How to Reduce Scrap Rate Using Your Production Data

Take a plastic injection shop running 40 tonnes of raw material a month at a 7 percent scrap rate. That is 2.8 tonnes leaving the building as waste instead of product. Pulling the rate down to 6 percent, a single percentage point, recovers 400 kg a month. At roughly 1.50 dollars a kilo that is about 7,200 dollars a year, and the calculation still ignores the electricity, tool wear and operator hours already spent on those rejected parts.

One point sounds modest. Its effect on the books is not. The real problem is that most shops know their scrap rate roughly, at month end, as a single number. What they do not know is which shift, which mould and which material lot the waste concentrates in. The job of AI here is not to eliminate scrap. It is to make that hidden breakdown visible.

This guide looks at how to reduce scrap rate starting from the data you already have: what records are genuinely required, how three very different AI jobs get confused with each other, what the budget actually looks like, and where to start. If you want the wider picture of which AI applications suit which sector, our AI by industry map covers the landscape.

What should my scrap rate be?

Short answer: there is no published, credible benchmark for manufacturing as a whole. The line you see everywhere, that under 5 percent is acceptable and under 1 percent is excellent, appears almost word for word across dozens of production software vendor blogs. None of them cite a study, a sample or a date. A number appearing in many places is evidence of copying, not evidence of accuracy.

The practical implication is reassuring: you do not need an external benchmark. The only meaningful comparison is your own last month and your own last quarter. How scrap moves for the same mould, the same material and the same shift is the thing worth measuring.

The same applies to sector-level figures. If someone quotes you an average scrap rate for your industry, ask where it comes from. More often than not the trail ends at a marketing page.

Unavoidable scrap versus defect scrap: which one does AI touch?

Scrap is not one thing, it is two. The first is process scrap: trim, flash, sprues, edge loss. It is the physical cost of making the part and it does not fall without a change in design or nesting. The second is defect scrap: out-of-tolerance parts, surface faults, mixed lots, bad settings. AI only touches the second category.

Projects that skip this distinction fail before they start. A target of "cut scrap from 7 percent to 2 percent" is physically impossible if 4 of those 7 points are trim waste. The team works for months, misses the target, and the conclusion becomes "AI did not work for us."

So the correct first move is unglamorous. Pull the last three months of scrap records and sort them into two buckets: waste that the design produces by definition, and parts that should not have existed. The project targets only the second bucket, and the target percentage is set against the size of that bucket alone.

Do I have enough data to start?

Almost certainly yes, and it is messier than you think rather than smaller. In the small and mid-sized plants we work with, production records usually live on paper forms and in spreadsheets rather than in a factory system. That is a starting point, not an obstacle.

Spreadsheets are data. The problem is never the format, it is consistency. If the scrap reason is typed as free text, then "mould broken", "mould related" and "tooling fault" become three separate categories describing one thing, and no model will find a pattern in that. Reducing scrap reasons to a fixed list of 8 to 12 options is worth more than any software you could buy.

The second issue gets discussed far less: incentives. If the operator who reports high scrap gets a hard conversation about it, the data degrades systematically. The quality of your measurement system depends on the person recording it feeling safe writing down the truth. That is a management decision rather than a technical one, and it is the single most important part of the project.

Scrap is the quietest cost line in manufacturing. It never arrives as an invoice, nobody sees it in the monthly report, and it leaks out of the accounts every single month.

Three different AI jobs, three different budgets

When people say "AI" in a scrap context they mean three unrelated things, and the gap between them in data requirements and cost is roughly tenfold. Confusing them leads either to an oversized investment or to expecting miracles from a small one.

• Scrap analysis (cheapest, fastest payback): finding patterns in records you already keep. Which shift, which mould, which material lot, which hour of the day. Three months of clean records is often enough for a first finding. This is more about disciplined data structuring than machine learning.
• Visual inspection (mid to high cost): catching the defect on the line with a camera. A serious model needs a lot of labelled data. In one published automotive case the system was trained on 1.2 million labelled images. The expectation that you show a model a handful of photos and it learns hits a wall here.
• Process optimisation (hardest): linking machine parameters such as temperature, pressure and speed to scrap and recommending settings. It requires a continuous data feed from the machine, so it cannot start on equipment without sensors.

The most common mistake we see is skipping the first step entirely and jumping to the second. A camera catches a defect, it does not stop the defect from occurring. Scrap falls only when detection is joined to root cause data. The camera is an input to that analysis, not a replacement for it.

What does three months of scrap analysis look like in practice?

Consider a metal parts shop running three shifts across twelve machines with about 40 people. Scrap records exist: the operator writes quantity and reason on a paper form at the end of each shift, and someone types it into a spreadsheet monthly. One number comes out at month end, and that is the end of it.

