Track scrap rate as units scrapped divided by total units started, times 100, counted the same way every week. For free-text reject notes, give AI the pile and ask it to propose six to twelve reason codes, then have operators correct them. Each week, count rejects per code and list the biggest first. That list is your Pareto.
The formula is the easy part; the reasons are where the value sits. A single scrap percentage tells you there is a problem and nothing about its cause. Free-text notes such as "chipped", "chip on edge", "tear out at the router" and "bad edge" are four phrases for one fault, and counting them separately hides the biggest cause under four small bars. AI is good at that grouping job, provided people who run the machines confirm the result.
Decide what counts as one unit and one scrap
Before any grouping, fix the definition. Most arguments about scrap figures are really arguments about what was counted.
Scrap rate = scrapped units / total units started x 100. Some workshops divide by units finished instead, which makes the rate look a little higher; both are valid if you never switch. Pick one and write it at the top of the sheet. The same goes for the unit: a door, a batch, a metre of extrusion, a kilogram of material. For processes where a fault ruins a whole batch (a bad mix, a spoiled dye lot), count in batches and also record units or weight lost.
Three more decisions help:
- Rework is not scrap. A door sanded and re-lacquered and then passed is rework. Keep a separate tally, because rework costs labour even when the part survives.
- Count at the point of rejection. Record the scrap when it goes into the bin, with the station where it was found. A defect found at final inspection may have been made three stations earlier, which is worth a second column: "where found" and "where made, if known".
- Weigh or count, not both loosely. If your material is sold by weight (resin, metal offcuts), use a weight-based rate and note the unit.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowSet up the sheet before touching AI
A scrap log needs only a few columns. The row is one reject or one small group of identical rejects.
| Column | Example | Why it is there |
|---|---|---|
| Date | 12/11 | Weekly grouping and trends |
| Shift or operator initials | B | Only to ask questions, never to blame |
| Product or job | Door D-45, oak | Different products scrap differently |
| Station where found | Spray booth | Shows where detection happens |
| Quantity scrapped | 2 | Feeds the formula |
| Free-text reason | "runs in lacquer, humid" | The raw note, never overwritten |
| Reason code | FIN-RUN | Filled in after grouping |
| Rework or scrap | Scrap | Keeps the two apart |
| Cost per unit | $22 | Optional, lets you price the Pareto |
Keep the free-text note exactly as written, in its own column, forever. The code column is an interpretation; the note is the evidence, and you will want to re-group it when you learn something new. Also record the total units started each day in a separate tab or sheet so the denominator is not guessed at the end of the week.
If the shop floor currently records scrap on scraps of paper, start with a clipboard at each station and a pen on a string. Ask for a tick and a few words. The aim is that writing a reason takes under ten seconds; if it takes longer, operators stop doing it.
A main example: one week in a cabinet-door workshop
This is an illustration with invented figures. A furniture maker's workshop produces flat-pack kitchen cabinet doors in oak. Last week they started 1,250 doors. Operators scrapped 87 of them, and rewrote each reason on the clipboard in their own words. The scrap rate is 87 / 1,250 x 100 = 6.96 per cent, which they round to 7.0 per cent. At roughly $22 of material and labour per door, that is about $1,914 thrown away in a week.
The owner types the 87 notes into the sheet. They look like this, a sample of fifteen:
runs in lacquer
dust in finish
chip on edge
tear out router
wrong size - cut short 4mm
bow in board
orange peel
joint gap at corner
edge chipped
cut under size
warped, came in like it
dust nibs
router tearout
gap in mitre
bad lacquer, humid dayTo a person scanning this list it already looks like five or six faults described in fifteen ways. Counting it by hand means reading eighty-seven lines and keeping a tally in your head. This is where AI earns its keep.
Ask AI to propose codes, not to decide them
The first pass is to let the model suggest a small set of reason codes, given the real notes. Use a prompt like this.
Below are free-text notes from a workshop's scrap log, one per line. Group them into between 6 and 10 reason categories.
For each category give: a short code (3 letters, a dash, 3 letters), a plain name, a one-line definition, and the list of notes that belong to it with a count.
