Cognitive Transition · Book One
Chapter 4 · Real Capabilities, Without the Narrative
“The system optimizes for plausibility, not truth.”
Summary
A report arrives flawless: well ordered, not a typo, in the tone of someone who knows. That is all the mechanism guarantees. This chapter takes the previous chapter's portrait down to the workshop and examines the system the way you would examine any component an organization wants to rely on: how it produces, what it lacks, and what it does better than anyone. It explains why fluency tells you nothing about truth, why the absence of accountability is a stable trait of the system and not a flaw the next version will fix, and on which fronts the machine genuinely outperforms the professional. It also shows the terrain where the expert remains indispensable, because there output without judgment is useless or dangerous. From that comes a thesis about your craft: the expert doesn't disappear, the role changes. The tool is a six-question checklist for deciding what gets automated, what gets assisted, and what gets protected.
Extended content
Liquid part: this text keeps growing after the printed edition.
Yesterday's output, under the lens
Open the last AI output you used without touching it. With a marker, underline every claim that could be checked outside the text: a date, a regulation, a figure, a name, a reference. Count how many you checked at the time. The ratio tends to be uncomfortable.
Now pick three of the underlined ones and actually check them. Write down where you looked, who maintains that source, and how current it is. What you are training is a gesture: stepping outside the text. The model doesn't step outside. That trip falls to you.
If none of the three fails, don't celebrate: you have checked three claims, not the document. If one does fail, look at how it was written. With the same aplomb as the others.
A worked example: Rocío's three functions
Rocío runs the sales area at an installation company, six people, and she puts the chapter's checklist to three functions.
The one already automated: sorting incoming emails by type of request. It goes well until question five. The person receiving the result is the client, and a client can't know that their urgent email landed in the wrong inbox. The error costs little per email and a great deal on the day the email is a formal complaint. The change: the model sorts, and every morning someone reviews a sample of what was tagged "routine."
The one she's considering: drafting quotes. She trips on question four. A good quote includes what the client didn't ask for: roof access, the municipal permit, the line item that always gets forgotten. The model organizes what it's given; what's missing gets caught by whoever has seen job sites where it was missing. She decides to automate the format and the standard terms, and leave the list of absences to a person with a seat at the table.
The one she thinks untouchable: the call with a client threatening to leave. The core confirms what she expected: ambiguous, expensive, relational. But around the call, pure execution turns up: summarizing the history beforehand, writing up the minutes afterward, preparing the follow-up draft. That gets delegated, and it frees an hour for the call.
Notice what Rocío never asked: which model is better. She asked who receives the result, what's missing from the page, and where the relationship lies. The checklist speaks about the function, and the function existed before the tool.
Where your checkpoint is
Pick a workflow where AI already takes part. Draw it as a single line with arrows, from request to effect: what goes in, what the model generates, what goes out to someone else. Answer in writing:
- At which arrow does someone with competence and authority look at the output before it takes effect?
- Does that person have real time to look, or do they sign out of inertia?
- What happens if they find an error: whom do they tell, and what gets corrected?
If any answer is blank, the workflow has no checkpoint. And the risk, with nobody to stop it, travels toward whoever is least able to detect it.
Questions to work with
- What framework did I hand the model this week without having questioned it myself first?
- What output of mine do others receive without being able to verify it?
- What absence in my current project am I the only one positioned to see?
- Who showed me where the model's edge lies this season, and who am I showing?
Execution and judgment, in two columns
For five days, at the end of the workday, write down what you did in two columns. In one, execution: producing, drafting, calculating, assembling. In the other, judgment: evaluating, correcting, contextualizing, deciding, answering for the result. Don't look for the ideal split. Look for which of the two columns grows from one day to the next, and which is harder to fill.
If by Friday almost everything sits in the first, nothing is broken yet. There is a relocation pending.
Moving on
The other half of the ledger remains: what it costs to use this at scale. Chapter 5 sticks to the verifiable risks.
© Pedro Mujica, 2026