Most articles about AI in manufacturing are written by people who have never been inside a factory. They describe automation in abstract terms — efficiency gains, cost reductions, machine learning applied to supply chains — without ever explaining what the actual work looks like before any of that technology touches it.
This is a different kind of article. We have been running woven label factories in Hong Kong, China and now also Vietnam since 1996. We operate over 60 Jacquard weaving lines. We produce labels for fashion brands ranging from independent designers ordering 50 pieces to established names ordering hundreds of thousands per season. I want to explain, from the inside, what the production process actually looks like — and where AI is already changing it, where it is about to change it, and where the real opportunity lies.
First: how a woven label actually gets made
To understand where AI fits, you need to understand the process it is being applied to. Most people outside the industry have no idea how many human decisions go into a single woven label. Here is what actually happens between a brand sending a logo and a finished label arriving at their studio.
A brand submits a design. Usually a logo, sometimes a combination of logo, brand name, and size or care information. The file arrives as a PDF, an AI, a JPG, or sometimes a low-resolution screenshot of something they found on their own website. This is the starting point.
A designer at the factory — a human, working in specialist weaving software — takes that file and converts it into a machine-readable weaving specification. This is not a simple file conversion. Woven labels are not printed. Every colour in the design must become a physical yarn colour. Every line, curve, and letter must be translated into a weaving pattern — a grid of thread intersections that, at the density of HD Damask weaving, runs to hundreds of thousands of individual decisions per square centimetre of label. The designer is not just copying the artwork. They are rebuilding it in a completely different medium, making judgement calls at every step about which curves can be reproduced accurately at the available thread density, which font weights will hold at small sizes, and which colour gradients need to be simplified to the available Pantone-matched yarn library.
That file goes to the factory floor. A technician selects the yarn colours — physically, from a library of hundreds of dyed polyester threads, each assigned to a Pantone reference — and sets up the loom. The machine is loaded with the correct yarn sequence. A sample run is made. A few centimetres of label come off the loom.
That sample is photographed or physically shipped to the brand. The brand looks at it against their original artwork. In most cases — for first samples on new designs — there are changes. The navy looks too purple. The logo mark is slightly heavier than intended. The brand name needs to be two thread counts thinner. The feedback comes back to the factory. The designer opens the weaving file, adjusts the specification, the technician may change a yarn or adjust the machine tension, and another sample is produced.
For a large brand with exacting standards — a luxury fashion house, a sportswear company with a brand standards manual — this cycle can repeat four or five times before a sample is approved. Each cycle takes days. Each one requires human expertise at multiple stages. And all of this happens before a single production label is made.
This is the reality of woven label production at scale. It is not a fast process. It is not a simple process. And almost none of it, until very recently, has been touched by any form of automated intelligence.
What AI is already doing: production scheduling and supply chain
The area where AI has had the most immediate practical impact in our factory is not the glamorous end of the process. It is scheduling.
A woven label factory running 60+ looms is a genuinely complex logistical system. Every loom runs a specific yarn configuration. Changing that configuration — swapping yarns, re-threading, adjusting density settings — takes time. The sequence in which orders are run matters. Running two orders that share the same yarn configuration back to back saves setup time. Running a long production run of a standard product before a complex short run of a specialty product is more efficient than the reverse. Scheduling decisions of this kind, across dozens of machines and hundreds of concurrent orders at different stages of the approval cycle, used to be made by experienced production managers based on pattern recognition built up over years.
AI-assisted scheduling changes this. A system that can model the full production queue — factoring in yarn inventory levels, loom availability, order deadlines, setup time costs, and the probability that any given order will require a resample — can produce a more efficient schedule than a human working from experience alone. Not because the AI understands weaving better than an experienced production manager, but because it can hold more variables simultaneously and optimise across the full queue rather than working order by order.
The downstream effects are real. Better scheduling means more accurate delivery dates — not the standard fifteen business day estimate we give to every order, but a genuinely calculated timeline based on where that specific order sits in the actual production queue. It means more efficient yarn procurement — ordering stock based on what is actually going to be needed in the next two weeks rather than maintaining a large safety buffer against uncertainty. It means fewer expedited orders, fewer apologies for delays, and fewer situations where a brand has to hold a production run because their labels have not arrived.
This is not the AI story that gets written about in fashion technology publications. It is not visual and it is not dramatic. But it is the part of AI adoption in manufacturing that is actually working right now, delivering measurable results, and changing what brands can expect from a well-run label supplier.