Rules: do not invent notes; every note goes in exactly one category; if a note is ambiguous, put it in "Unclear" and list it separately; keep Other below 10 per cent.
Use the workshop's own words where possible.
[paste the notes]A realistic output (illustrative) for the full 87 notes:
| Code | Plain name | Count |
|---|---|---|
| FIN-DEF | Finish defects: runs, dust, orange peel, humid-day lacquer | 31 |
| EDG-CHP | Chipped edges and router tear-out | 22 |
| CUT-SIZ | Cut to wrong size | 14 |
| MAT-WRP | Warped or bowed board as received | 9 |
| JNT-GAP | Joint or mitre gap | 7 |
| UNC | Unclear | 4 |
The counts add up to 87, which is the first thing to check: if a model's category counts do not sum to the number of notes you gave it, it dropped or doubled something. Always add them up yourself, or ask for the total and compare. Models are unreliable at counting long lists, so a second check matters.
The second check is whether the groups make sense to the people who made the scrap. That is the next step.
Check the codes with the operators, in ten minutes
Pin the proposed table to the wall or show it on a phone and walk the shop floor. Ask each operator three questions: does this list match what you throw away, is anything in the wrong group, and what is missing? In the example workshop, the spray-booth operator points out that "humid day lacquer" and "runs" are different problems: one is weather and the other is gun technique, with different fixes. The saw operator says the 14 wrong-size doors were all from one jig that slipped on Tuesday. And the joiner says the warped board is a supplier problem, not a workshop one, so it should stay separate and be easy to show to the supplier.
Update the list: split FIN-DEF into FIN-RUN (runs and sags, 12) and FIN-HUM (humidity, lacquer blushing or orange peel, 8), leaving FIN-DUS for dust nibs (11), which is 31 again. A split like that is the reason to run the check. A model can see words; the operator knows the cause.
Keep two rules for this meeting. The codes describe the fault, not the person. And the list is allowed to change, but only with a note saying when and why, so that old and new weeks can still be compared. If one code turns out to be a catch-all that nobody understands, rename it before it grows.
The weekly Pareto, built from the corrected codes
A Pareto list sorts causes from the biggest to the smallest and shows the cumulative share. The idea is that a few causes usually account for most of the loss, so you spend effort on those. It is a sorted table before it is a chart, and the table is enough.
| Rank | Code | Scrapped | Share | Cumulative | Approx. cost at $22 |
|---|---|---|---|---|---|
| 1 | EDG-CHP | 22 | 25.3% | 25.3% | $484 |
| 2 | CUT-SIZ | 14 | 16.1% | 41.4% | $308 |
| 3 | FIN-RUN | 12 | 13.8% | 55.2% | $264 |
| 4 | FIN-DUS | 11 | 12.6% | 67.8% | $242 |
| 5 | MAT-WRP | 9 | 10.3% | 78.2% | $198 |
| 6 | FIN-HUM | 8 | 9.2% | 87.4% | $176 |
| 7 | JNT-GAP | 7 | 8.0% | 95.4% | $154 |
| 8 | UNC | 4 | 4.6% | 100.0% | $88 |
Splitting the finish category changed the picture. Before the split, finish defects looked like the top cause at 31 doors and 35.6 per cent. After it, chipped edges are the biggest single code, and three finish codes need three different remedies. This is the advantage of a Pareto built from operator-checked codes rather than the first grouping. The cost column assumes every scrapped door cost the same; if cost varies, use each row's own figure.
Let the spreadsheet do the sums: a pivot table or a COUNTIF per code, then a sort, and a running total. Language models can be wrong on counts; the guide to cleaning messy data shows spreadsheet-side methods that do not depend on them.
Turn the top item into one action, then watch the next week
The Pareto is useful only if it ends in something somebody does. Take the top two or three codes and ask, with the operator concerned, what one change might cut them. For EDG-CHP: new router bit, slower feed, a backing board on the exit cut. Try one change, note the date on the sheet, and look at the next two weeks of data.
