Chapter 2 · The people in the room
Four people and a number nobody could explain
Days sales outstanding at Ostermill had crept up nine days over the previous year.
Four people decided what to do about that. Within a year the overdue-invoice letters were being drafted by something that chose which accounts to write to and which to hand back to a person, and the whole of Part Two is how that happened. Meet them first, because the project was shaped more by who was in the room than by anything in the technology.
Marcus Ellery is the controller, and the nine days bothered him less than the fact that he could not say why. He had the aging report, the collections figures, the write-offs, and four plausible explanations that each accounted for some of it. He has no objection to spending money on a problem. He has a standing objection to spending money on a version of a problem that nobody can state in a sentence.
When the AI is eventually wrong about money, and around here that is what everybody calls it, it will be wrong about his number. He knows that on the day the project starts, which is why he behaves better than most people in his seat.
Tom Brindle runs sales, and he reads the same aging report as a list of people he will have to call afterward. Every reminder Ostermill sends is a small message to a customer that the relationship is currently about money. His account managers had been complaining for a year that finance was antagonizing accounts they were mid-negotiation with, which was occasionally true and always remembered.
Tom is the reason this project exists. He is also, five months in, the person most impatient with it, and both come from the same place. He is protecting an asset that appears on no report.
Dana Okafor had been at Ostermill seven months. She came from a distributor twice the size, where she had watched two AI projects go the same way, and what she took from both was not technical. It was that the decision gets made on the afternoon somebody shows a working demo, and that every document produced after that afternoon is ratification.
Her instinct, when the subject came up, was to slow it down and write things first. That instinct is the argument of this book. It looked, to at least two people in the room, like obstruction.
Priya Nair is the architect. Her contribution to the entire build reduces to one question she asks in different words at least six times over the following year: you have written that it will never contact an account in dispute, so what stops it. Not what discourages it. What stops it.
Priya has spent a career learning which sentences in documents survive the trip into a system and which evaporate on contact. Her patience for the second kind ran out around 2019.
Those four and Ruth are the people who decide. They are not the people this is about.
The people who are never in the room
There is a retired machinist in Springfield with a disputed invoice from a job he finished in April. There is a family shop in Zanesville where the owner died in the summer and his daughter is opening the mail while she works out what to do with the business. There is the Dayton machine shop from the first pages of this book, which pays in the last week of every quarter and has never once been late by its own reckoning, because in its own reckoning the quarter is the unit.
None of them will ever see a dashboard. None of them agreed to any of this.
This book calls them the affected people, and it does not call them users, for two reasons that are worth separating. They never opted in, so no consent was given and none can be withdrawn. And they cannot tell the difference between a policy and a mistake. When a reminder arrives five days after a death, the daughter opening the mail does not experience a retrieval failure or a gap in an escalation rule. She experiences a company that sent it.
They recur in almost every chapter of this book. That is deliberate, and it is the only method I know for keeping an error rate attached to a face.
There is a habit in this industry of describing the third group as an edge case, and the word is doing work that should be examined. An edge case is rare from the system's point of view. From the point of view of the machinist in Springfield, the case is not rare. It is the only interaction he has ever had with Ostermill's receivables process, it is the one he will describe to other people in his trade, and it is a hundred percent of his experience of this company.
Rarity is a property of the aggregate. Harm is a property of the individual. Those two facts sit in different columns of the same spreadsheet and get added together constantly.
What the dashboards cannot see
Ostermill was not flying blind. This is the part that takes a minute, because the natural assumption is that a company about to make this mistake is a company with bad measurement, and Ostermill's measurement was fine.
Four numbers ran receivables, and every one of them is defensible. Days sales outstanding. Percentage of invoices aged past terms. Collections per clerk. Write-offs against the small-balance policy. All four had been in use for years, all four were accurate, and all four would appear on any competent finance dashboard in this industry.
Now put the agent behind them and read them again.
Days sales outstanding improves if more reminders go out sooner. Percentage past terms improves the same way. Collections per clerk improves dramatically, because the clerk is no longer writing the letters. Write-offs against policy stay flat, since the policy has not changed.
Every number on that dashboard improves if the agent sends the reminder to the family shop.
Follow that one case through, because the abstraction is easy to nod at.
The invoice is real. It is four hundred and ten dollars for parts delivered in May, it is forty-one days past terms, and there is no dispute flag on the account because nobody disputed anything. The owner died in July. His daughter has been opening the mail since, between a job of her own and a funeral and the question of whether the shop continues at all.
The reminder goes out on a Tuesday. It is correctly worded, correctly addressed, and it cites the right invoice number.