The real bottleneck: design file conversion
If production scheduling is where AI is already delivering value, design file conversion is where the biggest opportunity lies — and where the industry is furthest behind.
Consider what the design conversion step actually requires. A designer takes a brand's artwork and rebuilds it as a weaving specification. They make decisions about colour matching, line weight, type rendering, and pattern density. These decisions follow rules — rules about minimum thread counts for legible text, rules about how many colours can occupy adjacent weave positions without bleeding, rules about which curve geometries reproduce cleanly at different loom densities. These rules are learnable. They are, in the language of machine learning, a classification and optimisation problem with a defined input (brand artwork) and a defined output (production-ready weaving specification).
A trained model — trained on the library of approved weaving specifications that a factory like ours has accumulated over thirty years — could, in principle, take an artwork file and produce a first-pass weaving specification automatically. Not a perfect specification. Not one that eliminates the human designer. But one that reduces the conversion work from hours to minutes, that catches the most common problems before they reach the sample stage, and that gives the human designer a starting point rather than a blank canvas.
The colour matching step is even more tractable. Matching a brand's Pantone reference to the closest available yarn in a factory's thread library is a visual similarity problem. Computer vision systems solve visual similarity problems well. A system that can photograph a yarn sample, read its spectrophotometric properties, and match it against a brand's specified Pantone value — flagging cases where the closest available match falls outside acceptable tolerance before any yarn is loaded onto a loom — would eliminate an entire category of first-sample failure.
Neither of these systems exists yet as a commercial product for the woven label industry. The industry is small enough, and specialised enough, that the major AI platform providers have not built for it. The factories that will benefit most from these tools are the ones that build them — or contribute to building them — based on their own production data.
This is what we are working on at Labeloom. The design conversion problem is where we see the largest single efficiency gain available in our production process, and where solving it well would change the experience of ordering custom woven labels more than any other single improvement.
Quality control: the next frontier
After design conversion, the other area where AI has clear near-term applicability is quality control.
Every label that leaves our factory is inspected before shipping. That inspection is visual — a human examiner checks for weave defects, colour inconsistencies, dimensional accuracy, and edge finish quality. It is effective. It is also slow, and it is limited by human attention span and fatigue in ways that a machine vision system is not.
Automated visual inspection for woven textiles is a solved problem in principle — machine vision systems have been applied to fabric defect detection in broader textile manufacturing for years. Applying it to the specific requirements of woven labels — checking that a logo reproduces accurately against the approved specification, that colour density is consistent across a production run, that edge cuts are clean — is an engineering problem rather than a research problem. The question is not whether it can be done but whether the volume and margin structure of label production justifies the investment.
At sufficient scale, it does. And the data output from automated QC — records of which designs produce consistent results across production runs and which require repeated correction — feeds directly back into the design conversion problem. A system that knows which weaving specifications have historically produced QC failures is a system that can flag similar specifications at the design stage before they reach production.
What this means for brands ordering labels
If you are a brand ordering custom woven labels today, the AI developments described above are not yet visible to you. The process you experience — submitting artwork, waiting for a digital proof, reviewing a sample, requesting changes, and waiting again — has not fundamentally changed from how it worked ten years ago.
What will change, as these tools mature, is the speed and accuracy of the early stages of that process. A first digital proof generated automatically from your artwork file, with colour matches identified and potential reproduction issues flagged before a human designer has even opened the file. A sample approval process where the factory can tell you, before the loom runs, that your design will reproduce to a defined accuracy level based on the weaving specification generated from your file. A production timeline that is calculated from your specific order's position in the actual queue rather than estimated from a standard lead time.
None of this removes the craft from the process. The loom still runs thread by thread. The yarn still has physical properties that no software can override. The judgement calls about what looks right on a label — whether the logo feels balanced at the chosen size, whether the colour depth reads correctly against the label background — still require human eyes and human taste. But the mechanical and computational parts of the process, the parts that currently consume the most time and create the most friction, are precisely the parts that AI is best positioned to improve.
We are at the beginning of that change. The factories that invest in these tools now — and that have the production data to train them effectively — will produce labels faster, more accurately, and at lower cost than those that do not. That gap will become visible to brands within the next few years, in the form of shorter lead times, fewer sample cycles, and more consistent results across reorders.
At Labeloom, we are not waiting for that gap to open before us. We are working to be on the right side of it.
— Gabriele Limonta, Labeloom, Hong Kong
