A by-product, or a tax

Measure the extra second, not the whole task

The business case for Line 4 said a scan takes about three seconds. Ramez timed it, sixty observations, and found something more useful than a number.

A scan does not take three seconds. A bundle hand-off takes 3.2 seconds without a terminal and 11.4 seconds with one. The operator has to put the last piece down, turn, wake the screen, present the ticket, wait for the beep, and turn back. So the record adds 8.2 seconds. That is the only figure that means anything, because the first 3.2 seconds were being spent either way.

Time the difference, never the task. This is the first place a capture estimate goes wrong, and it goes wrong by a factor of nearly four.

The arithmetic everybody does

Line 4 makes 41 bundles a day across nineteen operations: 779 records.

records a day                                     779
extra seconds each                                8.2
seconds a day                                 6,387.8
minutes a day                                  106.46
share of 20,160 attended minutes                0.53%
at USD 0.0295 a loaded operator minute      USD 3.14 a day
eleven lines, 250 days                    USD 8,637 a year

A loaded minute is the wage plus everything the factory pays on top of it. Half a per cent of the floor's time, and less than nine thousand dollars a year, against a plant turning over far more. Every capture business case in the industry contains this calculation. It is the wrong one.

Why it is the wrong one

Operators are paid for attended time. They are paid the same whether the minute is spent sewing or spent at a terminal. So a scan taken by an operator with slack in her cycle costs the factory nothing at all. Not USD 0.0295. Nothing. The wage was already committed.

A scan taken at the station that sets the line's speed is different. It costs the whole line's output for those seconds, because there is no slack there to absorb it. That is what a constraint is: the one station the whole line waits for.

Course 5.3 covers how a line is balanced and how the constraint station is found. Take it as given here that on HD-745 the constraint is the sleeve-attach station, and that Line 4's output is set by it.

bundles a day at the constraint                    41
extra seconds each                                8.2
minutes a day lost at the constraint            5.603
constraint minutes a piece  (480 / 820)        0.5854
pieces a day not made                           9.572
share of the line                               1.17%
at USD 0.94 contribution                     USD 9.00 a day
eleven lines, 250 days                    USD 24,745 a year

USD 24,745 against USD 8,637. The honest figure is 2.9 times the one in the business case. The two calculations do not disagree about seconds. Both use the same 8.2 seconds. They price different minutes. One prices an hour of labour that was already bought. The other prices an hour of capacity that was not.

The record fails exactly where you need it

Ramez checked completeness the only way it can honestly be checked: against the payroll coupons, which are counted for a different purpose by a different person. Records made, over records that should have been made.

Twenty-six points of the record disappear in the hour after a breakdown. Read that as a sentence rather than as a table: the real-time system is at its blindest in the hour it exists to see. When the line is catching up, the scan is the first thing dropped, because it is the only part of the job that makes no garments.

A screen reading that hour will report the recovery as the worst hour of the day. It was the best one.

The ranking that inverts

Two operators work the same operation on opposite sides of Line 4. Over twelve days:

The system says Basma is ahead of Faten by 12.0%. The bundles on the floor say Faten is ahead of Basma by 7.1%.

The reason is not dishonesty and it is not laziness. Faten is faster. So Faten finishes more bundles. So Faten has more scans to make. So Faten has more scans to drop when she is pressed. The penalty for not recording falls hardest on the operator who has the most to record. The measurement is biased against speed by the way it is built.

Say the general form out loud, because it is the point of the whole lesson: a system that asks an operator to stop and record is measuring the operators who comply, not the line.

Then they paid for it, and got exactly what they paid for

Elmarj pays an incentive of USD 0.021 for each piece above the daily standard. From month two it paid that on the recorded figure rather than on the supervisor's count.

Faten's 1,104 unrecorded pieces over twelve days are worth USD 23.18. Scaled to a month that is USD 48.30, against a wage of USD 268. So 18.0% of her pay now depended on whether she stopped to scan.

She stopped to scan. Everyone did. Over six weeks completeness rose from 91.4% to 99.1%, and Line 4's output fell from 820 pieces a day to 810.

That is a fall of 1.2%. The constraint arithmetic in this lesson predicted 9.572 pieces, which is 1.17%. The wage arithmetic in the business case predicted no fall at all, because it never priced a minute of capacity.

The two calculations were not opinions. One of them predicted the future and the other did not.

The fix is not a faster scanner

A faster terminal takes the 8.2 seconds down to perhaps 5. It does not change which minute is being spent.

Here is what did change it. At the sleeve-attach station the record is now made by the helper who feeds the station, at the moment she puts the next bundle down. That hand-off was already happening, and she has slack. The operator does nothing. Capture at the constraint went from 5.603 minutes a day to zero, and completeness at that station went to 100%, because the helper is not the one under pressure.

The record did not get cheaper. It moved onto somebody whose minutes were not the ones being counted. That is what "a by-product of the work" means in practice. Wherever you cannot move it, you are buying data with output, and you should say so in the paper.

Act as an industrial engineer who has installed shop-floor data capture in garment factories and has seen the business case go wrong in the same way every time. Help me price a proposed capture scheme honestly. Facts about my line: operators [NUMBER], attended minutes a day [NUMBER], output a day [PIECES], bundle size [PIECES], sewing operations [NUMBER], the station that sets the line speed [WHICH], lines in the plant [NUMBER], working days a year [NUMBER], contribution a piece [AMOUNT], full cost of an attended operator minute, wage plus everything the factory pays on top [AMOUNT]. The proposed record is [DESCRIBE: what event, at what unit, who makes it, on what device]. Do the following. First, tell me exactly what to time on my own floor to get the EXTRA cost of the record, which is the difference between the hand-off with the record and the same hand-off without it, and warn me about the ways that measurement usually goes wrong. Second, compute the cost two ways: the wage-rate way that prices every operator minute, and the constraint way that prices only the records made at the station setting the line speed, at contribution. Give both, per day, per line-year and per plant-year, and tell me the ratio between them. Third, say plainly which of the two is the number I should put in the paper, and why the other one is not worthless but is answering a different question. Fourth, list every place in my scheme where the record could be moved onto somebody with slack: a feeder, a helper, a quality checker, a fixed reader the garment passes anyway. For each one, say what it would cost the constraint after the move. Fifth, tell me what completeness figure I should demand, what it must be measured against, and why the system's own count of its own records is not an acceptable comparison. Do not give me a range where a number is possible, and list every assumption you made at the end.