The first task was converting those free-text reasons into a fixed list and rebuilding three months of records as a table broken down by date, shift, machine, operator and material lot. There is no AI in that step at all; it is about two weeks of structuring work. Then the picture emerged: 60 percent of the scrap sat on three machines, most of that on the night shift, and the night shift peak clustered around lots from one particular supplier.

The resulting action was not a software purchase. Incoming lots from that supplier started getting measured at goods-in, and an extra setup check was added on the night shift. A vision camera only entered the conversation afterwards, for one station and one defect type. That order matters: find out which defect costs money first, then buy the hardware that catches it.

What does AI visual inspection cost?

For a realistic range: an entry-level system covering one station and one defect type, using a standard 2D camera and a compact edge computer, sits in the low thousands of dollars for hardware. These figures come from vendor price lists, though independent suppliers quote overlapping bands. Once 3D inspection or robot guidance enters the scope, the range climbs quickly into the tens of thousands.

Hardware is not the main budget line, though. Installation, conveyor integration, connection to your existing systems and above all the labour of labelling images make up the bulk of the total. Figures circulating online in the range of 110,000 to 200,000 dollars per line come from automotive and medical device scale; presenting that band to a small manufacturer is both misleading and discouraging.

Treat claims like "AI cuts scrap by up to 50 percent" or "the investment pays for itself in 6 to 12 months" the same way. These are sales arguments rather than industry findings, and there is no sample behind them. Run your own payback calculation with your own numbers, the way we did at the top of this article. Without knowing what one point of scrap is worth to you, no quote can be evaluated.

How do you build the business case internally?

Start with the number that survives scrutiny: the cost of one percentage point. Raw material price times monthly volume gives you a defensible floor. Add the conversion cost of scrapped parts, which is usually machine time and labour, and you have a range rather than a single figure. A range you can defend beats a precise number you cannot.

Then be explicit about what the project does and does not promise. A scrap analysis project promises visibility into where waste concentrates. It does not promise a specific percentage reduction, because the reduction depends on actions taken afterwards, and those actions may be commercial rather than technical, such as changing a supplier.

Finally, set a review point. Three months in, either the analysis produced actions or it did not. If not, the usual cause is not the technology but the data, and that is worth naming early instead of quietly extending the project.

When should you not try to reduce scrap rate at all?

Vendors will not ask this question, so ask it yourself. The right moment for a scrap project is when scrap is measurable and production is reasonably stable. In the following situations, postponing is the better call.

• Your product mix changes constantly. A shop making different parts with different moulds and materials every month never accumulates the repetition needed for a pattern. Simplifying the product groups comes first.
• No scrap records exist at all. Historical data cannot be manufactured retroactively. This becomes a record-keeping project lasting at least three months, and that should be said out loud rather than sold as analysis.
• There is a known, unfixed technical problem. If everyone already knows the tooling needs replacing, no analysis will improve on that knowledge. Do the known thing first; analysis is for finding the unknown.
• Operators are under pressure about scrap numbers. If reporting waste invites blame, the records come in systematically incomplete and the project produces confidently wrong conclusions.

Frequently asked questions

How many months of data are enough?

For scrap analysis, three months of clean records usually produces a first set of findings; capturing seasonal effects takes a year. Visual inspection models work on a completely different unit of measurement, where what matters is the number of labelled images rather than the number of months.

I run a small shop. Is this for me?

The analysis side, yes. If you have a spreadsheet you can begin, and the cost is essentially effort. Cameras and process optimisation depend on your equipment and volume; where a single station produces a few hundred parts a day, the return on vision inspection is weak.

Does it have to connect to my ERP?

Not at the start. The first analysis works fine on an exported table. Integration earns its place once the findings prove useful and the process becomes permanent. Projects that run this in reverse tend to stall in integration and never reach a result.

Scrap dropped. How do I know what caused it?

By not changing five things at once. The classic mistake is seeing the analysis and implementing every action immediately: scrap falls, and nobody knows which change did it. Sequence the changes and give each at least two weeks before judging it.

So what should you actually do?

• Reduce scrap reasons to a fixed list. Eight to twelve options, no free text. This costs nothing and everything else depends on it.
• Separate the two buckets. Do not set a target before splitting unavoidable process scrap from defect scrap, and set the target against the second bucket only.
• Rebuild three months of records as a breakdown. Date, shift, machine, operator, material lot. Be able to produce that table before buying anything.
• Analysis before hardware. Do not buy a camera before knowing which defect costs money, and when you do, start with one station and one defect type.
• Make it safe to report scrap honestly. If high scrap reports are punished, your data is corrupted at the source and no model compensates for that.

Making that leak visible does not require building a new factory. The paper forms already on your shop floor, once given a proper structure, can tell you where to look within the first month. If you want to talk through where your production data could take you, get in touch and bring a sample of those records. For a closely related starting point on the equipment side, our starter guide to predictive maintenance and AI quality control walks through the same logic from the machine's perspective.