profile
Door
More view more view
Recording the apparel supply chain
through shop-floor numbers and data

How Far to Trust an AI Draft: Five Lines You Check Yourself

AI Comment card showing Needs Review 2 rows and Front · Assembly (Tool missing) (example data)
The summary card marks the rows the draft could not fill (example data)

AI Process Analysis series 6/6 · (1) Process analysis is cost design · (2) How to read an operation breakdown · (3) What Man Power 0.5 means · (4) Why the first week loses money · (5) When the lowest unit price is not the optimal factory · (6) How far to trust an AI draft

Key takeaways

An AI draft marks the lines it could not fill, so the work goes to the rows that passed through unmarked. Five lines get checked by hand: operation count and missing work, machines and tooling, the two numbers behind Takt, the factory constants, and the pricing policy. Transcription, calculation and gap flagging get no second look. The test is not accuracy but whether you can walk back from a confirmed price to its basis.

The first time I watched 48 operations fill the screen a few seconds after a button press, I was pleased. A moment later I was uneasy.

Less because a table that used to take me days had appeared at once, and more because of the question underneath it: how would I know the table was right? Miss a wrong line and the error travels through the price and into the contract.

The last post in this series is my answer to that. To put the conclusion first, there is no need to review all of it. Five lines will do. Every screen and figure is example data.

The draft marks the rows it could not fill

When the operation breakdown is generated on the example screen, a summary card comes with it: Process 48 ea (OP 42 · HP 6), SMV 1,505 sec, Net Time 1,309 sec. And one more line under that.

Needs Review 2 rows, Front · Assembly (Tool missing), followed by a comment.

“Assign a tool to No. 18 and No. 44, then run line balancing.”

Forty-eight lines are written, and two of them are marked as unfilled, with the positions given: No. 18 and No. 44, the tooling column.

That is why I came to trust this way of working. A draft that fills every blank plausibly leaves nothing to review against. A draft with its gaps marked can be reviewed. The flagged rows are already half resolved, and the work goes to the rows that passed through unmarked.

Checklist of the five lines you check yourself on an AI process analysis draft
There is no second look at the rest. These five lines get checked by hand

The first line. Operation count and missing work

The breakdown on the example screen is grouped by part, as in Back (15) and Front (16).

What matters here is not the size of the number but the comparison against the tech pack. Some details sit in the tech pack but barely show in the sketches: lining treatment, label positions, special sewing instructions. The remarks field on the example order carries one of these, “Left shoulder seam with silicone tape, analysis follows W.I.”, which means the work instruction has to be read separately.

Lines like that drop out of a draft easily. An operation that is missing from the table is missing from the SMV, from the headcount and from the price. And on the floor, somebody always does that work anyway. Work that is not in the cost happens on the line.

Compare the operation count per part against the detail in the tech pack yourself, and confirm that the special operations named in the remarks and the work instruction made it into the table.

The second line. Machines and tooling

The table has a Machine column and a Tool column. Machines appear as SN, OD, OL4, OVL3, DN, IRN and MAN; tooling as Patch Limit, Pattern and Scissor.

This is where the table meets the factory. A draft proposes the machine normally used for this style, but whether that factory owns it is a separate question. If there are not enough units, two lines end up sharing one machine, which creates waiting time that is nowhere in the table.

Tooling is more sensitive still. As the third post showed, one attachment can be the way a bottleneck gets broken. Run a schedule built on that attachment without it, and the operation is over the line from the start.

Check the machine column against the factory’s equipment list, pull the operations that need tooling and confirm the factory can prepare it. The rows AI flagged as Tool missing get cleared here as well.

The third line. The two numbers behind Takt

The second post made the point that Takt does not come out of the operations. It is worked back from the delivery date. The same reason makes it a separate item at review.

The example screen shows Takt Time 48 sec, from 28,800 sec divided by Estimated Target 603 pcs. The 28,800 seconds are an assumption about the hours available in a day, and the 603 pcs is the daily target.

