part is running a little too loose. No machine replaces that.
But before and after the measurement, there's a layer of purely mechanical work. Reading a value printed on a drawing and retyping it. Recalculating a min and a max from a tolerance. Carrying measurements from a paper form into a clean file. Reformatting all of it so it looks like a deliverable document. That work calls for no judgment. It calls only for attention — and human attention, on a repetitive task, is exactly where errors slip in.
That's the hidden cost. It prints nowhere. It shows up on no quote. It's paid in hours of your most qualified staff, busy doing transcription work. And it's paid a second time, later, when a mis-copied dimension triggers a customer return.
Now multiply. A shop that produces a serious volume of reports a month doesn't lose "a few minutes." Month after month, it ties up the equivalent of whole days of an inspector paid at an expert's rate — to do work that, at bottom, doesn't call for that expertise.
The calculation, only you can do, because only your numbers count. But the form is always the same: (a few hours of transcription per week) × (the hourly rate of an experienced inspector) × (fifty-two weeks). Put your own values into that equation. The product is a very real annual cost — the one nobody writes down anywhere, on no quote, in no production report. The result often surprises the owners I show it to.
What AI does well here — and why, alone, it isn't enough
The repetitive layer I've just described is precisely what a document-reading AI knows how to do. You show it a zone of the drawing; it reads the dimension, the tolerance, the symbols. It proposes a table row: dimension, min, nominal, max. It places a numbered balloon on the drawing at the exact location of the dimension — the ballooned drawing.
On paper, it looks like a revolution. In practice, if you stop there, you've built a trap.
Because an AI left to itself has a fault that's brutal for quality: it fails silently. It doesn't tell you "I'm not sure about this dimension." It gives you a number, with the same confidence whether it's right or wrong. A slightly blurred drawing, a dimension on a busy background, a tolerance written in an unusual way — and it hands you a value that's plausible but false. On a deliverable that serves as proof, a plausible error is far worse than an obvious blank.
I've seen that kind of trap up close. On a handwritten-form reading module I designed, an old technical ceiling meant that beyond a certain number of lines, a multi-page scan came back with zero values — without a single error message. The document looked processed. It wasn't. That's the kind of failure you can't see with the naked eye, and that jumps out at you on audit day or on a customer return.
The lesson, I repeat it every time: an AI without a frame is a stage set. It's convincing as long as no one leans on it. What makes it reliable isn't a better model. It's the trade-specific software around it.
Framing the AI: what a well-built quality software enforces
Concretely, what does "framing" mean? Not words. Hard rules, coded into the tool, that stop the AI from doing damage. Here are the ones that matter for an inspection report.
Make doubt visible. A good tool doesn't settle for an answer: it knows its confidence level. Below a threshold, the value is highlighted for human review, not quietly slipped into the table. The inspector sees at a glance the five dimensions to check instead of re-reading the fifty. You don't remove human control: you concentrate it where it's useful.
Link, don't guess. When the AI reads a dimension number on a hand-filled form, good software doesn't let it guess which table row that number belongs to. It reads the printed number word for word, then a deterministic mechanism — fixed rules, not intuition — makes the link. Anything a certain calculation can establish should never be handed to a model's hunch. That's the border between the 80% the AI takes on and the 20% of judgment that stays human.
Understand the shop's language. A dimensioned drawing isn't text. A tolerance can be written as a standardized fit, an angle in degrees-minutes-seconds, an imperial fraction, a one-sided dimension of the "X max" type. Geometric tolerancing symbols each carry a precise meaning. A generic tool reads characters; a trade-specific tool understands a dimension. The difference is the errors you don't make.
We have to be honest on one point: reading a dimensioned and ballooned drawing reliably is a genuine technical challenge — dense geometry, overlapping balloons, mixed notations, busy backgrounds — and it's precisely because this isn't trivial that the software frame around the AI is not a luxury, but the condition for the result to hold.
Guarantee that what you see is what you ship. The ballooned drawing shown on screen must be exactly the one in the final PDF. That seems obvious. It isn't: in my own tool, I had to forbid certain structural changes directly in the editor, because they shifted the balloons between the screen and the exported document. A numbered balloon exists only attached to its dimension — never on its own. That deliberate rigidity is what guarantees that a balloon on the drawing always points to the right row in the table.
None of these rules is spectacular. Together, they make the difference between a gadget that dazzles in a demo and a tool a shop dares to build a contractual deliverable on.
| Task | Who does it best |
|---|---|
| Read the drawing's dimensions, build the table, calculate min/max | The AI, framed |
| Place the numbered balloons on the drawing | The AI, framed |
| Flag doubtful dimensions for review | The software (confidence threshold) |
| Choose the instrument, measure the part | The human |
| Settle an ambiguous dimension, judge a borderline case | The human |
| Authorize release of the product (clause 8.6) | The human |
The report doesn't live alone: the metrology behind every measurement
There's a blind spot we rarely talk about when we talk about inspection reports: with what were the measurements taken?
A dimensional inspection report rests entirely on the instruments that produced its numbers. If the caliper or micrometer used wasn't calibrated, or its calibration had lapsed, every value in the report loses its worth as proof. The standard is plain on this: monitoring and measuring resources (clause 7.1.5) must be verified, calibrated, and their status known. The traceability of the measurement is part of the traceability of the part.
In many shops, this management still lives in a spreadsheet someone updates "when they think of it." It holds up as long as no one asks. The day a customer asks for the calibration certificate of the instrument used to measure a given dimension on a given part, the scramble begins.
A well-designed software ecosystem closes that gap with no extra effort: the instrument used for a measurement is tied to its calibration record, due dates trigger alerts before they lapse, and an instrument with an expired certificate doesn't quietly slip into a report. The inspection report and the metrology aren't two separate files: they're two ends of the same chain of proof. Good software keeps them linked.
What a shop in the Sorel region understood
I think of a well-regarded machine shop in the Sorel region, the kind where thirty-odd people run machines all day and the inspection report is an everyday deliverable, not an exception.
What convinced them wasn't a demonstration of artificial intelligence. It was a very down-to-earth observation: their most experienced inspectors were spending an unreasonable share of their time on transcription work, and that share showed up in no production report. The time was real; it was simply invisible.
The turning point wasn't "AI will do the reports." That, rightly, worried them. The turning point was seeing that the tool showed its uncertainty: that it highlighted what it wasn't sure of, that it left the inspector to validate, that it kept the release signature on the human side. In other words, that responsibility didn't change hands. The AI prepared; the human checked and signed. The report stayed theirs.
That's always the moment the real gain appears. Not because we "added AI," but because we gave the inspector back the hours he was losing to copying, without taking an ounce of control over what leaves the shop.
In closing: what if you costed what you don't see?
The inspection report's cost is hidden because it runs through no accounting line. It's paid in expert time diverted, in transcription errors that come back as returns, in calibration certificates hunted for at the last minute. None of it shows up on a quote. All of it weighs on your margin.
So the questions I leave you with are simple, and they answer with your own numbers, not mine. Out of the total time of an inspection at your shop, what share is truly judgment, and what share is nothing but copying? How many of your recent nonconformities come from a defective part, and how many from mis-reported data? And if a customer asked you tomorrow for the calibration certificate behind a measurement from three months ago, how long would it take you to produce it?
If those answers give you pause, it isn't AI you need first. It's to look at the part of the work no one counts — then to decide what, in that part, still deserves a human's time, and what can be framed by a trade-specific tool to give it back.
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 piling on administrative work.