Days sales outstanding: improved, because the reminder went out eleven days earlier than Ruth would have sent one, and the daughter pays it that Friday because she is the kind of person who pays things. Percentage past terms: improved, same reason. Collections per clerk: improved substantially, because no clerk was involved at any point. Write-offs against the small-balance policy: unchanged, correctly, since four hundred and ten dollars sits above the floor and always did.
Four green numbers, all of them accurate, none of them wrong.
And Ostermill would lose a customer of nineteen years, which would show up nine months later as an account that stopped ordering, in a quarter where several accounts stopped ordering, attributed by everyone who looks at it to the economy.
The error did not go undetected. It was never a candidate for detection.
That is not an oversight by whoever designed the metrics, and calling it one gets the lesson backwards. Those four numbers measure the work that was visible at the time they were designed, and the visible work was the sending. Ruth's forty-one judgment calls a week were invisible in a very specific sense: they happened inside the same person as the sending, so they never needed an instrument of their own. You could not lose the judgment without losing Ruth.
Ruth was not going anywhere.
Separate the two, hand the sending to a machine, and the judgment becomes a thing that can degrade quietly while every number on the dashboard gets better. The measures this company needs do not exist yet, and the reason they do not exist is that the failure they would catch was previously impossible.
So Ostermill ends up carrying two sets of numbers through the build, and the second set is the one this book is about.
The first set is the one the business already runs on. Days sales outstanding, past-terms percentage, collections per clerk. The agent must not damage them, and if it does the project is over on the arithmetic alone.
The second set has to be invented, and every number in it is about the holds rather than the sends. How often the agent proposes contacting somebody it should have left alone. How often a person overrides it, and in which direction, because an override that always goes one way is telling you something different from one that goes both. How long it takes anybody to notice when the answers get worse.
None of those existed at Ostermill in February. None of them exists at most companies now.
What this book does with them
One thing about the shape of what follows, so it reads as evidence rather than as a story with its ending arranged in advance.
Ostermill launches at the lowest autonomy that could prove anything, and stays there longer than anyone on the project expects. Four of the five phases end in a decision to do less than was proposed. One capability that everybody agreed to, and that carried the best-argued justification in the whole design, is taken away eleven weeks later by the person who approved it, and earned back a year after that.
A near-miss from Ruth's ordinary week becomes, months afterward, a graded case, which is a test with her judgment attached to it, and it stops a promotion.
The agent is still running on the last page of this book, and it still does not touch disputes, bankruptcy, or death. I am going to argue that this is the design working rather than the design falling short, and the argument is the same one Chapter 1 made about the way Ruth divides. What she knew became the agent. What she noticed became the supervision. Neither half was ever going to cover the cases where the information does not exist in any system, and a design that pretended otherwise would have been a worse design with better slides.
One more thing about Ostermill, and it is a declaration rather than a note.
The company does not exist. Neither do Ruth, Marcus, Tom, Dana or Priya, and neither does the machine shop in Dayton. Ostermill is a designed composite, built out of receivables operations at several real distributors, and every number attached to it was chosen to be internally consistent rather than to be impressive.
That has a cost and it buys something, and both are worth stating plainly.
What it costs is the authority of a real case. Nothing here proves anything happened. A composite cannot be checked, cannot be audited, and cannot be cited by a reader arguing with a colleague.
What it buys is the paperwork. A real company will not show you its go memo, its eval readout with the failing cases still in it, or the meeting where somebody revoked a capability they had approved eleven weeks earlier. Those documents exist and they are the most instructive artifacts in this entire field, and they are exactly the ones that never leave the building. The composite lets this book put them on the page in full.
The rule I have followed is that the fiction carries the shape of the work and never carries the evidence. Where a claim rests on something outside this book, it is sourced, identified and checkable, and where it rests on Ostermill it is teaching material and says so. When you reach a figure in one of the exhibits, it is there because it is consistent with the case, not because it happened.
If that trade seems wrong to you, the four books this one draws on, named at the front, are built the other way, out of documented incidents and cited research, and they are the place to check the arguments this volume compresses.
One thing the composite does not buy is a project that behaves well.
What follows is not an academic exercise, and it is not a research proposal with the answer worked out at the back. It is a project, and it behaves like one. Some of it goes the way the plan said it would. A good deal of it does not, and those are the parts worth reading, because a version with only the successes in it would be an advertisement. Nobody in it is stupid and nobody is a villain. They decide with what they have on the day they have to decide, and some of what they decide turns out to be wrong in a way that only becomes visible two phases later.
What happens at Ostermill happens because somebody says one sentence out loud in a meeting, in late February, at the end of a monthly operations review. It is a sentence somebody is saying in your company this quarter, and it sounds like the least controversial thing anyone said that day.