Neither is a result of process analysis. They are premises. When working conditions or working days change, they have to change with them. Premises rarely do change, because whatever was in the template comes along with it.

Confirm that the numerator matches the working conditions of this order and that the denominator comes from this order’s quantity and working days. If Takt moves, the whole operator assignment has to be recalculated.

The fourth line. Factory constants

The CM basis on the example screen carries the values that describe the factory’s condition: Loss Time 15.00 %, Factory Rate 70.00 %, Working Day 7 days, Manpower 45 ppl / 48 ea.

Style attributes rebuilt every order compared with factory attributes that sit untouched (example data)
Style attributes and factory attributes are refreshed on different cycles (example data)

These belong to the factory, not the style, so once they are registered they tend to sit untouched for a long time. The problem is that factories change. If the share of new operators has risen, if equipment was replaced, if a line was added last year, the numbers should have moved.

The Factory Rate in particular feeds straight into the price. Quote this year on last year’s figure and you are quoting last year’s factory.

Check the last revision date and the basis for each constant, and compare them against results from recently completed orders of the same kind.

The fifth line. Pricing policy

The fifth is not a technical item. It is a management call.

The price section of the example screen reads CM Price $3.28, New Standard × 0.9 (-10%), Final CM $2.95. What process analysis produced gets the company policy applied once, and that is the confirmed figure.

Reviewing this line does not mean checking the arithmetic. It means confirming whose decision that factor was and whether it still holds for this order. As the fourth post showed, whatever comes off the price has to come from somewhere: the learning curve, the headcount, or the factory quietly absorbing it.

Confirm the conditions the factor applies under, and if they do not match this order, confirm the price without it.

What no longer needs reviewing

There is an other side to this. These are the items I no longer look at twice.

What AI handles as draft work compared with what people handle as judgment
What changes is where the review time goes

Transcription. Quantity, delivery cargo date and fabric details move from the tech pack into the fields without a person in between. There is nothing to mistype.

Calculation. SMV totals, headcount totals and the price come out of the same formula every time.

Gap flagging. Whatever is missing gets flagged first. There is nothing to hunt for by eye.

Most of my review time used to go to those three. It now goes to the five lines above. The hours are the same and the content of the review is not.

You have to be able to trace it back

Whether an AI draft can be trusted is not settled by what percentage of it is right. Tables people build are not 100% right either. The test is whether you can trace where it went wrong when it does.

The path back from a confirmed CM through capacity verdict, factory simulation, line balancing and the operation breakdown (example data)
You have to be able to walk back from a confirmed price to its basis (example data)

The CM row on the example cost sheet carries a CM Source next to it, reading AI Process Analysis · Confirmed. It records which process analysis the price came from and who confirmed it. Months later, when a claim arrives, the operation breakdown, the balancing result, the factory constants and the capacity verdict for that order can be opened in order.

A price that came out of a spreadsheet on somebody’s laptop cannot be walked back. When that person leaves, the basis leaves with them. That worries me more than accuracy does.

Closing the series

Across six posts I seem to have been making the same point. How to read the table, how to assign operators, how to put the curve into a quote, the order for selecting a factory, and how to review a draft. All of them were about what a person has to judge.

When writing the draft shrinks to seconds, what is left is judgment. And judgment still belongs to someone who knows the floor. I do not think that is a change for the worse. It is the change that turns the person who used to write process analysis into the person who designs cost.

Line to check by handWhat you are looking atHow to check
Operation countOperations in the tech pack but not in the tableCompare counts by part against the tech pack
Machines and toolingWhether the factory owns the equipmentCheck against the equipment list
TaktAssumptions on hours and daily targetConfirm the source of both numbers
Factory constantsLoss Time, Factory Rate, Working DayLast revision date against recent results
Pricing policyConditions the New Standard factor applies underWho decided it and whether it holds

Previous: When the Lowest Unit Price Is Not the Optimal Factory: Checking the Line Schedule First

To see Monolis AI process analysis for yourself, request a demo on the SIJE website.

back icon Back