Lessons · Lesson 2 of 6
What a wrong action actually costs
Break down the cost of one wrong autonomous action, and find out how much of it the undo button can reach.
Lesson 2 of 6 · 21 min
One wrong allocation, followed to the end
On 14 May a delivery from Barsana Mills arrived against a purchase order shared by two Bexholt orders. The receipt had 412 metres on one line. The agent allocated the line to order BX-7714. It belonged to BX-7802.
Nothing about that decision was reckless. Two orders, one purchase order, one receipt line, and a document that does not say whose the cloth is. Course 10.3 shows how that ambiguity arises and why allocation belongs on the receipt. The agent guessed the way the desk usually guesses, and it guessed wrong.
Ask the question everybody asks: is it reversible? In the software, completely. A store clerk opens the receipt line, changes the order code, and saves. Three minutes. Serdana's own figure for that is USD 0.58 of somebody's time.
Here is what it actually cost.
| What had to be undone | Cost | Share |
|---|---|---|
| The record itself — reopen the line, change the code, save | 0.58 | 0.22% |
| The physical position — 412 metres already issued to the wrong cut plan, returned and reissued | 32.20 | 12.38% |
| The work other people did on the wrong number — a short-fall enquiry raised with the mill and withdrawn, a marker re-planned, a loading sequence rebuilt | 186.40 | 71.69% |
| Attention — the meeting, the message to the buyer's merchandiser, and the afternoon spent checking forty other orders nobody had doubted the day before | 40.82 | 15.70% |
| Total | 260.00 | 100% |
The undo button reached 0.22% of it. Everything else was the cost of unwinding work that other people had already done, correctly, on a number that was wrong.
That ratio is the most portable thing in this lesson. When somebody says an action is safe to automate because it can be undone, they have answered about a fifth of one per cent of the question. And they have answered the part that was never in doubt.
Where the cost actually lives
Look at the largest line again. 71.69% of the loss was other people's correct work. Not the machine's mistake, and not the correction, but the completely reasonable things four colleagues did in the hours between the wrong write and its discovery.
That leads somewhere useful. It is the rule this course keeps coming back to:
The cost of a wrong action is set by how many people read the record and what they are entitled to do on it. It is a property of the audience, not of the action.
This is why two actions can be the same size, the same difficulty and the same accuracy, and still differ in cost by a hundred times. A field nobody reads is nearly free to get wrong. A field that seven people and two systems read is expensive to get wrong, in exact proportion to how quickly they read it. Lesson 3 is about measuring exactly that.
The five nobody found
Twelve of the seventeen were noticed. Five were not. Not that week, and not that month. They surfaced only when Ashour ran the six-month replay for this exercise.
The tempting conclusion is that those five were free. They were not. They surfaced at the quarterly count, as part of a fabric variance of 1,847 metres that nobody could explain. The five accounted for 1,204 metres of it. The variance was booked where unexplained fabric variance is always booked: to cutting waste.
1,204 m of misallocated cloth at USD 3.85 a metre USD 4,635.40
per undetected incident USD 927.08
against a DETECTED incident USD 260.00An undetected wrong action costs 3.57 times a detected one, because nothing stops the chain of correct work built on top of it. It also does one more thing that no incident log will ever show you. Booked as waste, those metres raised next season's wastage allowance on that cloth by 0.6 points. Across 214,000 metres of the same fabric, that is 1,284 metres bought and not needed, or USD 4,943.40.
That second figure is deliberately not carried into this course's arithmetic. It is real and it is traceable, but it is a consequence of a consequence. Including it would make every number in lesson 4 depend on a chain of attribution the desk cannot audit. It is named here so you know the boundary arithmetic is conservative, and in which direction.
What C is, for the arithmetic
Serdana ended up with two figures for the same action type, and it matters which one goes into the inequality.
mean of the 12 detected USD 260.00
mean of all 17, undetected ones at their variance cost USD 456.20The desk drew its first boundary in June with 260.00, because that was the number it had. The fabric-variance work landed in September. Lesson 6 redraws the line with 456.20, and one row moves. It is the row nobody was worried about.
Across the nine action types, C ranged from USD 40.00 for a duplicate chasing mail to USD 6,300.00 for releasing an order from credit hold to a buyer who then did not pay. That is a hundred-fold spread, on the same desk, from the same system, while p varies by a factor of six. `C` is the term that decides the boundary, and it is the term nobody measures, because measuring it means costing an incident properly instead of closing it.
Check yourselfA vendor tells you every action their agent takes is fully reversible from the interface. What have they told you?Show the answer
That the record can be put back, which was never the expensive part. At Serdana it was 0.22% of the loss. What matters is who read the record before it was put back and what they were entitled to do on it, how long it took anybody to notice, and whether the correction itself can go wrong. None of those three is a property of the vendor's software. Two are properties of your own process and one is a property of your people's habits. Ask instead for the audit trail that would let you replay a closed period and count what would have happened. That is a real answer to a real question, and it is a question a vendor can help with.
What you own at the end of this lesson
A method for costing one wrong action all the way to the end: record, world, other people's work, attention. The knowledge that the undo reaches almost none of it. A reason never to take an error rate from a complaint log. And a first honest range for C on your own action list.
Next: the test that decides whether an action is reversible at all, which is not a question about the software.