The Last Reader

The Job Was Never the Query

#

There was an analyst at a company I know who kept a bowl of green apples on her desk and never ate one.

The case for replacing her is easy to make and mostly correct. She spent her days writing queries against a warehouse, and writing queries is a mechanical skill that a competent model now performs faster and with fewer errors. The turnaround on a question was three days. It could be eight seconds. Nobody who has sat in a meeting waiting on a number wants to defend three days.

The usual objection to this is that the analyst knows things the schema does not contain. It is a true objection and it is weaker than the people making it believe. She knew there were six definitions of revenue in the warehouse and that the finance director meant the fourth one. She knew the Q3 2024 numbers were unusable because of a migration nobody documented. She knew that the table with the authoritative-sounding name had been abandoned in 2023 and that people still joined to it. All of that is tacit, all of it lives in one head, and all of it is exactly the kind of thing that gets written down eventually, gets absorbed, and stops being scarce. Betting the case against automation on the persistence of undocumented knowledge is betting on documentation never happening, which is a bet against the only trend in the industry that has ever been reliable.

So concede it. Assume the system learns the domain. Assume it knows about the migration and the abandoned table and the four other definitions of revenue, and it flags them, and it is right more often than she was.

Something is still gone, and it is not knowledge.

Somebody Had To Be Convinced

#

Every question that reached that desk was read by a person before it produced a number.

This was not a review step. There was no review step. It was a consequence of the physical arrangement: a human being with finite hours stood between a question and its answer, and that human being had opinions about the question. Some questions she answered immediately. Some she pushed back on. Some she reformulated into a different question and answered that one instead, and told the requester why. A large number she killed in the doorway by asking what the person was going to do with the number, and discovering that they did not know.

Nobody assigned her this. It is not in any job description written for the role, at that company or anywhere. It happened because asking a person for something requires you to state what you want out loud to someone who might say no, and stating things out loud to a potential objector is a different cognitive operation than typing them into a box.

The psychology here is not folklore. The finding that reasoning improves under accountability to an audience is one of the most consistent results in the literature, and it holds specifically when the audience’s views are unknown and the person expects to have to justify themselves. The argumentative account of reasoning goes further and says that the faculty is built for this: it evolved for producing and evaluating justifications in front of other people, it performs poorly in isolation, and its characteristic failure in isolation is generating support for whatever the reasoner already wanted.

An interface that answers in eight seconds is isolation with a fast response time.

Reasoning degrades when nobody has to be convinced, and the product is being sold on the elimination of the person who had to be convinced.

The Question With the Answer Inside It

#

The way this fails is specific and it is not about accuracy.

Consider a question of the form: show me why the West region underperformed. The question contains its conclusion. An analyst hearing it in a doorway says the West region did not underperform, or asks against what, or points out that the comparison period includes the reorganization. The question dies or improves.

A system receiving the same question complies. It is built to comply; compliance is the product. It will find something, because there is always something, and it will render it as a chart, and the chart will appear in a meeting the following morning attached to a premise nobody examined. The number will be correct. The analysis will be sound. The finding will be an artifact of the question.

Then there is the volume problem, which is worse and almost nobody is discussing it.

The statistical machinery organizations use to decide whether a pattern is real was calibrated for a world where asking was expensive. When a question costs three days of somebody’s time, a team runs perhaps five comparisons a quarter, and the odds of one of them being noise are tolerable. When a question costs eight seconds, the same team runs four hundred, and reports the three that came back interesting, and applies no correction of any kind, because the four hundred were never counted and no one experienced them as a search. Gelman and Loken’s description of the garden of forking paths covers exactly this: you do not need to be p-hacking deliberately, you only need a flexible analysis path and a lot of choices, and the choices produce significance out of nothing.

Cheap questions do not just increase the number of answers. They change the base rate at which answers are true, and they do it invisibly, and the dashboards look the same.

Latency was rationing the questions, and the rationing was holding up the statistics.

The Filter Was Bad and It Was a Filter

#

The honest complication is that the arrangement being dismantled was unfair in ways that deserve stating plainly.

The analyst was a bottleneck with preferences. Questions from senior people got answered and questions from junior people waited. Questions from departments she liked got the careful version. Requests she found stupid got a slow, technically compliant answer that discouraged repeat business. Whole categories of curiosity never got expressed at all, because people learned what it cost to ask and stopped asking, and the things they stopped asking were disproportionately the exploratory, low-status, might-be-nothing questions that occasionally turn out to be the important ones.

That filter was regressive. It rationed access to evidence by seniority and social capital. Removing it is a real gain and the people celebrating it are not wrong to celebrate.

