What AI can do, and what it cannot

Before any framework, name the reflex the framework exists to interrupt. When a person reads a well-structured, confidently worded answer, they transfer their trust from the form of the answer to its content. The prose is clean, therefore the reasoning must be sound. The citation is formatted correctly, therefore the source must exist. The number carries two decimal places, therefore someone calculated it.

You cannot train Fluency Transfer out of yourself by resolving to be more careful. Vigilance decays under deadline, and the outputs that most deserve scrutiny are the ones that read most smoothly. What works instead is a gate applied before the work starts, while you are still calm and the answer does not yet exist to seduce you.

Definition

The Reversibility Gate is a two-question test applied before you delegate a task to an AI system, never after. First: if this output is wrong and acted on, what does it cost to undo, and who pays that cost? Second: what specific evidence, named now, would make this output trustworthy? The two answers together determine one of three delegation verdicts, and the verdict is binding.

The gate is not a risk score and it is not a policy. It is a sentence you must be able to complete out loud before you open the tool. If you cannot complete it, you have not yet understood the task well enough to delegate any part of it.

When to Use It

Use the gate every time an AI output will leave your own screen. That includes a draft you will send, a summary someone else will act on, a classification that will route real work, and a number that will appear in a document with your name on it. It does not include exploratory use where nothing you produce reaches another person or system.

Run it again whenever the task changes shape. The same tool that safely drafts internal notes becomes a different problem the moment those notes start going to customers. Delegation permissions attach to tasks, not to tools.

How to Apply It

1. 1State the undo cost in units someone recognises: minutes, dollars, a customer conversation, a regulatory notification.
2. 2Name the person who pays that cost. If it is not you, the gate tightens.
3. 3Name the evidence in advance, before the output exists, and be specific: the recording, the source contract, the ledger, the original report.
4. 4Choose the verdict, write it down, and hold to it even when the output looks excellent.

Worked example 1 of 3

Marisa Delgado runs servicing operations at OmniCorp Financial. Her team handles roughly nine hundred customer calls a week, and every call needs a written summary in the servicing record. She wants to delegate the summaries. She runs the gate before she runs the pilot.

MDMarisa Delgado

If a summary is wrong, what does it cost to undo?

SSServicing supervisor

Depends what is wrong. A clumsy sentence costs nothing. A commitment we never made costs us the argument six months later, when the customer quotes our own record back at us.

MDMarisa Delgado

So the undo cost is a disputed obligation, and the customer pays first. That is not Draft. What evidence would make one of these trustworthy?

SSServicing supervisor

The call recording. Specifically, any sentence in the summary that says we agreed to something has to appear in the recording.

MDMarisa Delgado

Then this is Assist, and the evidence rule is that every commitment sentence gets checked against the audio before the record is saved. Write that down.

The pilot ran for three weeks. The model produced good summaries. In week two it also produced a summary asserting that OmniCorp Financial had agreed to waive a fee, which the customer had requested and the agent had declined. The reviewer caught it in eleven seconds, because she was looking for exactly one thing and knew where to look.

Why This Works

Naming the evidence in advance converts verification from an act of judgment into an act of comparison. Judgment is expensive, subjective, and the first thing to go under time pressure. Comparison is cheap, mechanical, and survives a bad Tuesday. Marisa's reviewer did not have to decide whether the summary was good: she had to check one class of sentence against one artifact.

Naming the payer does the other half of the work. Fluency Transfer thrives when the cost of being wrong is abstract. It weakens sharply when the person bearing that cost has a name and is not you.

Priya Raghunathan manages dispatch operations at OmniCorp Logistics. She is asked to use AI to draft the daily exception report that tells three regional managers which shipments slipped and why. She works the gate in order rather than debating the tool.

Undo cost: a manager reallocates two drivers based on a slippage that did not happen, and the real slippage goes uncovered. Payer: the customer whose shipment misses its window, then the regional manager. Evidence: every shipment identifier and every stated delay reason must match the dispatch system record. Verdict: Assist, with the identifier check mandatory and the narrative left to the model.

In the first week the drafts were unremarkable and the check took four minutes. In the second week a draft attributed a delay to weather at a depot that had reported a mechanical fault. The identifier matched; the reason did not. Priya widened the evidence rule from identifiers to identifiers and reason codes, which cost another two minutes a day, and kept the verdict at Assist. She did not withdraw the tool, and she did not loosen the check. Both would have been easier.

One gate, applied consistently, will catch most of what goes wrong in individual AI use. It will not tell you what to do when the task is shared across a team, when the evidence is contested, or when the verdict has to survive a regulator asking how it was reached. Building oversight that holds up under that pressure is the capability the paid programs develop next.