Say the workshop changes the router bit on Monday and the following week records 12 chipped edges in 1,300 doors started, instead of 22 in 1,250. The edge chip rate per hundred doors went from 1.76 to 0.92. Because the codes did not change between weeks, the comparison is fair. If the total scrap rate rises because humidity spiked, the code-level view still shows that edge chips fell.
A weekly rhythm that suits a small shop:
- Friday, ten minutes: enter the week's clipboard notes and units started.
- Friday, five minutes: AI proposes a code for each new note, using the existing code list only.
- Friday, ten minutes: a supervisor reads the proposed codes against the notes and fixes errors; "Unclear" items are asked about on Monday.
- Monday, five minutes at the start-of-week meeting: show the top three codes, last week's change and the result.
For the weekly coding step, use a prompt that restricts the model to your list.
Here is our current reason-code list (code, name, definition). For each note below, assign exactly one code from this list. Do not create new codes. If none fits, answer UNC and say why in five words. Return a table: note, code, and your confidence (high, medium, low).
[code list]
[this week's notes]Review every "low" and every UNC line; trust the "high" ones after you have spot-checked a sample of ten. If more than a tenth come back UNC for three weeks running, it is time for a new code.
A different shop: a garage rejecting parts, and a clothing line
The method does not belong to woodwork. A car repair garage that refurbishes alloy wheels can count wheels rejected after the powder-coat stage with reasons like "orange peel", "fisheye", "pit in casting" and "bent, not recoverable". After grouping, the garage may find that a third of the rejects are casting pits found only after stripping, which is not a coating fault at all, and the fix is a better incoming check rather than a different oven setting.
An online clothing store with a small in-house sewing room can code rejects as "needle holes", "uneven hem", "wrong colour thread", "stain from handling" and "fabric flaw". The same list shows whether the problem lives in sewing or in the fabric that arrived. If it is the latter, the supplier sample scorecard in this tutorial on sample requests is the matching tool for scoring the material before it is cut.
When a note names two causes, or none
Real notes are rarely tidy. Three kinds cause most of the trouble, and each needs a rule written down once so everyone codes the same way.
Two causes in one note. "Chipped and dust in finish" is two faults on one door. The rule that works best is to code the first fault in the process order, because the earlier fault is the one that made the door a reject, and to put the second in a "also" column if you want it. Do not log one door twice, or your scrap count will exceed the doors thrown away and the rate will be wrong. Tell the model the rule in the prompt: "if a note names more than one fault, choose the one that happened earliest in this process order: cut, joint, route, sand, finish".
A note with no cause. "Rejected by [first name]" or "not good" cannot be coded. Let it fall to Unclear, and on Monday ask what was wrong while the door is still in the bin. Over a month the number of these tells you how well the clipboard habit is holding.
A cause outside the workshop. A board that arrived warped was never the workshop's fault, but it still costs money. Keep it as a code, because the Pareto should show total loss, and because it gives you a number to put in front of the supplier.
Weighing the rate against a target you can defend
Once you have a month of weekly figures, you can set a target. A simple way is to take the average of the last four weeks, say 7.0, 6.4, 7.8 and 6.9 per cent, which averages 7.0 per cent, and aim for a modest fall, such as 6 per cent in a quarter. Ask what the fall is worth: on 1,250 doors a week at $22 each, one percentage point is 12.5 doors, or $275 a week, which is roughly $14,300 a year. That figure is what you compare with the cost of a new router bit, a jig or an hour of training.
The calculation is also a way to avoid chasing the wrong problem. If the biggest code costs $484 a week and the fix costs $2,000, it pays back in about four weeks; if the fix removes only a third of the loss (about $161 a week), it takes about twelve and a half weeks. Put both numbers in the decision, and remember that the measured fall in the sheet, not the promise, tells you whether the change worked.
Keeping the data honest
Numbers about waste have a way of improving when people are afraid of them. Three habits protect your data.
- Never use the log to blame. If the initials column is used in a conversation about disciplining someone, operators will stop writing things down or begin writing "Other". Use it to ask who knows about a pattern, not who caused it.