But what replaces a bad filter is not a good filter. It is no filter, and the difference between a bad filter and no filter is not the difference between bad and good. It is the difference between a system that rejects the wrong things and a system that rejects nothing, which are failures of different kinds and only one of them is visible from inside.

The fix is not restoring the analyst. It is building the challenge into the thing that answers: a system that asks what you intend to do with the number, that reports how many comparisons the session has run, that says the premise in your question is not supported before it complies with the question. All of that is technically trivial. None of it will ship, because every part of it makes the demonstration worse, and the demonstration is what gets bought.

I wonder whether any organization will notice the challenge function was gone before it has been gone for a decade.

Why This One Disappears Quietly

#

The general shape here is a function that no one assigned, no one measured, and everyone experienced as an obstacle.

Merton described the category almost a century ago and the corpus has been circling it for a long time under other headings. A practice generates an effect nobody intended; the effect is load-bearing; the practice is removed for reasons entirely unrelated to the effect; the effect vanishes with no entry anywhere recording that it existed. Chesterton’s version asks you to find out why the fence was put up before removing it, which is good advice and useless here, because nobody put this fence up. It condensed out of the fact that a person was in the way.

Things that were never anyone’s job cannot be defended in a business case, cannot be assigned an owner during a transition, and cannot appear in a before-and-after comparison, because there is no metric that was ever tracking them. The transition will show faster answers, more answers, higher satisfaction among people who ask questions, and reduced headcount. Every one of those numbers will be true. None of them measures whether the questions got better or worse, and no one is going to build the instrument that would, because building it requires having the idea first, and the whole difficulty is that the idea does not occur to anyone until long after.

A function nobody owns cannot be cut, because cutting requires a decision, and this will not be a decision.

The meetings will be faster. Somebody will ask a question in the room and get a chart in the time it takes to finish the sentence, and the chart will be right, and the premise underneath it will be whatever the person who asked already believed on the way in. Everyone will call this an improvement, and for a while it will be one.

She kept the apples because people took them. They would stop at the desk, pick one up, stand there for a second turning it over, and then, having already stopped, say what they had come to ask. She got the question standing up, in person, with the requester still in front of her, before it was a ticket and before it was a specification, at the only moment it was still soft enough to change. It was the cheapest interrogation instrument anyone in that building ever deployed and it cost about nine dollars a week and no one, including her, would have described it as part of the job.

This essay applies the argument form named in the essay on unnamed absences (What the Gaps Were Doing) to a function that no organization has ever assigned to anyone: the pushback a question receives from the person who has to answer it. It follows Parts 107 and 108 (The Older Approximation and The One That Always Answers) on what happens when a responsive system enters a room, and connects to the Unit of One material on firms that stopped guessing, where the same speed appears as a clean gain.

References

#

Accountability and the Social Function of Reasoning:

Lerner, Jennifer S., and Philip E. Tetlock. “Accounting for the Effects of Accountability.” Psychological Bulletin, vol. 125, no. 2, 1999, pp. 255-275.

Mercier, Hugo, and Dan Sperber. The Enigma of Reason. Harvard University Press, 2017.

Nickerson, Raymond S. “Confirmation Bias: A Ubiquitous Phenomenon in Many Guises.” Review of General Psychology, vol. 2, no. 2, 1998, pp. 175-220.

The Statistics of Cheap Questions:

Gelman, Andrew, and Eric Loken. “The Statistical Crisis in Science.” American Scientist, vol. 102, no. 6, 2014, pp. 460-465.

Ioannidis, John P. A. “Why Most Published Research Findings Are False.” PLoS Medicine, vol. 2, no. 8, 2005, e124.

Simmons, Joseph P., Leif D. Nelson, and Uri Simonsohn. “False-Positive Psychology: Undisclosed Flexibility in Data Collection and Analysis Allows Presenting Anything as Significant.” Psychological Science, vol. 22, no. 11, 2011, pp. 1359-1366.

Benjamini, Yoav, and Yosef Hochberg. “Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing.” Journal of the Royal Statistical Society: Series B, vol. 57, no. 1, 1995, pp. 289-300.

Tacit Knowledge in Organizations:

Polanyi, Michael. The Tacit Dimension. Doubleday, 1966.

Nonaka, Ikujiro, and Hirotaka Takeuchi. The Knowledge-Creating Company. Oxford University Press, 1995.

Unintended Function and Its Removal:

Merton, Robert K. “The Unanticipated Consequences of Purposive Social Action.” American Sociological Review, vol. 1, no. 6, 1936, pp. 894-904.

Chesterton, G. K. The Thing: Why I Am a Catholic. Sheed and Ward, 1929.

Perrow, Charles. Normal Accidents: Living with High-Risk Technologies. Basic Books, 1984.