The Last Reader — Summary

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

The case for replacing her is easy and mostly correct. Writing queries is a mechanical skill a model performs faster with fewer errors, and nobody who has waited three days for a number wants to defend three days. The usual objection, that she knows what the schema does not contain, is weaker than its defenders think. Six definitions of revenue, the migration that ruined Q3 2024, the authoritative-sounding table abandoned in 2023: all tacit, all eventually written down. Betting against automation on the persistence of undocumented knowledge is betting against documentation.

So concede it. Assume the system learns the domain and is right more often than she was. Something is still gone, and it is not knowledge.

Every question that reached that desk was read by a person before it produced a number. There was no review step. A human with finite hours stood between a question and its answer and had opinions about it. She pushed back, reformulated, and killed a large number in the doorway by asking what the requester intended to do with the number and finding they did not know. Nobody assigned her this. Reasoning improves under accountability to an audience whose views are unknown, and the argumentative account says the faculty is built for that and fails in isolation by generating support for whatever the reasoner already wanted. An interface answering in eight seconds is isolation with a fast response time.

The failure is not about accuracy. Show me why the West region underperformed contains its conclusion. An analyst kills it or improves it. A system complies, because compliance is the product, and finds something, because there is always something. Worse is the volume problem. When a question costs three days a team runs five comparisons a quarter. At eight seconds it runs four hundred, reports the three that came back interesting, and corrects for nothing, because the four hundred were never counted. Cheap questions change the base rate at which answers are true, invisibly.

The complication is that the filter was regressive. Senior people got answered, junior people waited, and the exploratory low-status question never got asked. Removing it is a real gain. But what replaces a bad filter is no filter, and only one of those two failures is visible from inside. Building the challenge into the thing that answers is trivial and will not ship, because every version makes the demonstration worse.

Nobody put this fence up. It condensed out of the fact that a person was in the way, and functions nobody owns cannot be cut, because cutting requires a decision.

She kept the apples because people took them. They would stop, pick one up, and having already stopped, say what they had come to ask. She got the question standing up, before it was a ticket, at the only moment it was still soft enough to change.