- Reward the unflattering entry. A week with a high scrap rate and clear notes is better than a low rate and empty clipboards. Say so out loud.
- Audit with a bin check. Once a month, count what is physically in the scrap bin when the shift closes and compare it with the log. A gap of more than a few per cent means some scrap goes unrecorded.
The same warning applies to AI summaries: a clean grouping can look authoritative while hiding a miscounted line. Read the tutorial on checking AI work for a routine that fits here, and keep the raw notes column so you can always re-run the grouping.
When scrap numbers belong in a daily report or a quote
Once the weekly list works, two things follow naturally. One is a daily summary of counts, scrap and stops for the owner, covered in the daily production report tutorial. The other is pricing: if your scrap rate on a product is 7 per cent, a quote that assumes none underprices the job by about the same amount. A tutorial such as the injection moulding quote check shows where an allowance for scrap enters an estimate, and you can use your measured figure rather than a guess.
For high volumes or very fine defects, a camera system may beat human eyes at catching problems early; the visual inspection cost tutorial sets out what that costs a small manufacturer, so you can compare it with the dollars on your Pareto. If your weekly scrap bill is $1,900 and one new system would remove a quarter of it, you can see the break-even without a sales pitch.
Checking that tracking itself is working
After four weeks, ask five questions of your own system. Are the units started recorded daily, not reconstructed? Is Other or Unclear below ten per cent? Do two different people, given the same ten notes, choose the same codes at least eight times? Does the bin check agree with the log? Did at least one action come out of the Pareto, with a date and a result?
If the answers are mostly yes, you have a process that survives holidays and staff changes, and a list you can show a supplier, an insurer or a buyer. If not, fix the weakest answer first, usually the daily count, because every other figure divides by it.
Scrap tracking questions from small workshops
What is a normal scrap rate?
There is no single normal figure: it depends on the process, material and how much rework you count as scrap. A rate that is steady week to week and falling matters more than matching somebody else's number. Measure your own for four to six weeks before setting a target.
Should rework count as scrap?
Track them separately. Scrap is a unit thrown away; rework is a unit fixed and passed, which still costs labour. If you merge them, a fall in scrap may just mean more rework. Add a second code column or a separate tally for reworked units.
How many reason codes should I have?
Between six and twelve. Fewer hides detail; more means operators cannot remember them and start choosing Other. Keep Other under about ten per cent of rejects, and review the notes inside it each month for a new code.
Can AI read handwritten shop notes?
It can read clear handwriting from a photo, but misreads are common with smudged or shorthand notes. Type or dictate the notes first where you can, or check every transcribed line against the paper. Do not upload anything with customer or personal details.
Do I need special software?
No. A spreadsheet with one row per reject and a few columns is enough to start. Move to your production software's own scrap field when the spreadsheet becomes the bottleneck, and keep the same reason codes so the history carries over.
Further reads
- AI for Coffee Roasters: Batch Logs, Wholesale Orders, Labels — Batch logs in another small-production setting.
- AI Quality Control for Small Manufacturers: Where to Start — Wider options for quality checks on a small line.
- AI Supplier Management: Track Prices, Lead Times, and Risk — Take warped or faulty incoming material up with the supplier.
- How to Compare Supplier Quotes Side by Side With AI — Compare replacement-material quotes side by side.
- Why Is AI Bad at Maths? What to Check in Quotes and Invoices — Why counts and rates belong in a spreadsheet formula.
- AI Bias in Small Business Decisions: Hiring, Pricing and Credit — Why operator-blaming codes skew the data.
- Supplier Audit Checklist: Plan and Write Up Visits With AI — A small maker's audit of a powder-coating supplier: pre-visit questions from the NCR log, a 3-hour agenda, an on-site checklist and a findings report.
- Supplier Corrective Action Request: Draft a SCAR With AI — A SCAR built around one failed bracket lot: evidence pack, containment ask, root-cause questions, due dates and a closure test the supplier can meet.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: standard quality-management practice for scrap rate and Pareto analysis; no vendor pricing is quoted in this tutorial.